An empirical analysis of poverty, inequality and the labour market in Malawi
by
Anderson Sawira Gondwe
Dissertation presented for the degree of Doctor of Philosophy (Economics) in the Faculty of Economic
and Management Sciences at Stellenbosch University
Supervisor: Prof. Servaas van der Berg
December 2016
Stellenbosch University https://scholar.sun.ac.za
Declaration
By submitting this dissertation electronically, I declare that the entirety of the work contained therein is
my own, original work, that I am the sole author thereof (save to the extent explicitly otherwise stated),
that reproduction and publication thereof by Stellenbosch University will not infringe any third party
rights and that I have not previously in its entirety or in part submitted it for obtaining any qualification.
December 2016
_____________________________
(Anderson Gondwe)
Copyright © 2016 Stellenbosch University
All rights reserved
i
Stellenbosch University https://scholar.sun.ac.za
Abstract
This thesis is a consolidation of three related studies on Malawi. The first study contains spatial and
temporal comparisons of poverty and inequality in Malawi using two non-monetary dimensions, namely
an asset index and child nutritional status. Through stochastic dominance tests, the study establishes that
poverty and inequality are unambiguously higher in rural areas, which contain 85% of the population,
in the Southern region and among households headed by females. Results indicate that poverty has
significantly declined over time and that the gains from growth have been pro-poor. We show that
welfare does not vary much across regions and areas with respect to child nutritional status but there are
large differences in asset poverty. Stunting is a bigger problem among children under the age of five
than body wasting and being underweight. Econometric analysis shows that asset ownership is
positively associated with household size, the age of household head and education attainment. Age
dependency ratio and incidence of sickness are negatively associated with asset ownership. Multivariate
analysis of child nutrition reveals that malnutrition first worsens before improving at some critical age.
This is consistent with possible recovery found in some of the studies that track children over time. Also
in accordance with some literature, we find that boys have weaker nutritional status than girls.
The second study looks at the role of education in poverty reduction identified through the labour
market. This study contributes to research on returns to education by including self-employment
activities and non-farm business enterprises. Unlike previous studies, this study uses panel data which
has many advantages, as acknowledged in the literature. We find large and positive returns to education
in Malawi suggesting that education is a good investment. The returns increase with the levels of
education. Interestingly, females have higher returns to education than males with similar skills. Since
the Malawian labour market is not homogeneous, our analysis distinguishes between the formal and
informal employment sectors. Furthermore, studying Malawi’s informal sector is important as it
accounts for 78% of total employment. Our results show that education externalities exist and play an
important role in non-farm enterprises. The findings are robust to sample selection and treatment of
outliers. We further show that dealing with inconsistencies in the data helps improves the quality and
reliability of the results.
The third study applies spatial panel data econometric techniques to the study of migration and
employment in Malawi. The study shows that the magnitudes of coefficients drop after taking into
account spatial dependencies. This confirms that studies that fail to take into account the spatial effects
tend to overstate the results. By matching geographical codes that are consistent over time, it is now
feasible to integrate census data with other data for similar spatial analysis. The study further evaluates
the impact of land reform policy on spatial migration and employment using a difference-in-difference
ii
Stellenbosch University https://scholar.sun.ac.za
estimation strategy. Results show that the policy has had significant effects on migration and
employment patterns in Malawi.
iii
Stellenbosch University https://scholar.sun.ac.za
Opsomming
Hierdie tesis is a konsolidasie van drie verwante studies oor Malawi. Die eerste studie ondersoek
armoede en ongelykheid in Malawi oor tyd en ruimte heen deur twee nie-monetêre dimensies, naamlik
'n bate-indeks en die voedingstatus van kinders, te gebruik. Deur middel van stogastiese dominansietoetse word ondubbelsinnig getoon dat armoede en ongelykheid hoër is in landelike gebiede, wat 85%
van die bevolking huisves, in die Suidelike streek en onder huishoudings met vroue as hoof van die huis.
Resultate toon dat armoede beduidend afgeneem het en dat groei tot voordeel van die armes strek. Ons
resultate toon weinig verskille in welsyn tussen streke en gebiede met betrekking tot die voeding status
van kinders, maar groot verskille in bate-armoede . Vertraagde groei is 'n groter probleem by kinders
onder die ouderdom van vyf jaar as kwyning en ondergewig. Ekonometriese ontleding toon dat batebesit positief verband hou met die grootte van die huishouding en die ouderdom en opvoedingsvlak van
die hoof van die huishouding . Die ouderdom-afhanklikheidslas en die voorkoms van siekte hou negatief
verband met bate-besit. Regressie-analise wys dat wanvoeding onder kinders eers met ouderdom
toeneem voordat dit by hoër ouderdomme afneem, wat konsekwent is met die moontlikheid van herstel
soos party studies wat kinders oor 'n tydperk volg bevind. Ook, in ooreenstemming met party studies,
word bevind dat die voedingstatus van dogters beter is as dié van seuns.
Die tweede studie bestudeer die rol van onderwys in die vermindering van armoede in die arbeidsmark.
Deur die insluiting van selfwerksaamheidsaktiwiteite en nie-landbou sakeondernemings dra die studie
by tot navorsing oor die voordele van opvoeding in Malawi. Anders as in vorige studies, gebruik hierdie
studie paneeldata, wat baie voordele inhou, soos in die literatuur bevestig. Ons vind groot en positiewe
opbrengste op onderwys, wat daarop dui dat dit 'n goeie belegging is. Opbrengste neem toe met vlakke
van onderwys. Interessant genoeg, ervaar vroue hoër opbrengste op belegging in onderwys as mans met
dieselfde vaardighede. Aangesien die arbeidsmark in Malawi nie homogeen is nie, tref ons analise ‘n
onderskeid tussen die formele en informele indiensnemingsektore. Dit belangrik om Malawi se
informele sektor in ag te neem, aangesien dit 78% van die totale indiensneming uitmaak. Ons resultate
wys dat daar eksternaliteite van onderwys bestaan wat 'n belangrike rol speel in nie-landbou
ondernemings. Ons resultate is robuust virsteekproefseleksie en die hantering van uitskieters. Die
uitstryk van data-onreëlmatighede dra tot 'n verbetering in die kwaliteit en betroubaarheid van die
resultate by.
Die derde studie pas ruimtelike paneeldata ekonometriese tegnieke toe op migrasie en indiensneming in
Malawi. Die grootte van koëffisiënte neem af as ruimtelike afhanklikhede in ag geneem word. Dit
bevestig dat studies wat nalaat om ruimtelike aspekte in berekening te bring geneig is om effekte te
oorskat. Deur konsekwente geografiese kodes oor tyd te verbind is dit nou moontlik om sensusdata met
ander data te integreer vir verdere ruimtelike analise. Die studie evalueer ook die uitwerking van die
iv
Stellenbosch University https://scholar.sun.ac.za
grondhervormingsbeleid op ruimtelike migrasie en indiensneming deur die gebruik van 'n verskil-inverskille metodeevalueer. Die resultate dui daarop dat hierdie beleid 'n beduidende uitwerking op
migrasie en werkloosheid in Malawi het.
v
Stellenbosch University https://scholar.sun.ac.za
Acknowledgements
Firstly, I would like to thank Professor Servaas van der Berg from whom I have received tremendous
feedback and guidance throughout my PhD study. I will forever remember him for his positive
contribution to my life and for opening my mind to the big picture of research. I know Servaas as a very
humble, patient and wise man with outstanding fatherly care. Considering my non-academic
background, the fact that I am completing my studies within three years speaks volumes about how
Servaas is able to identify and develop talent in people.
Secondly, I acknowledge the contribution from Research on Social Economic Policy (RESEP) which is
led by Servaas. RESEP has been important to my progress and timely completion of the studies. Apart
from the networking, the organisation has been the source of the much needed additional funding for
my studies and academic conferences. The RESEP team also boasts an excellent pool of researchers
from which I have greatly benefited. It has provided a platform for the transfer of new skills and refining
my work through the weekly departmental seminars and training workshops. I particularly thank Dr
Dieter von Fintel for his insights that have also helped shape my study.
Thirdly, I also say my heart-felt thanks to The Stellenbosch Institute for Advanced Study (STIAS) for
the scholarship funding provided through The Graduate School of Economic and Management Sciences
(GEMS). The GEMS cohort system, which recruits PhD researchers from different parts of the world,
has enabled me to develop important networks with colleagues from three different continents, namely
Africa, Asia and Europe. In this regard, special mention goes to Dr Jaco Franken, the manager of the
graduate school programme and fellow PhD students in the programme.
Lastly, I thank my family and friends who have provided encouragement to me during my physical
absence from them.
I have made it to the glory of the LORD and Jesus Christ.
vi
Stellenbosch University https://scholar.sun.ac.za
Dedication
I dedicate this work to my family and particularly Mary Gondwe. My wish is that there may never cease
to be people who attain PhD education throughout our family generations.
vii
Stellenbosch University https://scholar.sun.ac.za
Table of Contents
Declaration ............................................................................................................................................... i
Abstract ................................................................................................................................................... ii
Opsomming ............................................................................................................................................ iv
Acknowledgements ................................................................................................................................ vi
Dedication ............................................................................................................................................. vii
Table of Contents ................................................................................................................................. viii
List of Figures ........................................................................................................................................ xi
List of Tables ......................................................................................................................................... xii
Chapter 1 ................................................................................................................................................. 1
1.1
Introduction ............................................................................................................................. 1
1.2
Geography and history ............................................................................................................ 1
1.3
Economy.................................................................................................................................. 2
1.4
Problem statement ................................................................................................................... 5
1.5
National data sources............................................................................................................... 6
1.6
Thesis structure........................................................................................................................ 7
Chapter 2 ............................................................................................................................................... 10
2.1
Introduction ........................................................................................................................... 10
2.2
Theoretical considerations in poverty measurement ............................................................. 12
2.3
Inequality measurement ........................................................................................................ 14
2.4
Stochastic dominance analysis .............................................................................................. 15
2.5
Poverty and inequality decomposition .................................................................................. 17
2.6
Pro-poor growth analysis....................................................................................................... 17
2.7
Data ....................................................................................................................................... 20
2.8
Poverty lines .......................................................................................................................... 27
2.9
Cumulative density curves .................................................................................................... 27
2.10
FGT poverty estimates .......................................................................................................... 28
2.11
Poverty dominance analysis .................................................................................................. 30
2.12
Gini and GE inequality estimates .......................................................................................... 31
2.13
Inequality dominance analysis .............................................................................................. 33
2.14
Poverty decomposition .......................................................................................................... 34
2.15
Subgroup inequality decomposition ...................................................................................... 34
2.16
Spatial distribution of poverty and inequality ....................................................................... 37
2.17
Factors affecting asset poverty .............................................................................................. 40
viii
Stellenbosch University https://scholar.sun.ac.za
2.18
Child nutritional status in Malawi ......................................................................................... 42
2.19
Multivariate analysis of child nutrition ................................................................................. 44
2.20
Asset index and pro-poor growth analysis ............................................................................ 47
2.21
Pro-poor growth in child nutritional status............................................................................ 52
2.22
Conclusions ........................................................................................................................... 55
2.23
Policy discussion ................................................................................................................... 55
Chapter 3 ............................................................................................................................................... 57
3.1
Introduction ........................................................................................................................... 57
3.2
Methodology ......................................................................................................................... 59
3.2.1
Theoretical framework .................................................................................................. 59
3.2.2
Estimating returns to education ..................................................................................... 62
3.2.3
Sample selection ............................................................................................................ 63
3.2.4
Modelling unobserved heterogeneity ............................................................................ 64
3.3
Description of the data .......................................................................................................... 65
3.3.1
Work and non-work activities of the employed ............................................................ 66
3.3.2
Describing employment structure and hours worked .................................................... 67
3.3.3
Treatment of outliers, missing data and zero earnings .................................................. 69
3.3.4
Dealing with inconsistencies ......................................................................................... 71
3.4
Labour force participation ..................................................................................................... 71
3.4.1
Size of labour force and labour force participation rates............................................... 72
3.4.2
Changes in the labour force according to background characteristics .......................... 73
3.4.3
Shares in the labour force .............................................................................................. 75
3.4.4
Multivariate analysis of labour force participation........................................................ 76
3.5
Unemployment ...................................................................................................................... 78
3.6
Employment trends and characteristics ................................................................................. 79
3.6.1
Employment shares and growth rates ............................................................................ 79
3.6.2
Multivariate analysis of employment likelihood ........................................................... 80
3.7
Earnings and changes in employment status ......................................................................... 81
3.7.1
Employed in either wave ............................................................................................... 82
3.7.2
Employed in both waves ............................................................................................... 82
3.7.3
Identifying sources of increases in earnings .................................................................. 83
3.8
Econometric analysis of returns to education........................................................................ 85
3.8.1
Wage employment ......................................................................................................... 85
3.8.2
Household non-farm enterprise earnings ....................................................................... 97
3.9
Measurement error using panel data...................................................................................... 99
3.10
Comparing income and consumption .................................................................................. 100
ix
Stellenbosch University https://scholar.sun.ac.za
3.11
Dynamics of household percapita consumption .................................................................. 102
3.12
Conclusion and policy implications .................................................................................... 104
Chapter 4 ............................................................................................................................................. 106
4.1
Introduction ......................................................................................................................... 106
4.2
Data and methods ................................................................................................................ 107
4.2.1
Data ............................................................................................................................. 107
4.2.2
Theoretical framework ................................................................................................ 108
4.2.3
Spatial autocorrelation ................................................................................................. 109
4.2.4
Spatial panel data models ............................................................................................ 111
4.3
Descriptive analysis ............................................................................................................. 114
4.3.1
Spatial distribution of employment, education, assets and fertility ............................. 114
4.3.2
Fertility trends ............................................................................................................. 116
4.3.3
Changes in population age structure and labour supply .............................................. 118
4.3.4
Spatial and temporal patterns of migration.................................................................. 122
4.3.5
Spatial autocorrelation in variables ............................................................................. 125
4.4
Spatial panel regression results ........................................................................................... 127
4.4.1
Effects of land reform policy on migration ................................................................. 128
4.4.2
Effects of land reform policy on employment ............................................................. 130
4.5
Conclusions ......................................................................................................................... 133
Chapter 5 ............................................................................................................................................. 135
5.1
Introduction ......................................................................................................................... 135
5.2
Summary of findings ........................................................................................................... 135
5.3
Conclusions ......................................................................................................................... 138
5.4
Implications of the research ................................................................................................ 139
5.5
Summary of contributions ................................................................................................... 143
5.6
Future research .................................................................................................................... 144
List of references ................................................................................................................................. 146
Appendices .......................................................................................................................................... 154
Appendix A1: Coefficient comparison test ................................................................................. 154
Appendix A2: Dealing with data inconsistencies ........................................................................ 154
x
Stellenbosch University https://scholar.sun.ac.za
List of Figures
Figure 1.1: Political map of Malawi ........................................................................................................ 2
Figure 1.2: Malawi’s GDP per capita annual growth rates (1961-2015) ................................................ 5
Figure 2.1: Adjusted and unadjusted asset indices by population subgroups........................................ 24
Figure 2.2: Distribution of anthropometric Z-scores for HAZ, WAZ and WHZ .................................. 25
Figure 2.3: MCA asset index cumulative density curves by population groups ................................... 28
Figure 2.4: MCA asset index Lorenz curves by population subgroups................................................. 31
Figure 2.5: Spatial distribution of poverty using asset index and child-nutritional status .................... 37
Figure 2.6: Spatial distribution of inequality for asset index and child-nutritional status ..................... 39
Figure 2.7: Asset poverty incidence curves by survey year .................................................................. 48
Figure 2.8: Access to assets by type and survey period ........................................................................ 52
Figure 2.9: Poverty incidence curves for HAZ by DHS survey year .................................................... 53
Figure 3.1: Histogram of real monthly earnings by survey year and occupation .................................. 84
Figure 3.2: Rates of return on education by gender .............................................................................. 90
Figure 3.3: Kernel densities for annual household percapita consumption and income by year ........ 101
Figure 4.1: Map of Malawi showing administrative boundaries ......................................................... 108
Figure 4.2: Spatial distribution of status in employment for 2008 ...................................................... 115
Figure 4.3: Spatial distribution of employment type for 2008 ............................................................ 115
Figure 4.4: Spatial distribution of educational attainment for 2008 .................................................... 116
Figure 4.5: Spatial distribution of asset index, schooling and fertility for 2008 ................................. 116
Figure 4.6 : Pyramids showing populations and labour force participation by sex (1987) ................. 120
Figure 4.7: Pyramids showing populations and labour force participation by sex (1998) .................. 120
Figure 4.8: Pyramids showing populations and labour force participation by sex (2008) .................. 121
Figure 4.9: Proportions that moved between districts ......................................................................... 124
Figure 4.10: Spatial autocorrelation for employment .......................................................................... 126
Figure 4.11: Spatial autocorrelation for migration .............................................................................. 126
Figure 4.12: Measures of local spatial autocorrelation for employment and migration ...................... 127
Figure A 1: FISP evaluation panel data on real ganyu wages ............................................................. 155
xi
Stellenbosch University https://scholar.sun.ac.za
List of Tables
Table 1.1: Malawi’s HDI trends based on consistent methodology and data ......................................... 4
Table 2.1: Summary of Malawi DHS data sets used ............................................................................. 21
Table 2.2: Descriptive statistics for the poverty and asset indices ........................................................ 23
Table 2.3: Child malnutrition rates by population groups ..................................................................... 26
Table 2.4: Poverty headcount, average poverty gap and poverty severity estimates ............................ 29
Table 2.5: Poverty stochastic dominance test results for population subgroups ................................... 30
Table 2.6: Inequality estimates across population subgroups ............................................................... 32
Table 2.7: Generalised Lorenz dominance test results across population subgroups............................ 33
Table 2.8: FGT poverty sub-group decomposition ............................................................................... 35
Table 2.9: GE inequality decomposition for asset index and child nutrition ........................................ 36
Table 2.10: Summary descriptive statistics for the asset model ............................................................ 40
Table 2.11: OLS regression results for asset poverty ............................................................................ 42
Table 2.12: Descriptive statistics for the nutritional models ................................................................. 43
Table 2.13: OLS regression results for child nutritional status ............................................................. 45
Table 2.14: Descriptive statistics for asset index scores ....................................................................... 47
Table 2.15: Differences in poverty headcount indices for household asset ownership ......................... 49
Table 2.16: Indices of pro-poorness in child nutritional status between 1992 and 2010 ...................... 49
Table 2.17: Pooled mean access of assets by area and region, all periods ............................................ 51
Table 2.18: Descriptive statistics for HAZ ............................................................................................ 52
Table 2.19: Differences in the FGT poverty headcount index for HAZ ............................................... 54
Table 2.20: Indices of pro-poorness for HAZ ....................................................................................... 54
Table 3.1: Employment and occupation structures ............................................................................... 67
Table 3.2: Changes in the average weekly hours worked and years of education by year ................... 69
Table 3.3: Labour force participation rates according to characteristics ............................................... 73
Table 3.4: Changes in the labour force according to characteristics ..................................................... 74
Table 3.5: Shares of working-age population and labour force according to characteristics ................ 75
Table 3.6: Probit regressions on labour force participation................................................................... 77
Table 3.7: Unemployment shares and rates by year .............................................................................. 78
Table 3.8: Broad and narrow employment shares and growth by year ................................................. 80
Table 3.9: Two-step Heckman probit results on employment likelihood ............................................. 81
Table 3.10: Mean monthly total wages by employment status and survey period ................................ 82
Table 3.11: Mean monthly ganyu wages by employment status and survey period ............................. 83
Table 3.12: Employment status and education attainment of individuals ............................................. 83
Table 3.13: OLS results for log of real monthly wages ........................................................................ 86
Table 3.14: OLS, Fixed effects and random effects results for log of real monthly wages .................. 87
Table 3.15: OLS and random effects results for log of monthly wages using education categories ..... 89
Table 3.16: Regressions of log monthly wages by gender and employees in both waves .................... 91
Table 3.17: Random effect results based on alternative treatment of outliers....................................... 92
Table 3.18: Wage functions corrected for sample selection.................................................................. 93
Table 3.19: Probit on choice of employment sector .............................................................................. 94
Table 3.20: Regression results on log of monthly wages by sector without sample selection .............. 95
Table 3.21: Regression results on log of monthly wages by sector with sample selection ................... 96
Table 3.22: Regressions for monthly household non-farm enterprise earnings .................................... 98
Table 3.23: Average household percapita consumption by components and year .............................. 102
xii
Stellenbosch University https://scholar.sun.ac.za
Table 3.24: Consumption transition matrix ......................................................................................... 103
Table 4.1: Summary of spatial panel model options ........................................................................... 114
Table 4.2: Age-specific fertility rates (ASFR) and total fertility rates (TRF) ..................................... 117
Table 4.3: Changes in economically active population and labour force by sex (1987-2008)............ 119
Table 4.4: Proportions of population and labour force by age group .................................................. 122
Table 4.5: Inter-regional migration transition frequencies by gender ................................................. 123
Table 4.6: Inter-regional migration transition probabilities by gender ............................................... 123
Table 4.7: Results showing spatial dependencies in variables (pooled 1987, 1998 and 2008) ........... 125
Table 4.8: Effects of land reform policy on migration ........................................................................ 129
Table 4.9: Balancing tests for base period........................................................................................... 130
Table 4.10: Effects of land reform policy on agricultural employment .............................................. 131
Table 4.11: Effects of land reform policy on government employment ............................................. 132
Table A 1: Table showing the distribution of households by year and population subgroups ............ 156
Table A 2: Table showing the distribution of children by year and population subgroups................. 156
Table A 3: WHZ regression results by age category ........................................................................... 157
Table A 4: List of regions, districts and traditional areas consistent over time................................... 158
Table A 5: Results showing spatial dependencies in variables for 1987............................................. 163
Table A 6: Results showing spatial dependencies in variables for 1998............................................. 163
Table A 7: Results showing spatial dependencies in variables for 2008............................................. 163
Table A 8: Population figures for individuals aged (15 years and older)............................................ 164
Table A 9: Number of people who migrated to other districts (15 years and older) ........................... 165
Table A 10: Proportions of people who migrated to other districts (15 years and older) ................... 166
Table A 11: Effects of land reform policy on migration ..................................................................... 167
Table A 12: Effects of land reform policy on agricultural employment ............................................. 168
Table A 13: Effects of land reform policy on government employment ............................................. 169
Table A 14: Effects of land reform policy on self, wage and private employment (OLS only) ......... 170
xiii
Stellenbosch University https://scholar.sun.ac.za
Chapter 1
Introduction
1.1
Introduction
This Chapter provides some background information on Malawi to help the reader contextualise the
study before turning to the topics addressed in this thesis. In Section 1.2, we look at Malawi’s geography
and history. Section 1.3 discusses the economy. The problem statement is discussed in Section 1.4. The
background information on the national data sources is given in Section 1.5. Finally, Section 1.6 gives
the thesis structure.
1.2
Geography and history
Malawi is a landlocked country located in Southern Africa. It is bordered by Zambia to the north-west,
Tanzania to the northeast and Mozambique to the west, south and east as shown in Figure 1.1. The
country is a long strip of about 901 km and ranges between 80 and 161 km in width1. The total surface
area is about 118,484 km2 of which 94,276 km2 is made up of land. The remaining area is largely made
up of Lake Malawi, Africa’s third-largest fresh-water lake, about 475 km long (National Statistical
Office & ICF Macro, 2011).
Administratively, Malawi is divided into three regions, namely the Northern, Central and Southern
regions. Regions are also divided into districts and there are 28 districts in total: six, nine and 13 districts
in the North, Centre and South, respectively. Each of the 28 districts is further subdivided into traditional
authorities (TAs), ruled by senior chiefs. The smallest units of administration are villages typically
governed by village headmen or women. Only a small proportion (about 15%) of Malawi’s population
resides in urban areas (National Statistical Office & ICF Macro, 2011). The geographical division of
Malawi into regions, districts and traditional authorities provides an important dimension for
decomposition analysis as we will see later.
Malawi became a protectorate of Britain in 1891. From 1953 to 1963, Malawi (formerly called
Nyasaland) was part of the Federation of Rhodesia and Nyasaland together with Zambia (formerly
Northern Rhodesia) and Zimbabwe (formerly Southern Rhodesia). Malawi became independent in 1964
and gained the status of a republic in 1966 (National Statistical Office & ICF Macro, 2011). Although
politically independent for 52 years, the country is still highly dependent on foreign aid; the United
Kingdom, the European Commission, the Global Fund and the World Bank continue to make up the
four largest donors (Organisation for Economic Co-operation and Development, 2008). Multiparty
democracy was abolished in 1966 but later reintroduced in 1993 after a national referendum. During the
1
In Chapter 4 we compute distance matrices for spatial analysis.
1
Stellenbosch University https://scholar.sun.ac.za
first year of a multiparty democracy in 1994, primary education in government schools became free and
this resulted in a large increase in school enrolment (Kadzamira & Rose, 2015).
Figure 1.1: Political map of Malawi
Source: Own construction from country shapefiles
1.3
Economy
The Malawian economy is largely dependent on agriculture, which made up about 30% of the gross
domestic product (GDP) in 2015 and continues to directly benefit more than 75% of the population. The
agricultural sector in Malawi consists of the smallholder sector mainly for subsistence production and
the estate sector for exportation. The main food crops are maize, rice and cassava. In 2014, tobacco
generated about 64% of foreign exchange. Other important export crops in Malawi are tea and sugar,
accounting for about 9% and 8% of total export value, respectively (Reserve Bank of Malawi, 2015).
However, more recently, tobacco, which is the main cash crop for export, has come under threat because
of the world-wide anti-smoking campaign.
2
Stellenbosch University https://scholar.sun.ac.za
According to Food and Agriculture Organisation (FAO) AQUASTAT (2015), less than 1% of cultivated
areas is under irrigation2. Since much of Malawi’s agriculture is dependent on rains, food security and
household incomes are threatened by flooding and droughts. Irrigation development, if combined with
advancements in good cropping systems and inputs, has the potential of improving farm incomes and
food security for the majority of the population who are involved in subsistence agriculture. The Green
Belt Irrigation is one of the key priority areas identified in the national development agenda with the
aim of facilitating growth and development in the coming years by taking advantage of the country’s
abundant water resources. Potentially, 1 million hectares can be irrigated in Malawi. The current project
is expected to expand the amount of land under irrigation from 90,000 to 400,000 hectares (Government
of Malawi, 2007, 2012).
The development policy agenda for Malawi is focussed on poverty reduction and is summarised in the
Malawi Growth and Development Strategy (MGDS), a five-year strategy. The first programme (MGDS
I) was launched in July 2007 and ran through to 2011. The second programme, the Malawi Growth and
Development Strategy II (MGDS II), was being implemented from 2011 and expired on 30 June 2016.
Some of the elements of MGDSII have been rolled over by the government. MGDS I and MGDS II
share the same themes but the latter has six themes instead of five. The themes or thematic areas are
Sustainable Economic Growth, Social Development, Social Support and Disaster Risk Management,
Infrastructure Development, Improved Governance and Cross-Cutting Issues. MGDS II includes a
cross-cutting theme that deals with gender imbalances, capacity development, and research and
development, HIV and AIDS, nutrition, environment, climate change, population and science and
technology (Government of Malawi, 2007, 2012).
Malawi’s national strategies are developed by the Government of Malawi in consultation with key
stakeholders, particularly the International Monetary Fund (IMF) and the World Bank, from whom the
country continues to receive technical support. In the 1980s and 1990s, Malawi embarked on economic
and structural reforms under the Structural Adjustment Policies (SAPS) of the World Bank and IMF.
The country was also granted debt relief under the 1996 Heavily Indebted Poor Countries (HIPC)
initiative which the IMF and World Bank implemented for a number of developing countries. Despite
these reforms and other targeted interventions, Malawi is still one of the least developed countries in the
world and its economy is still undiversified (Government of Malawi, 2012; Organisation for Economic
Co-operation and Development, 2008).
2
Available at http://www.fao.org/nr/water/aquastat/main/index.stm.
3
Stellenbosch University https://scholar.sun.ac.za
According to the Human Development Report (2015), the Human Development Index (HDI) value for
Malawi in 2014 was 0.445 and this puts the country on a ranking of 173 out of 188 countries and regions
recognised by the United Nations (UN). Table1.1 provides a review of Malawi’s progress between 1980
and 2014 in each of the three indicators that make up the HDI. The three dimensions of the HDI are a
long and healthy life (measured by life expectancy), access to knowledge (measured by expected and
mean years of schooling) and a decent standard of living (measured by Gross National Income (GNI))3.
Table 1.1: Malawi’s HDI trends based on consistent methodology and data
Year
1980
1985
1990
1995
2000
2005
2010
2011
2012
2013
2014
Life expectancy
at birth
44.8
45.1
43.8
43.5
44.1
48.3
56.9
58.6
60.1
61.5
62.8
Expected years of Mean years of
schooling
schooling
4.8
1.8
4.6
2.1
5.3
2.5
11.1
2.7
10.3
3.1
9.6
3.4
10.6
4.3
10.8
4.3
10.8
4.3
10.8
4.3
10.8
4.3
GNI per capita
(2011 PPP$)
705
643
612
556
613
601
722
732
717
726
747
HDI
Value
0.278
0.278
0.284
0.334
0.340
0.355
0.420
0.429
0.433
0.439
0.445
Source: Human Development Report (2015)
The table shows that from 1980 to 2014, life expectancy improved by 18 years, expected years of
schooling by 6 years, mean years of schooling by 2.5 years and GNI per capita by about 6%. Comparing
HDI changes over time using previously published reports would be erroneous due to the revisions and
updates that take place from time to time. However, the comparisons provided in the table are based on
consistent indicators and methodology developed by the United Nations Development Programme
(UNDP) for the purpose of analysis over time. Therefore, the figures in the tables show real changes in
values and Malawi’s actual progress over time (Human Development, 2015). In Figure 1.2 we show that
the annual GDP per capita growth for Malawi between 1961 and 2015 has been both volatile and dismal
in some of the years.
The average number of years of education is among individuals aged 25 years and older while the expected years
of schooling are the total number of years of schooling a child of school-entry age is expected to receive assuming
that the prevailing patterns of age-specific enrolment rates stay the same during the child's life.
3
4
Stellenbosch University https://scholar.sun.ac.za
Figure 1.2: Malawi’s GDP per capita annual growth rates (1961-2015)
Source: Own computation from World Bank country data4
Population growth, alongside nutrition, is another critical area identified under the MGDS II’s theme of
Social Development. Due to high levels of fertility rates, averaging 5.05 children per woman in 2008,
Malawi’s population is rising rapidly. During the last population and housing census in 2008, the
population was estimated at 13 million and projected at 17 million in 2015 (National Statistical Office,
2008). Malawi is now ranked at number 62 out of 196 countries by population with 149 people per km2.
This expanding population has implications for the labour market as well as the economy’s ability to
support jobs. Furthermore, in the MGDS II, labour and employment are identified as some of the subthemes for sustainable economic development. Within this framework, the focus is placed on the
creation of employment with a view to poverty reduction, incorporating gender specific issues in all
labour initiatives and interventions, reducing practices of discrimination in the labour market and
improved statistics on the labour market (Government of Malawi, 2012).
1.4
Problem statement
Our study is placed within the context of the critical issues facing Malawi as outlined in the national
development plan, namely the MGDS. The government identifies malnutrition as a crisis particularly
for the rural areas where most of the children suffer from high levels of stunting, wasting and
underweight. Agriculture naturally comes on the national development agenda because it is the major
source of employment for the Malawian population, as earlier stated. It is for this reason that the
government continues to invest in agriculture with the aim of improving food security which has
implications for nutritional status for the population. Land is identified as one of the critical resources
for agricultural development and the Government of Malawi recently embarked on the land reform
policy with the aim of improving production and incomes for the rural population which makes up 85%
of the population (World Bank, 2012). A related project is the Integrated Rural Development (IRD)
4
Available at http://www.worldbank.org/en/country/malawi.
5
Stellenbosch University https://scholar.sun.ac.za
programme which seeks to promote economic development in rural areas through the establishment of
rural growth centres and provision of social services. Other critical areas in the national development
plan include labour and employment, addressing gender imbalances and improving education
(Government of Malawi, 2012).
It is against this background that our study focuses on poverty, child-nutritional status, education,
employment and migration. The overarching theme in the thesis is poverty reduction and economic
development. This study’s contribution is towards an improved empirical understanding of the
aforementioned economic phenomena. It, therefore, can inform and have implications for policy and
development strategy.
1.5
National data sources
This study makes use of multiple data sources, all of which complement each other in the understanding
of the issues discussed in the thesis. The main agency for data collection in Malawi is the National
Statistical Office (NSO), responsible inter alia for conducting nationally representative surveys in the
country. Malawi has an Integrated Household Survey programme consisting of national censuses, the
Integrated Household Surveys (IHS) and Demographic Health Surveys (DHS). We identify these data
sets as the most suitable for this thesis.
Conducted every ten years since 1966, the national censuses provide a unique source of data for
understanding long-term patterns of economic phenomena in Malawi. Furthermore, they consist of
information at small geographical areas which is important for understanding issues and planning at low
levels of administration. Information collected in the censuses includes literacy, education, migration
and employment, among other population characteristics (National Statistical Office, 2008).
The IHSs are conducted every five years and collect information on consumption expenditure,
education, time use and labour, agriculture, health and child anthropometry. The first IHS was carried
out in 1990 and was called the Household Expenditure and Small Scale Economic Activities Survey
(HESSEA). Three rounds of integrated household surveys have been conducted after HESSEA, namely
IHS1 conducted in 1997/8, IHS2 conducted in 2004/5 and recently IHS3 conducted between March
2010 and March 2011 (National Statistical Office, 2012). In between the surveys are the Welfare
Monitoring Survey (WMS) normally conducted every year with the aim of tracking the living conditions
of people, identification of the vulnerable population groups and collecting indicators for monitoring
the attainment of national goals to which Malawi has committed itself, such as the MGDS and the
6
Stellenbosch University https://scholar.sun.ac.za
Millennium Development Goals (MDGs)5. So far, seven rounds of WMSs have been conducted in
Malawi, namely for 2005, 2006, 2007, 2008, 2009, 2011 and 2014 (National Statistical Office, 2015).
An alternative data source for the understanding of labour markets issues is the 2013 Malawi Labour
Force Survey (MLFS), which was conducted to provide a situational analysis of employment and
unemployment in Malawi. The previous labour force survey was conducted in 1983 but was not publicly
made available (National Statistical Office, 2012). Although it complements data from the IHSs and
WMSs, we do not use data from the labour force survey as it is stand-alone and, therefore, inadequate
for comparisons over time.
The DHSs are conducted every four years with main emphasis on health and nutrition, which is an area
of focus of study in this thesis. The first Malawi DHS was conducted in 1992 (National Statistical Office
& ICF Macro, 2011). Although the integrated household surveys also collect information on assets,
health and child-anthropometry, the DHSs provide a better source because of the detailed extent to which
these issues are dealt with in the questionnaire. In the context of Malawi, Verduzo-Gallo, Ecker, and
Pauw (2014) point out some serious inconsistencies and data quality issues in child-nutrition estimates
obtained using IHS data sets compared to anthropometric records from other nationally representative
data sets. Specifically, they find that while the 2010 DHS and the 2009 National Micronutrient Survey
(NMS) yield national child stunting rates of between 47% and 49%, estimates from the 2010/11 IHS3
suggest a prevalence rate of only about 30%. Similarly, estimates of child stunting levels based on the
IHS2 data in 2004/05 are about 9 percentage points below the incidence rates based on the 2004/05 DHS
and the 2006 Multiple Indicator Cluster Survey (MICS).
1.6
Thesis structure
This thesis is a consolidation of three related studies on Malawi. The introductory discussion provided
in the previous sections of this Chapter covers the issues related to all the three studies. In order to allow
for a detailed analysis, each of the studies forms a separate chapter. Similarly, the theoretical
underpinnings of the studies are also discussed separately in each of the main chapters. In Chapter 2, we
provide some theoretical considerations on poverty and inequality measurement, stochastic dominance
analysis and measurement of pro-poor growth. In Chapter 3, two main competing groups of theories for
explaining labour market outcomes are discussed, namely the traditional neoclassical model of labour
supply and the segmented labour market hypothesis. Examples of the theories discussed are the human
capital theory, Roy’s (1951) two sector model and the Harris-Todaro (1970) model of migration, among
others. Furthermore, we discuss the different theories or explanations as to what constitutes or gives rise
5
MDGs have now replaced with Sustainable Development Goals (SDGs) since 2015.
7
Stellenbosch University https://scholar.sun.ac.za
to informal labour markets. In Chapter 4, we discuss the theoretical perspectives of migration, broadly
grouped under either the disequilibrium or equilibrium perspectives. The specific theories discussed in
the chapter are the gravity models of migration, the human capital theory and the ‘spatial job-search
models’.
Chapter 2 contains spatial and temporal comparisons of multidimensional poverty and inequality in
Malawi based on two non-monetary dimensions of welfare, namely an asset index and child nutritional
status. Data for this study are drawn from the DHSs. Through this paper, we attempt to present Malawi’s
profile of poverty and inequality, including the progress made over time. We also show the extent to
which the observed changes over time have been pro-poor. Child nutritional status is identified as a
problem requiring attention in the country’s national development agenda (MGDS). The first part of the
chapter deals with the derivation of our two dimensions of poverty. The second part conducts robust
comparisons of economic welfare and an econometric analysis of underlying possible reasons behind
the observed changes. The third and final part analyses pro-poor changes in poverty over time.
In Chapter 3, we look at the role of education in poverty reduction in Malawi by using data from the
Malawi Integrated Household Panel Survey (IHPS). The linkage between education and poverty
reduction is identified through the labour market. The argument made here is twofold, namely that
education improves an individual’s chances of getting employment and that education positively
impacts on earnings. This partly justifies why governments invest in education, which is acknowledged
to have positive externalities on households and communities (e.g. Basu & Foster, 1998). In the context
of Malawi, primary education was universally made free in 1994, which is before the MDGs, while
university education is either directly subsidised or students are granted study loans. The overall
objective of this chapter is to estimate returns to education. Our analysis distinguishes between wage
employment and self-employment activities (household enterprises), which make up a large percentage
of total employment in Malawi. With respect to self-employment, the returns to education are calculated
at the household level using the maximum level of education in the household. Prior to the analysis, we
also conduct some consistency checks in the data to ensure data quality and meaningful comparability
over time.
The theme of employment continues through Chapter 4, where emphasis is now placed on gender related
issues and the role of spatial effects in employment. The data used are from the national censuses.
Segregation of results by gender is important because women form a large percentage of the labour force
in Malawi, with the majority of them engaged in the agricultural sector. Specifically, Chapter 4 applies
spatial panel data econometric techniques to the study of migration and employment in Malawi. It is
widely recognised in the literature that both geography (space) and time are important to the
understanding of economic phenomena. However, only few studies incorporate spatiotemporal analysis
8
Stellenbosch University https://scholar.sun.ac.za
and within the context of Malawi, this study is the first-time attempt. First, we first match geographical
codes so that they are consistent over time. Once this is done, we analyse long-term patterns of migration
and employment in Malawi. In 2004, the Government of Malawi introduced land reform with the aim
of increasing the incomes of poor rural families in four Southern region districts of the country. This
policy was aimed at poverty reduction through increase of incomes and improvement in food security
for the participating families. To be specific, willing individuals purchased agricultural land from willing
sellers and resettled in the new areas. Therefore, the second part of this study is dedicated to the analysis
of the effects that the land reform policy had on migration and employment. The findings from this study
are important because of the future joint plans by the World Bank and Malawi Government to scale up
the project to the rest of Malawi.
Finally, Chapter 5 concludes the thesis. The chapter provides a discussion on how the thesis addresses
the research questions developed in each of the studies. We also look at the significance of the thesis in
terms of the contributions made to research, implications of the research, its limitations and suggestions
for future study.
9
Stellenbosch University https://scholar.sun.ac.za
Chapter 2
Measuring poverty and pro-poor growth in Malawi
2.1
Introduction
Poverty and inequality remain big concerns in Malawi, a very poor country. Deprivation exists in a
number of dimensions such as education, consumption, child nutritional status and assets. Based on
household per capita consumption estimates from the Third Household Integrated Survey (IHS3), about
51% of households in Malawi are poor. In addition, the Gini coefficient shows that inequality has
increased over the past five survey years from 0.390 in 2005 to 0.452 in 2011. Estimates based on the
2010 Malawi Demographic Health Survey (MDHS) indicate that the incidence of child stunting in
Malawi stands at 47% (National Statistical Office & ICF Macro, 2011).
Following the works of Sen (1985, 1987), a number of approaches have been developed to measure
multidimensional poverty. A multidimensional view of poverty considers more than one aspect of
deprivation. Conventionally and for a long time, poverty has been looked at in terms of either income
or consumption. However, this view of poverty has been criticised for ignoring other important
dimensions of well-being such as health, education, empowerment and freedom of association. Based
on the literature, one can group the existing approaches to the measurement of multidimensional poverty
into three alternatives.
The first approach aggregates a number of dimensions of poverty such as life expectancy, literacy and
Gross Domestic Product (GDP) into a single one-dimensional index. Examples include the Human
Development Index (HDI) and the Multidimensional Poverty Index (MPI). Ravallion (2011) raises
questions as to whether such single-one dimensional indices (which he refers to as “mashup” indices)
are sufficient for poverty measurements as opposed to developing a set of poverty indicators that are
relevant within a particular setting. Specifically, Ravallion (2011) criticises the composite indices for
not being so useful for sound development policy making because they essentially ‘collapse’ important
dimensions into a single index which is difficult to interpret. Another criticism of the multidimensional
indices raised by Ravallion (2011) is how weights are applied in the construction of multidimensional
indices. Specifically, in as much as it is recognised that poverty is multidimensional, weights need not
be determined by the poverty analyst but should rather be consistent with the choices made by the poor
people. The second approach considers two or more dimensions of poverty such as income, education,
health, etc. but analyses each dimension independently, without taking into account the possible
correlations which may exist between the dimensions (e.g., Sahn & Stifel, 2003; Mussa, 2013). The
third approach also considers two or more dimensions but unlike the second approach takes into account
the potential interrelationships among dimensions (e.g., Gondwe, 2011; Duclos, Sahn, & Younger,
2006; Batana, 2008; Batana & Duclos, 2008). In this approach, a poverty line is set for each of the
10
Stellenbosch University https://scholar.sun.ac.za
dimensions and then a decision is made as to whether an individual is to be considered poor if deprived
in just one, some or all of the chosen indicators. In the literature, poverty has been found to be higher in
distributions with higher correlations between the measures of well-being than those with lower
correlations. Therefore, it is possible that univariate and multivariate analyses of poverty produce
different rankings of poverty between distributions (Bourguignon & Chakravarty, 2003).
The third approach, therefore, looks at poverty measures that make for possible substitutions and
complementarities between the levels of dimensions. Assuming two dimensions, for the substitutability
assumption, we expect that the more someone has of one dimension of poverty, the less is overall
poverty deemed to be reduced if their value of the other dimension is increased. On the other hand, for
the complementarity assumption, increasing one dimension would reduce overall poverty. For example,
transferring education from the poorly nourished to the better nourished would reduce overall poverty
because better-nourished children learn better (Bourguignon & Chakravarty, 2003).
Most previous studies on poverty in Malawi concentrated on unidimensional poverty analysis (e.g.,
Murkherjee & Benson, 2003; Bokosi, 2006). Mussa (2013) considered three dimensions of poverty and
inequality in Malawi, namely household per capita consumption, education and health. However, the
study looked at the three dimensions one-at-a-time (or independently) without taking into account the
correlations that exist between the dimensions of well-being. Gondwe (2011) did account for the
possible correlations that exist between the dimensions of poverty. Two dimensions were used,
household per capita consumption and education.
This study conducts spatial comparisons of multidimensional poverty and inequality using two nonmonetary dimensions, namely an asset index and child nutritional status. We look at the two dimensions
separately. It is the first time attempt to apply the asset index approach to the measurement of poverty
in Malawi and uses a more recent DHS data set compared to the 2004 DHS data used by Alkire and
Santos (2010). Also, the study conducts pro-poor growth analysis in the selected dimensions of living
standards over two decades, from 1992 to 2010. As pointed out in Grosse et al. (2008), pro-poor growth
analysis has recently become important to researchers and policy makers particularly with respect to
monitoring progress towards the attainment of MDGs (now SDGs). However, the current efforts of propoor growth analysis have largely focussed on monetary dimensions of poverty thereby ignoring the
non-monetary aspects of well-being. This study tries to reduce this shortcoming in the literature of the
current pro-poor growth analysis by using two non-monetary dimensions of poverty, namely assets and
nutritional status which are also central to the attainment of SDGs.
11
Stellenbosch University https://scholar.sun.ac.za
Household income, consumption expenditures and assets are the three main indicators of economic
status that exist in the literature. The use of assets has gained popularity in recent years (Filmer &
Pritchett, 1998; Sahn & Stifel, 2000; Booysen, Van der Berg, Burger, Von Maltiz, & Du Rand, 2008).
Using data from Demographic and Health Surveys (DHS), an index is computed from a number of asset
variables and this forms the basis for ranking households by their long-run socio-economic status.
Furthermore, assets as a measure of economic status have been found to have more advantages than
both income and expenditure. We discuss four main advantages. Firstly, unlike assets, both income and
household expenditures are associated with measurement problems. For example, many respondents
hide their incomes and only provide income figures in ranges. Income and consumption are also
associated with seasonality and, therefore, unreliable as long term measures of status. Secondly,
unearned income such as interest on loans, gambling, etc. is not reported. Income on home production
and self-employment activities is usually excluded just as expenditure on non-routine goods and
services. Thirdly, it is usually the income or consumption expenditure of the respondent (in most cases
household head) that is recorded as opposed to the rest of the household members. Fourthly, data
collected on income and expenditure is usually over the past month, week or day thereby raising
questions as to what period of time should be covered (Rutstein & Johnson, 2004).
Based on the foregoing discussion, wealth is not only said to represent a more permanent status than
income and expenditure but also more easily measured with only a single respondent required in most
cases. In addition, the collection of asset information requires fewer questions than in income and
expenditure surveys (Rutstein & Johnson, 2004).
The study achieves five objectives. First, it presents spatial poverty and inequality comparisons in assets
and child nutritional status across population groups (areas, regions and sex of household head) in
Malawi. Related to the first objective, we conduct poverty and inequality decompositions to see the
relative contributions by the respective population groups or distributions. Second, it establishes a robust
ranking of poverty and inequality across the groups that are compared. Third, it identifies the factors
associated with asset poverty and child nutritional status in Malawi. Fourth, it tracks the incidences of
asset poverty and child malnutrition in Malawi over the past two decades using a series of cross-sectional
data sets. Finally, it establishes if the observed changes in living standards and child nutritional status
over time have been pro-poor, absolutely and relatively speaking. Relative pro-poor changes in welfare
have implications for inequality since poor people benefit more from the changes than the rich.
2.2
Theoretical considerations in poverty measurement
Three conditions are necessary for poverty measurement, namely a set of welfare indicators, poverty
line and poverty measure (World Bank, 2004). The first condition is the choice of the welfare indicator
which can be grouped into two, namely monetary (e.g. consumption or income) and non-monetary
12
Stellenbosch University https://scholar.sun.ac.za
dimensions (e.g. assets or child-nutritional status). There is debate in the literature regarding which is a
better indicator of welfare. The second condition is the choice of the poverty which can be looked at as
the threshold separating the poor from the non-poor with the former falling below it. There exist two
definitions of the poverty line. On the one hand, we have the absolute poverty line which is set for a
particular group without reference to other members in the population. This poverty line is determined
with respect to the basic needs needed for a living by a household or individual. On the other hand, we
have the relative poverty line set with reference to the population, say at 60% of the average percapita
consumption. The choice of which poverty line to use depends on the population we are studying. For
poor countries, an absolute poverty line seems appropriate since emphasis is to ensure that the basic
needs of the population are met. However, for richer countries that have met the basic needs a relative
poverty line would make sense.
Having chosen the measure of welfare and poverty line, the third step is the choice of the poverty
measure to use (Haughton & Khandker, 2009). Several poverty measures are available in the literature
such as the Watts index and the Sen-Shorrocks-Thon index, among others. A good poverty measure is
supposed to satisfy some basic axioms to be considered reliable (see for example Sen, 1976; Kakwani,
1980; Foster, Greer, & Thorbecke, 1984). In this study, we use the Foster-Greer-Thorbecke (FGT)
measures because of the decomposability property which they possess in addition to other favourable
characteristics. We consider three FGT indices, namely the headcount index, poverty gap index and the
squared poverty gap or poverty severity index (Foster, Greer, & Thorbecke, 1984). The FGT measures
are given as:
1
P z ,
N
z yi
I yi z
z
i 1
N
(2.1)
Where: yi , z, N , are the welfare indicator, poverty line, population size and measure of poverty
aversion, respectively. I . is an indicator function that takes on a value of 1 if the expression is true
and 0 otherwise. When 0 , the result is the poverty headcount index which is a measure of the
proportion of the population that is poor. For 1 we have the poverty gap index which indicates the
extent to which individuals on average fall below the poverty line and expresses it as a percentage of
the poverty line. Finally, when 2 , we have the squared gap index which averages the squares of
the poverty gaps relative to the poverty line.
Although the headcount index P0 is easy to understand and measure, it does not indicate how poor the
poor are. Unlike the headcount index, the poverty gap index P1 measures the extent to which
13
Stellenbosch University https://scholar.sun.ac.za
individuals fall below the poverty line (the poverty gaps) as a proportion of the poverty line. The main
limitation of the poverty gap index is that it does not reflect changes in inequality among the poor. By
averaging the squares of the poverty gaps relative to the poverty line, the squared poverty gap index
(also called the poverty severity index, P2 ) is able to show the changes in inequality among the poor.
2.3
Inequality measurement
Unlike poverty analysis which only focuses on the poor individuals or households, inequality is defined
over the entire population and takes into account both the rich and the poor. Here we only consider the
Gini and Theil indices measures due to their desirable properties. The other measures of inequality
discussed in the literature include Decile Dispersion Ratio, Atkinson’s inequality measures and
Coefficient of Variation (see Haughton & Khandker, 2009; Duclos & Araar, 2009, for discussion).
The Theil indices are advantageous because they are additive across different population subgroups and
enable us to see between and with group inequalities. On the other hand, the Gini is easy to understand
and has a desirable graphical representation. It is for this reason that the Gini is preferred in most studies.
The Gini coefficient varies between 0 (representing equal distribution) and 1 (representing a complete
inequality). On the other hand, Theil index values vary between zero and infinity, which reflect complete
equality and inequality, respectively. Graphically, the Gini coefficient is calculated as the area above
the curve but below the line of perfect equality divided by the total area below the line of perfect equality.
Apart from measuring the level of inequality, the Lorenz curve is also be used to test for inequality
dominance between two distributions.
The Gini coefficient is calculated by the following formula:
N
G 1 xi xi 1 yi yi 1
(2.2)
i 1
Where: G refers to the Gini coefficient; x i is the cumulative proportion of the population (represented
on the x-axis) and y i be the cumulative proportion of the welfare indicator (in our case child-nutritional
status and asset index).
If there are N equal intervals on the x-axis, equation (2.2) collapses to:
Gini 1
1
N
N
y
i 1
i
y i 1
(2.3)
14
Stellenbosch University https://scholar.sun.ac.za
The Gini satisfies four main properties and these are mean independence, population size independence
and the Pigou-Dalton Transfer sensitivity. However, the Gini does not satisfy two important
characteristics, namely decomposability by population groups, dimensions or sources and statistical
testability over time although this is less problematic now due to the fact that confidence intervals can
typically be obtained through the use of bootstrap techniques. The Theil index measures satisfy all of
the six properties.
The Theil indices are part of a larger family of measures referred to as the Generalised Entropy (GE)
class of indices. The general specification of the GE measures is given as:
1 1 N yi
1
GE
1 N i 1 y
(2.4)
Where: y is the selected welfare indicator or dimension and y is its average or mean. The parameter
gives the weight given to distances between values of a given indicator at different parts of the
distribution, and can take any real value. The GE index is more sensitive to changes in the lower tail of
the distribution for lower values of , and for higher values, GE is more sensitive to changes occurring
at the upper tail. When 0 , we have the Theil-L index, also called mean the mean log deviation
measure, and when 1 , the result is the Theil-T index.
2.4
Stochastic dominance analysis
Dominance tests are necessary because poverty or inequality ranking can be reversed by different
choices of poverty lines, measures, aggregation procedures and samples. Stochastic dominance analysis
seeks to achieve non-ambiguous ranking in terms of welfare and inequality between any two
distributions (Araar, 2006; Davidson & Duclos, 2000).
First, we discuss poverty dominance. Assuming two distributions, A and B, for our dimensions of
poverty, namely asset index and child nutritional status, F A and F B will be the cumulative density
functions (CDFs). Distribution B is said to dominate distribution A stochastically at first order if, for
any argument y , F A y F B y . In terms of poverty, this means that there is (weakly) more poverty in
distribution A than there is in B. Higher orders of stochastic dominance are obtained through repeated
integrals of the CDF of each distribution (Davidson & Duclos, 2000). Generally, we have:
y
D 1 y F y , D s 1 y D s z dz, for s 1,2,3,...
0
15
(2.5)
Stellenbosch University https://scholar.sun.ac.za
Where:
D 1 is the CDF of the distribution under study;
D 2 y is the integral of D 1 from 0 to y;
D 3 y is the integral of D 2 from 0 to y, and so on.
By definition, distribution B dominates A at order s if D A y DB y for all arguments y [0, z max ]
s
s
. The lower limit of 0 represents the lowest value of the welfare indicator while zmax is the maximum
acceptable poverty line for each welfare indicator. First-order dominance implies dominance at all
higher orders (Davidson & Duclos, 2000). Where first order dominance is not established, we proceed
to higher levels but stop at third order dominance as is the practice in the literature (e.g., Mussa, 2013).
Lorenz curves are the widely used approach to testing stochastic dominance in inequality (Araar, 2006).
A given distribution is said to Lorenz dominate another distribution if the Lorenz curve of the first
distribution lies everywhere above that of the latter. We then say that there is less inequality in the
distribution with the higher curve than in that with a lower curve. Simply put, inequality is higher in A
than in B if L B p is everywhere above L A p . Distribution B dominates distribution A in inequality,
with the second order, if
L A p L B p p 0,1
(2.6)
Where p is the percentile. The Lorenz curve for the percentile p can be defined as follows:
p
Qq dq 1 Qq dq
L p
Qq dq
0
1
p
(2.7)
0
0
L p is the cumulative proportion of the welfare indicator (asset index or child-nutritional status) held
by a cumulative percentage p of the population, when individuals are ordered in increasing asset or
the bottom p
p
Qq dq gives the sum of the values of the welfare indicator of
proportion (the poorest 100 p % ) of the population. Q q dq gives the sum the welfare
child-nutritional values. The integral
0
1
0
indicator values of all (Duclos & Araar, 2006).
The inequality dominance tests used in this study are based on Araar’s (2006) theoretical developments.
Specifically, generalised Lorenz dominance tests are used, and these turn out to be the same thing as
second-order stochastic poverty dominance (Araar & Duclos, 2013).
16
Stellenbosch University https://scholar.sun.ac.za
2.5
Poverty and inequality decomposition
As indicated in Sections 2.2 and 2.3, the FGT and GE indices are decomposable by population groups.
In this study, we follow decompositions based on Araar and Duclos (2013). The decomposition of the
FGT index enables us to determine the absolute or relative contribution of each group such as area,
region or etc. It takes the following form:
G
Pˆ z ; ˆ g Pˆ z ; ; g
(2.8)
g 1
Where: G refers to the total number of population groups; Pˆ z ; ; g and ˆg are the estimated FGT
ˆ g Pˆ z; ; g
are the estimated
index and population share of subgroup g ; ˆ g Pˆ z; ; g and
Pˆ z;
absolute and relative contributions to total poverty by subgroup g .
GE decomposition takes the following form:
K
ˆ k ˆ
I k ; I
Iˆ ˆk
k 1
ˆ
(2.9)
Where: K refers to the total number of population groups; ˆk is population share of subgroup k ; ̂ k
is the mean of the selected indicator subgroup k ; Iˆk ; is the inequality within subgroup k ; I is
population inequality if each individual in subgroup k is given the mean for the poverty indicator of
subgroup k , ̂ k .
2.6
Pro-poor growth analysis
In the literature, outcomes of pro-poor growth between any two given periods are analysed by
calculating the growth rate (g) and five different pro-poor indices (Duclos & Verdier-Chouchane, 2010).
The first three of these indices are measures of absolute pro-poorness and they are: the Ravallion and
Chen (2003) index, the Kwakwani and Pernia (2000) index and the PEGR index. The other two indices
namely, the Ravallion and Chen (2003) index minus (g) and the PEGR index minus (g) are indices of
relative pro-poor growth.
There exist two different approaches to the definition of pro-poor growth, namely a relative and an
absolute approach. Growth is defined as pro-poor in the absolute sense if it reduces absolute poverty.
17
Stellenbosch University https://scholar.sun.ac.za
Using the relative approach, growth is pro-poor if reduces inequality and relative poverty. In this sense,
the poor proportionately benefit more from growth than the non-poor.
If the growth rate and the Ravallion and Chen (2003), the Kwakwani and Pernia (2000) and the Poverty
Equivalent Growth Rate (PEGR) indices are positive, there is absolute pro-poor growth from one period
to another. When g is positive and the Kwakwani and Pernia (2000) is negative or when g is negative
and the Kwakwani and Pernia (2000) index is positive, then the distributive change has increased
absolute poverty. Growth is said to be anti-poor when this is the case. When the Ravallion and Chen
(2003) minus g and the PEGR minus g are positive, the distributive change is considered to be relatively
pro-poor. A similar conclusion is arrived at if the Kakwani and Pernia (2000)'s index is larger than 1. In
this case, growth among the poor is higher than average growth. The poor have, therefore, been
favourably affected by the change.
In order to understand Ravallion and Chen (2003)'s growth incidence curves, we, first of all, explain
what a “quantile” is. Suppose there are n incomes in a given distribution ranked from the lowest to the
highest. A quantile of a given population is given by the income level that is found at a particular rank
in that distribution. The rank of the level of income yi will be given by i / n . Growth incidence in the
population can be understood by comparing quantile curves before and after a change in a distribution
has taken place. Let the pre-change distribution be given by y A and the post-change distribution be given
by y B , each of equal size n . We can build quantile curves for each of these distributions; these are given
by the incomes y iA and y iB found at different ranks i / n . We can then assess the incidence of growth at
any particular rank i / n by comparing the quantile curves at the point i / n . The absolute value change
is given by y iB y iA . The proportional change is given by
y iB y iA
y iA
.
The Ravallion and Chen (2003) growth incidence curve is a plot of the proportional change against all
possible values of ranks i / n . The incidence curve shows the rates of growth for various ranks in the
distribution. Absolute pro-poorness of growth is obtained when the absolute value change is everywhere
positive for the range of ranks over which the initially poor individuals or households are located.
Relative pro-poorness of growth is obtained when the growth incidence curve is everywhere above the
proportional change in the mean income.
18
Stellenbosch University https://scholar.sun.ac.za
The Kakwani and Pernia (2000) index compares the actual poverty outcome of a distributive change to
the outcome that would have been observed if the change had been distribution-neutral (Kakwani &
Pernia, 2000). Two main distribution-neutrality criteria are provided. The first one assumes that
everyone's income has changed by the same absolute amount while the second one considers that
everyone's income has changed by the same proportional amount. There exist several views on what
that proportion should be but the most common one is the proportional change in average income.
Suppose P A and P B be the actual pre- and post-change poverty levels, and P BN let be post change
PA PB
is the ratio of the actual change in poverty to
poverty under distribution neutrality. Then,
P A P BN
the change that would have been observed under distribution neutrality. Several poverty indices can be
chosen for P . In the main text, P is specified as the headcount ratio. Several scenarios of distribution
neutrality can also serve to specify BN . Let for instance A y 1A , y 2A ,..., y nA and B y 1B , y 2B ,..., y nB .
Kakwani and Pernia (2000)'s index uses the following definition for BN :
BN B y1A , B y 2A , B y nA
A
A
A
(2.10)
It says that a change is distribution neutral if incomes change in proportion to the proportional change
in average income. This index thus gives the ratio of the observed change in poverty to the change that
would have been observed under constant inequality.
The Poverty Equivalent Growth Rate (PEGR) index, also called the Kakwani, Khandker and Son (2003)
index, assesses the pro-poorness of growth by calculating “poverty equivalent growth rates”. PEGR is
the growth rate that would have resulted in the same observed level of poverty change if the distribution
of income shares had not changed. PEGR can be thought of as the counterfactual distribution of income
BN 1 PEGR y 1A , 1 PEGR y 2A ,..., 1 PEGR y nA as giving the same final level of poverty as
A
the one that is actually observed. When the growth rate PEGR is applied to all of the initial income y ,
B
poverty thus equals poverty with the distribution of y .We, therefore, have PB PBN and thus that
PA PB PA PBN
If PEGR is greater than 0, the distributive change is judged to be absolutely pro-poor by this approach.
This is the case if and only if PA PB 0 .
19
Stellenbosch University https://scholar.sun.ac.za
Let g
B A
and it is thus the actual rate of growth in average income. If income shares remain
A
constant in the movement from A to B , then we must have that y iB 1 g y iA for all i . Since
PA PBN , with constant income shares, it must be that PEGR g . The poverty equivalent growth rate
is, therefore, just the usual growth rate if inequality has remained unchanged. Movements in inequality
will, however, create a divergence between the poverty equivalent growth rate and the usual growth
rate. The greater the adverse effects of inequality on the poor, the greater the value of PB , and therefore
the lower the value of PEGR .
The difference, PEGR g , can, therefore, help assess whether the distributive change has affected the
income shares of the poor. If PEGR g is negative, growth among the poor is lower than average
growth, and the income shares of the poor have therefore been adversely affected by the change. The
converse is true when PEGR g is positive.
Although pro-poor growth analysis based on non-monetary dimensions has been shown to yield
important results, there exist a number of potential problems of extending pro-poor growth analysis to
non-monetary dimensions of welfare. Grosse et al. (2008) provide a good discussion of some of these
potential limitations. In the context of this study, the main challenge relates to comparing relative
changes in the asset index score or HAZ score on a linear scale. For example, in our data sets, HAZ
scores range from -6 to 6 which clearly includes negative, 0 and positive values. This raises the question
of how to compare a relative change from a score of -6 in 1992 to -5 in 2010 with an improvement from
5 to 6 over the same period. We overcome this challenge by transforming the z-scores into percentiles
in the distribution. Similarly, we convert the asset index scores so that they are all positive by adding to
the asset index a value slightly higher than the absolute value of the most negative number. The
conversions and transformations of the HAZ and asset index scores are further discussed in Section 2.7.
2.7
Data
Malawi has participated in four nationally representative Demographic Health Surveys (DHSs). These
surveys have provided up-to-date information on living conditions and health programmes in Malawi.
The surveys were conducted by the National Statistical Office (NSO) and the Community Health
Sciences Unit (CHSU) during different times of the calendar year. Traditionally, DHSs consist of data
sets relating to the household, men, women and children. This study makes use of the household and
children’s data sets only. Child nutritional status is particularly important because it affects their growth
and development, which has a direct link to their future health status as adult men and women. The
household data set contains information on assets which are found to be more reliable than consumption
20
Stellenbosch University https://scholar.sun.ac.za
expenditure and income as a measure of the long-term living standards for households, as we explain
shortly. Table 2.1 provides a summary of the DHS data sets used in the study. Detailed tables showing
the distribution of households and children by the different subgroups are respectively presented in
Tables A1 and A2 in the appendices.
Table 2.1: Summary of Malawi DHS data sets used
Year
Survey period
No. of households
No. of children (0-59 months)
1992
September-November 1992
5,323
3,353
2000
July-November 2000
14,213
9,753
2004
October 2004 -January 2005
13,664
8,707
2010
June-November 2010
24,825
4,801
Source: Own computation from MDHS data
From the data sets, we derive two non-monetary measures of welfare, namely household asset index and
child nutritional status in terms of anthropometric indices of children. The number of children indicated
in the table excludes children whose age, height and weight measurements are missing. Poverty and
inequality measurement is based on the 2010 data sets only. On the other hand, pro-poor growth analysis
is done using all the DHS survey periods from 1992 to 2010.
As earlier discussed, in the literature the use of assets as a measure of economic status has been found
to be advantageous over household income and consumption. Wealth6 is not only said to represent a
more permanent status than income and expenditure but is also more easily measured with only a single
respondent required in most cases (Rutstein & Johnson, 2004).
Since wealth cannot be directly observed, asset variables have to be identified to proxy wealth. There
exists no best approach for selecting which asset variables to use (Montgomery et al., 2000). Although
the choice of asset variables has varied by author, generally all assets and utility services reflecting the
economic status of a household need to be included. This broader criterion, as opposed to selecting a
few assets, is preferred because using a large number of variables potentially prevents a situation where
households are concentrated on certain index scores (Rutstein & Johnson, 2004). In adopting the broad
criterion, the study uses 31 asset variables and access to utilities including household assets, means of
transport, source of lighting, ownership of livestock, agricultural land, having separate room for kitchen,
rooms for sleeping, source of drinking water, type of toilet facility, type of material for the main floor,
main wall material, main roof material and type of cooking fuel.
6
In the literature, wealth and asset index are loosely used synonymously because the assets owned by a household
represent its asset wealth.
21
Stellenbosch University https://scholar.sun.ac.za
The derivation of the asset index requires that the indicator variables for asset ownership7 be captured
or transformed into binary form, e.g. 1 for “yes” if a household owns a given asset and 0 for “no” if a
household does not own the asset. It is, however, not necessary to transform variables that are already
categorical such as the source of drinking water, type of toilet facility, floor, wall, roof material and
source of cooking fuel. Stata 13.1, the software used, recognises the categorisation automatically even
in cases where there are more than two categories. After categorisation in the manner explained, the
mca command was applied to the asset variables. The index is weighted by household size and sample
weight. Principal components analysis (PCA) and factor analysis (FA) were used as robustness checks.
Results from MCA showed that assets and living conditions associated with good higher economic
status contribute negatively to the index while those associated with low economic status contribute
positively to the index. For example, ownership of a durable asset (e.g. radio, television, refrigerator,
etc.) and having piped water, flush toilet, carpet floor and good source of lighting and cooking fuel,
among others, contribute negatively to the asset index. On the other hand, not owning a durable asset
and living in poor conditions contribute positively to the asset index. The negative signs imply that the
index generated from MCA is not an asset index but rather a poverty index. This implies that if quintiles
were to be created from the index, the poorest households would be in the fifth quintile and the richest
households in the first quintile. This complicates the interpretation of the index and the suggested
solution in the literature is to generate an asset index by multiplying the poverty index by (-1). When
the transformation is done, the richest households would now be in the fifth quintile and the poorest
households in the first quintile, giving the normal interpretation of an asset index. This approach is
motivated by Greenacre (2007) and adopted in the other studies (e.g. Da Maia, 2012). In this case, the
poverty and asset indices are negative opposites of each other.
Asset and poverty indices typically contain negative values which are unsuitable for poverty and
inequality measurement. Following the literature, we adjusted the asset index by adding to it a value
slightly higher than the absolute value of the most negative number (e.g., Da Maia, 2012). In our case,
since -1.056288 was the minimum value, we added 1.056289 to the asset index. This results in a shift
of the distribution of the asset index to the right. The resultant asset index, which we call the adjusted
asset index, is what is used for poverty and inequality analysis. Some selected summary statistics for the
poverty and asset indices are provided in Table 2.2.
7
In this study, asset ownership is used to mean ownership of private assets as well as access to public services.
22
Stellenbosch University https://scholar.sun.ac.za
Table 2.2: Descriptive statistics for the poverty and asset indices
Description
Poverty index Asset index Adjusted asset index
Percentiles
1%
-3.65
-0.92
0.14
5%
-1.84
-0.85
0.20
10%
-1.00
-0.80
0.26
25%
-0.08
-0.67
0.38
50%
0.43
-0.43
0.63
75%
0.67
0.08
1.13
90%
0.80
1.00
2.06
95%
0.85
1.84
2.90
99%
0.92
3.65
4.71
Statistic
Minimum
-5.29
-1.06
0.00
Maximum
1.06
5.29
6.34
Mean
0.11
-0.11
0.94
Standard deviation
0.91
0.91
0.91
Variance
0.82
0.82
0.82
Skewness
-2.35
2.35
2.35
Kurtosis
9.36
9.36
9.36
Observations
24,825
24,825
24,825
Source: Own computation from MDHS 2010
In Figure 2.1, we show the adjusted and unadjusted asset indices in addition to some kernel density plots
across population sub-groups. The results show that urban areas have a higher average of asset index
scores compared to rural areas. Amongst regions, the Northern region has the highest asset score values
seconded by the Central region and finally the Southern region. A similar trend is observed amongst
Rural north, Rural centre and Rural south. On the other hand, Urban centre and Urban south seem to be
doing better than Urban north.
23
Stellenbosch University https://scholar.sun.ac.za
0
0
.5
.5
Density
Density
1
1
1.5
1.5
Figure 2.1: Adjusted and unadjusted asset indices by population subgroups
-2
0
2
MCA asset index
6
0
Unadjusted
2
MCA asset index
4
Urban areas
6
Rural areas
0
0
.5
.5
Density
Density
1
1
1.5
1.5
Adjusted
4
0
2
4
6
Central region
0
Southern region
2
4
MCA asset index
Rural north
Rural centre
6
Rural south
0
0
.1
.5
Density
Density
.2
1
.3
1.5
.4
Northern region
MCA asset index
0
2
Urban north
MCA asset index
Urban centre
4
6
0
Urban south
2
MCA asset index
Male head
4
6
Female head
Source: Own computation from MDHS 2010
With respect to our second poverty measure, i.e. child- nutritional status, we calculate three
anthropometric indices or z-scores, namely height-for-age (HAZ), weight-for-age (WAZ) and weightfor-height (WHZ) for children aged between 0 and 59 months. We discuss each of these three measures
separately in the next few paragraphs.
Firstly, HAZ measures stunted growth and reflects cumulative linear growth and failure to receive
adequate nutrition over a long period. HAZ, therefore, indicates the long-term effects of malnutrition in
24
Stellenbosch University https://scholar.sun.ac.za
a population. Secondly, WHZ is a measure body mass in relation to body height or length and reflects
the current nutritional status. It can be used to describe body wasting which represents the failure to
receive adequate nutrition in the period immediately preceding the survey. Wasting may result from
inadequate food intake or recent episodes of illness causing loss of weight and the onset of malnutrition.
Finally, WAZ (a composite index of HAZ and WHZ) gives the overall malnutrition level and takes into
account both acute and chronic malnutrition. It measures the state of body weight.
The z-scores are calculated using the new 2006 World Health Organisation (WHO) child growth
standards based on the Multicentre Growth Reference Study done on a ‘healthy’ sample size of 8,440
children drawn from six countries across the world. The analysis is done using the zscore 06 module8.
The z-scores express normal and abnormal departures of an individual child's height or weight from the
average height or weight of comparable children of the same sex and age in the standard reference
population. According to WHO (2006), z-scores below -2 standard deviations (SD) and -3 SD from the
median of the reference population indicate malnourishment and extreme malnourishment, respectively.
Normal children have z-scores of greater than or equal to -2 SD (>=-2SD). Figure 2.2 provides kernel
density estimates for HAZ, WAZ and WHZ. Summary descriptive statistics are provided in Table 2.3.
.3
.2
.1
0
Cumulative proportion of children
.4
Figure 2.2: Distribution of anthropometric Z-scores for HAZ, WAZ and WHZ
-6
-4
-3
-2
0
2
Anthropometric Z-scores
HAZ
WAZ
4
6
WHZ
Source: Own computation from MDHS 2010
Levels of child malnutrition seem to depend on the choice of measure. HAZ shows the highest levels of
child malnutrition, whereas WAZ and especially WHZ appear to reflect lower levels. Child malnutrition
rates for the sample (i.e. scores more than 2 standard deviations below the mean for the standard
8
zscore06: Stata command for the calculation of anthropometric z-scores using the 2006 WHO child growth
standards; http://www.ifpri.org/staffprofile/jefleroy.
25
Stellenbosch University https://scholar.sun.ac.za
reference population) stand at 46%, 14% and 4% as based on HAZ, WAZ and WHZ, respectively.
Extreme child malnutrition (more than 3 standard deviations below the reference population mean)
stands at 19%, 3% and 2% based on HAZ, WAZ and WHZ, respectively.
Table 2.3: Child malnutrition rates by population groups
Description
HAZ
<-2SD
< -3SD
Age
0-23
38.60% 18.20%
24-59
50.70% 19.80%
Sex
Male
49.00% 22.20%
Female
42.20% 16.20%
Areas
Urban
39.40% 15.30%
Rural
46.60% 19.80%
Region
Northern
42.50% 17.90%
Central
45.40% 18.50%
Southern
46.40% 20.00%
Residence
Rural North
42.50% 18.30%
Rural Centre
46.30% 19.20%
Rural South
48.10% 20.80%
Urban North
41.70% 14.90%
Urban Centre
40.50% 14.40%
Urban South
38.10% 16.20%
Total
45.50% 19.10%
Source: Own computation from MDHS 2010
WAZ
<-2SD
< -3SD
WHZ
<-2SD < -3SD
13.50%
13.80%
3.70%
3.30%
6.40%
2.30%
2.60%
0.70%
14.80%
12.60%
3.10%
3.80%
4.40%
3.80%
1.80%
1.30%
11.40%
14.10%
3.20%
3.50%
2.40%
4.40%
0.60%
1.70%
12.70%
14.10%
13.40%
2.50%
4.00%
3.20%
2.70%
4.40%
4.10%
0.50%
1.80%
1.50%
13.80%
14.30%
13.90%
3.40%
13.20%
11.10%
13.70%
2.50%
4.00%
3.30%
2.50%
3.60%
2.80%
3.50%
3.00%
4.60%
4.50%
0.80%
2.80%
2.20%
4.10%
0.60%
2.00%
1.70%
0.00%
0.60%
0.60%
1.50%
The level of malnutrition is higher amongst children aged above 24 months compared to those below 24
months. We follow the WHO age-group comparison although a more detailed analysis can be done
across smaller age groups. Malnutrition levels are also lower amongst girls compared to boys. Urban
residents have a lower incidence of child malnutrition compared to rural residents. Amongst regions,
child malnutrition is the highest in the Southern region followed by the Central region. Similar
observations are made across areas and regions combined.
However, it is worth noting that the gap in the levels of child mal-nutrition across regions, areas and
regions are smaller compared to those obtained by age, sex and area; we get very similar incidence levels
of child mal-nutrition amongst the three regions, rural areas (Rural north, Rural Centre and Rural South)
and urban areas (Urban north, Urban Centre and Urban South). This seems to suggest that Malawi is
uniform in terms of the incidence of child malnutrition. On the other hand, we get a different picture
26
Stellenbosch University https://scholar.sun.ac.za
with assets where the incidence levels seem much different not only between areas but also across
regions and by sex of the household head.
2.8
Poverty lines
Based on household expenditure data, 50.7% of households in Malawi live below the poverty line. The
asset poverty line is set at an equivalent of 50.7% since it is assumed that the appropriate asset poverty
line should place the same proportion of households in poverty. The absolute poverty line for the asset
index is, therefore, the value at the 50.7th percentile and this turns out to be 0.63399.
With respect to child nutritional status, we convert each of the three anthropometric z-scores into
percentiles. This approach has also been previously used by Mussa (2010). The conversion involves
calculating the area under the standard normal curve to the left of the z-score. The area under the curve
adds up to unity and the mean (z-score of 0) splits the area into two halves of 0.5 each. The conversion
is monotonic and does not affect the ranking of the children. Therefore, for each point of the SDs, there
is a corresponding percentile or cumulative probability which is fixed along the x-axis. For example, if
we look up in the standard normal distribution tables, a z-score of -2 gives 0.0228 as the area to the left
of -2 or simply the 2.3rd percentile. Similarly, -3 corresponds to 0.0013 or the 0.13th percentile.
In this study, we are interested in malnutrition (<-2 SD) as opposed to extreme malnutrition (<-3 SD)
and, therefore, use 2.3 as our poverty line. A percentile gives the value of a variable below which a
certain percentage of observations (or population) falls. In our case, as we can see from Figure 2.2,
almost half of the data points for HAZ lie to the left of -2. We also note that all our three anthropometric
measures (HAZ, WAZ and WHZ) follow the standard normal distribution pattern.
2.9
Cumulative density curves
Cumulative density curves (CDCs) indicate how poverty incidence varies with the level of poverty lines
but are also used to test for dominance between two distributions. We show the CDCs for the asset index
(when 0 ) in Figure 2.3.A distribution whose curve lies above the other reflects a higher level of
poverty. The figures generally show that there is poverty dominance in the poverty relevant range. The
CDCs only cross each other at very high asset levels where it is difficult to conclude that poverty is
higher in one population subgroup than the other. However, it is only the poverty relevant range that we
are interested in for practical purposes. The figures also show that the gap between the CDCs is largest
between urban and rural areas; small differences exist between regions and by sex of household head.
27
Stellenbosch University https://scholar.sun.ac.za
1
.6
.2
0
0
.2
.4
.4
.6
.8
.8
Cumulative proportion of households
Cumulative proportion of households
1
Figure 2.3: MCA asset index cumulative density curves by population groups
0
.63399
1.4
2.8
4.2
Poverty line (MCA asset index)
7
0
.63399
Rural areas
1.4
2.8
4.2
Poverty line (MCA asset index)
Northern region
Central region
5.6
7
Southern region
0
.2
.2
.4
.4
.6
.6
.8
.8
C um u la tive p rop ortio n of h o use ho ld s
C um u la tive p rop ortio n of h o use h o ld s
1
1
Urban areas
5.6
.63399
1.4
2.8
4.2
Poverty line (MCA asset index)
Rural north
Urban north
Rural centre
Urban centre
5.6
7
0
0
0
Rural south
Urban south
.63399
1.4
2.8
4.2
Poverty line (MCA asset index)
Male head
5.6
7
Female head
Source: Own computation from MDHS 2010
2.10
FGT poverty estimates
Table 2.4 shows our poverty estimates for the three FGT classes, namely the headcount index ( 0) ,
the average poverty gap index ( 1) and the poverty severity index ( 2) , respectively. Household
observations are weighted by sampling weights and household size. Sampling weights are used so that
the chosen households are representative of all households in Malawi. On the other hand, household size
takes into account the effect of size on welfare. Thus, a poor household with more members is given a
higher weight in the analysis than a similarly poor household with fewer members. For the children’s
data set, we only apply sampling weights since household size is not applicable for individual level data.
Without the use of weights, our results would be either overestimated or underestimated.
The difference between the results in Table 2.4 and Figure 2.3 is that the latter is calculated for the whole
range of the poverty lines while the former is calculated at a specific and chosen level of the poverty
line. For example with the poverty line set at as asset index of 0.63399, the table shows that 46.0% of
28
Stellenbosch University https://scholar.sun.ac.za
the Malawian population is asset poor for 0 . For the same measure (headcount index), asset poverty
is higher in rural areas (53.3%) compared to urban areas (7.2%). The Central region has the highest
levels of asset poverty amongst the three regions with the incidence of household poverty at 51.3%. A
similar observation is made for rural centre and urban centre where 59.4% and 9.2% of the households
are living below the poverty line, respectively. Asset poverty is also higher in female-headed households
than male headed households.
Child malnutrition estimates are dependent on the type of measure used except for rural-urban
population group where rural areas are found to be the poorer than urban areas for all the three indicators,
namely HAZ, WAZ and WHZ. At a national level ( 0) , 47.1% of the children are malnourished
with respect to HAZ compared to 13.2% and 4.2% for WAZ and WHZ, respectively.
Table 2.4: Poverty headcount, average poverty gap and poverty severity estimates
Description
Area
Urban
Rural
Region
Northern
Central
Southern
Residence
Rural north
Rural centre
Rural south
Urban north
Urban centre
Urban south
Sex
Male head
Female head
Malawi
α=2
α=0
WHZ
α=1
40.8% 31.0% 26.4%
48.3% 38.0% 33.2%
10.5% 6.7% 5.1%
13.7% 8.7% 7.0%
2.5%
4.6%
1.8% 1.5%
3.3% 2.8%
7.2% 2.4% 1.1%
53.3% 20.3% 10.1%
44.8% 34.6% 30.0%
47.2% 37.2% 32.3%
47.6% 37.3% 32.6%
11.8% 6.8% 5.2%
13.5% 9.0% 7.2%
13.2% 8.3% 6.5%
2.9%
4.5%
4.3%
1.5% 1.0%
3.4% 2.9%
3.0% 2.6%
32.4% 11.0% 5.1%
51.3% 20.6% 10.6%
44.5% 16.2% 7.7%
44.9%
48.1%
49.3%
43.8%
42.1%
39.2%
12.9%
13.8%
13.8%
2.9%
12.3%
10.0%
5.6%
7.3%
6.9%
2.0%
6.6%
4.2%
3.1%
4.8%
4.7%
0.9%
2.9%
2.3%
1.6%
3.6%
3.3%
0.6%
2.1%
1.8%
1.1%
3.2%
2.8%
0.5%
1.6%
1.6%
34.5% 11.8% 5.5%
59.4% 24.0% 12.3%
52.9% 19.3% 9.2%
8.6% 2.0% 0.7%
9.2% 3.1% 1.5%
5.3% 1.8% 0.8%
12.9% 8.4% 6.7%
17.0% 9.3% 6.8%
13.2% 8.4% 6.7%
4.3%
4.1%
4.2%
3.0% 2.6%
3.2% 2.7%
3.0% 2.6%
41.7% 14.8% 6.9%
59.1% 25.8% 13.9%
46.0% 17.5% 8.6%
α=0
HAZ
α=1
35.0%
38.0%
38.9%
32.0%
32.4%
29.5%
α=2
30.5%
33.1%
34.1%
26.0%
27.7%
25.3%
47.1% 36.8% 32.1%
47.6% 38.5% 34.0%
47.1% 37.0% 32.2%
α=0
WAZ
α=1
7.3%
9.1%
8.7%
2.3%
8.4%
5.9%
α=2
α=0
Asset index
α=1
α=2
Source: Own computation from MDHS 2010
It is, however, important to state that the national poverty estimates based on HAZ and asset index are
similar in magnitude suggesting a similar national poverty profile for Malawi. We also note that the
differences in the incidence of poverty between groups (e.g. rural and urban areas) seem to be larger for
the asset index than for the three anthropometric measures. Consequently, the results tell us that levels
of malnutrition are similar in magnitude across many population subgroups in Malawi. On the other
hand, the incidence of asset poverty varies much by population groups and a similar pattern is obtained
based on household consumption expenditure studies such as Mussa (2013). Although we have not
29
Stellenbosch University https://scholar.sun.ac.za
presented the CDCs of HAZ, WAZ and WHZ, we find them to be much closer to one another than is
the case for the asset index in Figure 2.3.
Poverty ranking by population group is the same for all the three indices, 0 , 1 and 2 .
However, this result only holds at the specifically chosen level of poverty line and not all poverty lines.
As discussed, it is difficult to conclusively identify which population groups have higher poverty levels
because the CDCs cross at different levels. The estimates depend on the choice of the poverty measure
used as is the case with child nutritional status. This problem is resolved through stochastic dominance
analysis dealt with in Section 2.11 which follows.
2.11
Poverty dominance analysis
Our dominance tests are done up to the third order, as is the tradition in the literature (e.g., Mussa, 2013).
Naturally, where first order dominance exists, there is no need to test for second or third order
dominance, or where second order dominance exists, there is no need to test for third order dominance,
as it would automatically hold. Table 2.5 shows the results. Poverty dominance is measured and
interpreted in terms of welfare or standard of living. A distribution that dominates the other has lower
levels of poverty than the one that is dominated.
Table 2.5: Poverty stochastic dominance test results for population subgroups
Population group pair
HAZ
Area
Urban v. Rural
U>R: order 2
Region
Northern v. Central
ND
Northern v. Southern
ND
Central v. Southern
ND
Residence
Rural north v. Rural centre
ND
Rural north v. Rural south
ND
Rural centre v. Rural south
ND
Urban north v. Urban centre
UN>UC: order 2
Urban north v. Urban south
ND
Urban centre v. Urban south
ND
Sex
Male head v. female head
ND
Source: Own computation from MDHS 2010
WAZ
WHZ
Asset
ND
U>R: order 2
U>R: order 1
ND
ND
ND
N>C: order 2
N>S: order 1
ND
ND
N>S: order 3
ND
ND
ND
ND
ND
ND
ND
RN>RC: order 3
RN>RS: order 1
ND
UN>UC: order 2
UN>US: order 1
ND
ND
RN>RS: order 2
ND
ND
ND
UC>US: order 2
ND
ND
MH>FH
Notes: U=Urban areas, R=Rural areas, ND= no dominance up to third order, UN=Urban north,
UC=Urban centre, US=Urban south, N=Northern region, C=Central region, S=Southern region,
RN=Rural north, RC=Rural centre, RS=Rural south, MH=male-headed household, FH=femaleheaded household
30
Stellenbosch University https://scholar.sun.ac.za
As with the FGT estimates, dominance test results for child nutritional status depend on the measure
used. Where dominance is established, the order of dominance is not reversed. For example, all measures
establish that urban areas are better in terms of welfare than rural areas and not the opposite. Using HAZ,
urban areas dominate rural areas at second order dominance. Likewise, urban north dominates urban
centre. No dominance is established for the rest of the subgroups. When WAZ is used, no dominance is
established at all for all the pairs of subgroups. Finally, when WHZ is used, non-dominance is found for
the Central and Southern regions, Rural centre and Rural south and Urban centre and Urban south. No
dominance is established between male and female-headed households- the same finding for HAZ and
WAZ. It is, however, important to note that in most of these cases, non-dominance occurs outside the
poverty relevant ranges of the poverty line as illustrated in the CDCs.
Using the asset index, urban areas dominate rural areas at first order. The Northern region dominates
the Southern region. When areas and regions are further broken down, dominance is only established
for two pairs, namely rural north dominates rural south and urban centre dominates urban south. Male
headed households dominate those headed by females.
2.12
Gini and GE inequality estimates
In Figure 2.4 are Lorenz curves for Malawi’s population groups based on the asset index. The Lorenz
curve maps the cumulative share of the asset index, on the vertical axis against the population
distribution on the horizontal axis. The 45-degree line indicates perfect equality, a case in which each
household has the same share of assets. The numerical inequality estimates for both the asset index and
child malnutrition are presented in Table 2.6 that follows.
0
0
.2
.2
.4
.4
.6
.6
.8
.8
Cumulative share of the (MCA) asset index
Cumulative share of the (MCA) asset index
1
1
Figure 2.4: MCA asset index Lorenz curves by population subgroups
0
.2
.4
.6
Cumulative proportion of households
.8
1
0
.2
.4
.6
Cumulative proportion of households
45° line
Population
45° line
Population
Urban areas
Rural areas
Central region
Southern region
31
.8
Northern region
1
0
.2
.2
.4
.4
.6
.6
.8
C um ulative share of the (M C A) asset index
.8
1
1
C um ulative share of the (M C A) asset index
Stellenbosch University https://scholar.sun.ac.za
.2
.4
.6
Cumulative proportion of households
.8
1
0
0
0
.2
.4
.6
.8
45° line
Population
Rural north
Rural centre
Rural south
Urban north
45° line
Population
Urban centre
Urban south
Male head
Female head
Cumulative proportion of households
1
Source: Own computation from MDHS 2010
Table 2.6: Inequality estimates across population subgroups
Description
Gini
Theil L (theta=0)
HAZ WAZ WHZ Asset HAZ WAZ WHZ Asset
Theil T (theta=1)
HAZ WAZ WHZ Asset
2.165 0.756 0.279 0.205
2.563 0.889 0.429 0.261
0.852 0.364 0.149 0.170
1.058 0.459 0.186 0.267
2.416 0.787 0.292 0.244
2.547 0.918 0.430 0.353
2.481 0.840 0.408 0.362
1.018 0.424 0.159 0.230
1.021 0.447 0.181 0.356
1.029 0.445 0.186 0.361
2.487
2.608
2.532
1.811
2.189
2.182
1.040
1.051
1.070
0.845
0.850
0.842
Area
Urban
0.683 0.461 0.281 0.326
Rural
0.744 0.521 0.313 0.397
Region
Northern region 0.733 0.502 0.290 0.376
Central region 0.734 0.513 0.306 0.457
Southern region 0.736 0.513 0.314 0.462
Residence
Rural north
0.739 0.512 0.300 0.361
Rural centre
0.742 0.519 0.309 0.394
Rural south
0.747 0.523 0.321 0.397
Urban north
0.680 0.413 0.197 0.337
Urban centre
0.681 0.475 0.292 0.338
Urban south
0.678 0.454 0.281 0.308
Sex
Male head
0.734 0.512 0.308 0.441
Female head
0.741 0.519 0.313 0.478
Malawi
0.735 0.513 0.308 0.453
Source: Own computation from MDHS 2010
0.794
0.943
0.851
0.691
0.768
0.752
0.310
0.450
0.437
0.116
0.316
0.268
0.225
0.257
0.259
0.200
0.220
0.184
2.502 0.872 0.402 0.327
2.528 0.856 0.453 0.385
2.505 0.871 0.406 0.348
0.441
0.457
0.462
0.282
0.390
0.353
0.168
0.183
0.194
0.075
0.166
0.145
0.210
0.265
0.272
0.180
0.182
0.152
1.022 0.443 0.180 0.325
1.048 0.459 0.186 0.400
1.024 0.444 0.181 0.346
The results show that inequality is higher in rural areas compared to urban areas and this is consistent
for all the three inequality measures used, namely Gini, Theil L and Theil T. Nevertheless, amongst the
three measures of child nutrition, HAZ generally yields the highest levels of inequality. Just as with
poverty measurement, levels of inequalities depend on the measure used. Consequently, it is difficult to
conclude without ambiguity which population groups have higher levels of inequality. We, therefore,
32
Stellenbosch University https://scholar.sun.ac.za
turn to inequality dominance testing as a solution. The inequality dominance test results are presented
in Section 2.13 and act as robustness checks.
2.13
Inequality dominance analysis
Results of the Lorenz dominance tests are presented in Table 2.7. Using HAZ, inequality dominance is
only established for the urban and rural area pair. Since the Lorenz curve for urban areas lies above that
of rural areas, inequality is said to be lower in urban areas compared to rural areas. This finding is
confirmed with WHZ as well as the asset index. No dominance is established using WAZ.
Table 2.7: Generalised Lorenz dominance test results across population subgroups
Population group pair
HAZ
Area
Urban v. Rural
U>R: order 2
Region
Northern v. Central
ND
Northern v. Southern
ND
Central v. Southern
ND
Residence
Rural north v. Rural centre
ND
Rural north v. Rural south
ND
Rural centre v. Rural south
ND
Urban north v. Urban centre
ND
Urban north v. Urban south
ND
Urban centre v. Urban south
ND
Sex
Male head v. female head
ND
Source: Own computation from MDHS 2010
WAZ
WHZ
Asset
ND
U>R: order 2
ND
ND
ND
N<C
N<S
ND
ND
ND
ND
ND
ND
ND
ND
RN<RS: order 2
ND
UN<UC: order 2
UN<US: order 2
ND
ND
RN<RS: order 2
ND
ND
ND
UC<US: order 2
ND
ND
MH<FH: order 2
U>R: order 2
ND
ND
ND
Notes: U=Urban areas, R=Rural areas, ND= no dominance, UN=Urban north, UC=Urban centre,
US=Urban south, N=Northern region, C=Central region, S=Southern region, RN=Rural north,
RC=Rural centre, RS=Rural south, MH=male-headed household, FH=female-headed household
Amongst the three regions, no inequality dominance is established using HAZ, WAZ and asset index.
For WHZ, inequality is found to be the highest in the Northern region compared to both the Central and
Southern regions. With respect to the sex of the household head, dominance is established for the asset
index only, in which case inequality is found to be higher among male-headed households. WHZ and
asset index establish dominance between some areas such as Rural north, Rural south, Urban north,
Urban centre and Urban south.
33
Stellenbosch University https://scholar.sun.ac.za
2.14
Poverty decomposition
Table 2.8 shows results of the FGT decompositions across rural-urban areas, regions, areas and sex of
the household head for the three anthropometric indicators and the asset index. We provide the estimated
population share of each subgroup as well as the estimated relative contribution of each subgroup to
total poverty. Three FGT measures are used, namely the poverty headcount, average poverty gap and
poverty severity indices. The population shares and relative contributions to poverty sum up to unity.
The objective here is to establish if the most populated areas are also the ones with the highest incidence
of poverty.
Our decomposition results for the asset index indicate that most poor people in Malawi are also in the
rural areas-where the population is large. For example, the table shows that 84% of the households are
based in rural areas which contribute 97.5% to the poverty headcount when the asset index is used as
the measure of welfare. The contributions are slightly higher with the poverty gap and severity indices
at 97.8% and 98%, respectively. This does not only indicate that the incidence of poverty is above the
national average in rural areas but also that the poorer parts of the population are more concentrated in
rural areas. Amongst the three regions, much of the poverty is contributed by the Central region followed
by the Northern region. A similar observation is made within rural and urban areas, e.g. Rural north and
Rural centre. Households headed by males contribute more to poverty than those headed by females.
Using the nutritional measures, rural areas are also found to contribute more to poverty than urban areas
but the difference between the rural population share and the contribution of rural areas to these
measures of nutritional poverty is small. This is surprising since it appears as if nutritional status is not
much better in urban than in rural areas, despite the fact that asset holdings in urban areas are definitely
considerably greater and asset poverty less severe. With respect to the three regions of the country, it is
the Central region that contributes the highest to poverty in both absolutes and relative terms. Rural
centre and urban centre also contribute more to poverty than the other population subgroups. Finally,
just as is the case with the asset index, male-headed households contribute more to poverty than those
headed by females.
2.15
Subgroup inequality decomposition
Table 2.9 presents generalised entropy inequality estimates decomposed to indicate within (vertical) and
between (horizontal) subgroup inequalities. Irrespective of the measures used, inequalities are largely
driven by within population subgroups as opposed to between population subgroups. Between sub-group
inequality is only a greater magnitude for the urban-rural areas comparison using the Theil L or Theil T
measure for asset inequality, when it contributes somewhere between one-quarter and one-third to
overall asset inequality. All other between group contributions to inequality are extremely small,
reflecting the fact that locational differences in nutritional status appear to be non-systematic.
34
Stellenbosch University https://scholar.sun.ac.za
Table 2.8: FGT poverty sub-group decomposition
Description
Population
share
HAZ
α=1
α=0
α=2
Area
Urban
15.0%
13.0% 12.6% 12.4%
Rural
85.0%
87.0% 87.4% 87.6%
Region
Northern
10.7%
10.0%
9.8%
9.8%
Central
46.4%
46.3% 46.5% 46.4%
Southern
42.9%
43.7% 43.7% 43.8%
Residence
Rural north
9.6%
9.0%
8.9%
8.9%
Rural centre
39.6%
40.2% 40.5% 40.5%
Rural south
35.8%
37.7% 38.0% 38.2%
Urban north
1.1%
1.0%
1.0%
0.9%
Urban centre
6.7%
6.0%
5.9%
5.8%
Urban south
7.2%
6.0%
5.7%
5.6%
Sex
Male head
91.9%
91.9% 91.6% 91.5%
Female head
8.1%
8.1%
8.4%
8.5%
Malawi
100.0%
100.0% 100.0% 100.0%
Source: Own computation from MDHS 2010
α=0
WAZ
α=1
α=0
WHZ
α=1
α=2
12.6%
87.4%
12.9%
87.1%
10.0%
47.8%
42.3%
9.7%
41.3%
36.4%
0.3%
6.4%
5.9%
α=2
12.6%
87.4%
8.6%
91.4%
8.8%
91.2%
8.4%
91.6%
2.5%
97.5%
2.2%
97.8%
2.0%
98.0%
9.4%
48.9%
41.7%
9.4%
49.1%
41.5%
7.1%
49.6%
43.3%
5.1%
52.3%
42.7%
4.3%
53.0%
42.7%
8.4%
48.4%
43.3%
7.5%
51.1%
41.4%
7.0%
53.0%
40.0%
9.1%
42.2%
35.8%
0.4%
6.6%
5.9%
9.0%
42.7%
35.8%
0.4%
6.4%
5.7%
6.9%
45.1%
39.4%
0.2%
4.5%
3.9%
4.8%
47.8%
38.5%
0.2%
4.5%
4.1%
4.1%
49.1%
38.4%
0.2%
3.9%
4.3%
8.2%
47.0%
42.3%
0.2%
1.4%
0.9%
7.4%
49.9%
40.6%
0.1%
1.3%
0.8%
6.9%
51.8%
39.2%
0.1%
1.2%
0.7%
90.0% 91.4% 91.9%
10.0%
8.6%
8.1%
100.0% 100.0% 100.0%
35
91.6% 90.6% 90.4%
8.4%
9.4%
9.6%
100.0% 100.0% 100.0%
Asset index
α=0
α=1
α=2
68.5% 63.9% 60.6%
31.5% 36.1% 39.4%
100.0% 100.0% 100.0%
Stellenbosch University https://scholar.sun.ac.za
Table 2.9: GE inequality decomposition for asset index and child nutrition
Description
GE index
Sub-group
Type
Areas
Absolute Relative
Within
0.254
Asset Theil L
Between
0.098
Total
0.353
Within
0.236
Between
0.116
Asset Theil T
Total
0.352
Within
2.503
HAZ Theil L
Between
0.002
Total
2.505
Within
1.022
HAZ Theil T
Between
0.002
Total
1.024
Within
0.869
WAZ Theil L
Between
0.002
Total
0.871
Within
0.442
WAZ Theil T
Between
0.002
Total
0.444
Within
0.406
WHZ Theil L
Between
0.000
Total
0.406
Within
0.180
WHZ Theil T
Between
0.000
Total
0.181
Source: Own computation from MDHS 2010
72.2%
27.8%
100.0%
66.9%
33.1%
100.0%
99.9%
0.1%
100.0%
99.8%
0.2%
100.0%
99.7%
0.3%
100.0%
99.5%
0.5%
100.0%
100.0%
0.1%
100.0%
99.9%
0.2%
100.0%
Regions
Absolute Relative
0.349
0.004
0.353
0.348
0.004
0.352
2.505
0.000
2.505
1.024
0.000
1.024
0.871
0.001
0.871
0.444
0.001
0.444
0.406
0.000
0.406
0.180
0.000
0.180
99.0%
1.0%
100.0%
99.0%
1.0%
100.0%
100.0%
0.0%
100.0%
100.0%
0.0%
100.0%
99.9%
0.1%
100.0%
99.9%
0.1%
100.0%
100.0%
0.0%
100.0%
99.9%
0.0%
100.0%
36
Within areas/regions
Absolute Relative
0.248
0.105
0.353
0.230
0.122
0.352
2.501
0.003
2.505
1.020
0.003
1.024
0.868
0.004
0.871
0.441
0.004
0.444
0.406
0.000
0.406
0.180
0.000
0.180
70.3%
29.7%
100.0%
65.3%
34.7%
100.0%
99.9%
0.1%
100.0%
99.7%
0.3%
100.0%
99.6%
0.4%
100.0%
99.2%
0.8%
100.0%
99.9%
0.1%
100.0%
99.7%
0.2%
100.0%
Household head
Absolute Relative
0.344
0.009
0.353
0.344
0.008
0.352
2.505
0.000
2.505
1.024
0.000
1.024
0.871
0.000
0.871
0.444
0.000
0.444
0.406
0.000
0.406
0.180
0.000
0.180
97.5%
2.5%
100.0%
97.6%
2.4%
100.0%
100.0%
0.0%
100.0%
100.0%
0.0%
100.0%
100.0%
0.0%
100.0%
100.0%
0.0%
100.0%
100.0%
0.0%
100.0%
100.0%
0.0%
100.0%
Stellenbosch University https://scholar.sun.ac.za
2.16
Spatial distribution of poverty and inequality
The maps are drawn to provide a visual representation of the FGT headcount and Gini inequality
estimates presented in Table 2.4 and Table 2.6. They are drawn for Malawi’s three regions (Northern,
Central and Southern) using StatPlanet, an interactive data visualisation and mapping software9. The
colours on the maps vary positively with the levels of poverty and inequality. Therefore, darker and
lighter colours represent areas with higher and lower levels of poverty and inequality, respectively.
Figure 2.5 and Figure 2.6 provide the spatial distribution of poverty and inequality using our four
selected indicators, namely the asset index, HAZ, WAZ and WHZ. The mapping shows little variation
in terms of both poverty and inequality estimates across Malawi’s regions.
Figure 2.5: Spatial distribution of poverty using asset index and child-nutritional status
Poverty headcount (asset index)
9
Poverty headcount (HAZ)
We multiply our estimates by 100 for easy customisation of the map colours in StatPlanet.
37
Stellenbosch University https://scholar.sun.ac.za
Poverty headcount (WAZ)
Poverty headcount (WHZ)
Source: Own computation from MDHS 2010
The maps also confirm that regardless of the indicator used, the poverty incidence is the lowest in the
Northern region. The ranking between the Central and Southern regions is dependent on the indicator
used. Based on the asset index, WAZ and WHZ, the Central region has the highest incidence of poverty.
On the other hand, the Southern region has the highest incidence of poverty when HAZ is used as the
indicator.
With respect to Gini mapping, it is also shown that inequality is lowest in the Northern region for all the
measures used. The Central and Southern regions fairly rank the same for all the measures used apart
from HAZ where inequality is highest in the Southern region. This ranking is similar to the one found
in the poverty headcount when HAZ is used.
38
Stellenbosch University https://scholar.sun.ac.za
Figure 2.6: Spatial distribution of inequality for asset index and child-nutritional status
Gini inequality (asset index)
Gini inequality (HAZ)
Gini inequality (WAZ)
Gini inequality (WHZ)
Source: Own computation from MDHS 2010
39
Stellenbosch University https://scholar.sun.ac.za
2.17
Factors affecting asset poverty
So far, we have only looked at the poverty profile for Malawi without explaining the factors associated
with poverty. As acknowledged in the literature, a satisfactory explanation of why households are poor
is important if we are to be able to deal with the roots of poverty (Haughton & Khandker, 2009). This
chapter, therefore, addresses the question of what factors are associated with asset poverty in Malawi
using the 2010 MDHS data. Our dependent variable is the asset index, reflecting long-term economic
status. The choice of the explanatory variables is informed by the literature.
The correlates include household size, age dependency ratio, age of the household head, sex of the
household head (male=1, female=2), incidence of sickness in the household (no=0, yes=1) and the levels
of education in the household. We also include dummies for area (urban=1, rural=2) and region
(1=Northern, 2=Central and 3=Southern region) to control for the location in which the households
reside. The educational status of household members has four categories, namely 0=no education and
preschool, 1=primary education, 2=secondary education and 3=post-secondary education. Shocks to the
household are represented by the incidence of sickness in the household, captured as (no=0,
yes=1).Table 2.10 shows descriptive statistics of the variables used in our analysis of asset poverty.
Table 2.10: Summary descriptive statistics for the asset model
Variables
Obs
Mean/Prop.
Std. Dev.
Asset index
24,825
0.95
0.903
Age of head
24,798
43.34
16.368
Household size
24,825
4.79
2.295
Dependency ratio
24,825
49.2%
0.245
Female head
24,825
28.4%
0.451
Sickness in the household
19,947
2.5%
0.157
Primary
24,724
61.0%
0.488
Secondary
24,724
17.3%
0.378
Higher
24,724
2.6%
0.160
24,825
88.3%
0.322
Central
24,825
33.7%
0.473
Southern
24,825
48.8%
0.500
Household's education
Rural area
Region
Source: Own computation from MDHS 2010
The proportions relate to categorical variables. The asset index has a mean value of about 0.95 and
standard deviation of 0.903. On average, the household head is aged about 43 years with standard
40
Stellenbosch University https://scholar.sun.ac.za
deviation of 16.368. Based on the sample, the average household size in Malawi is about 4.79 persons.
Based on an alternative data set, the Malawi IHS3 data, the average household size is 4.6 members. The
age dependency ratio stands at 49% and is calculated as the number of children aged 0 to 14 years and
the number of persons aged 65 years divided by the household size10. About 28% of the households are
headed by females. With respect to the incidence of sickness, only about 2.5% of the households
reported not to have been very sick for 3 or more months in the year prior to the survey. With regards
to the levels of education, most households (about 61%) have primary education as the highest level of
qualification. This is followed by no education or preschool education at 19%, secondary education at
17.3% and finally post-secondary education at 2.6%. With regard to area of residence, about 88% of the
people reside in households which are rural compared to 12% which are urban. The Southern region has
the highest number of households with 49%, followed by the Central region with 34% and finally the
Northern region with 17%.
Table 2.11 presents the OLS regression results of the asset index. The results show that about 55% of
the variance in the asset index is explained by the explanatory variables. All the variables have the
expected signs and statistically significant at conventional levels. There is also satisfactory performance
for our control variables.
We checked for the robustness of the model based on an alternative definition of the age dependency
ratio and use of sickness of the father and mother in place of sickness for any of the household members.
Our preferred model, which is presented, was chosen because it gives the largest size of R-squared and
age of the household head had the expected positive sign.
The coefficient for household size is positive and statistically significant. It indicates that a unit increase
in household size is associated with an increase of 0.071 standard deviations in the asset index while
holding all the other factors constant. Put differently, larger households are associated with more assets.
The coefficient for age dependency ratio is also significant coefficient but negative indicating that the
greater the proportion of economically inactive members in the household, the poorer is that household
in terms of asset ownership. Households that are headed by older people are associated with better longterm economic status by about 0.005 standard deviations. The regression results indicate that the there
is no statistically significant difference in the asset ownership between male and female headed
households.
10
Strictly speaking, age dependency ratio is given as the ratio of dependents (people younger than 15 or older than
64) to the working-age population (those aged 15-64). However, this strict definition generates missing values
since some of the households do not have people in either of the three age categories.
41
Stellenbosch University https://scholar.sun.ac.za
With respect to education, it is found that higher levels of education are positively associated with asset
ownership in Malawi when compared to the base category (no education). It is worth noting with great
interest that the size of the education coefficient increases with the level of education. Urban areas are
associated with higher levels of asset ownership than rural areas by about 1.051 standard deviations.
Regional dummies have mixed performance with the Central region being statistically different from
the Northern region which is the base. Controlling for the 27 districts in the data does not add much to
the model so we do not report in the tables.
Table 2.11: OLS regression results for asset poverty
Description
Household size
Age dependency ratio
Age of household head
Sex of household head
Household member is sick
Household education
Primary
Secondary
Higher
Rural area
Region
Central
Southern
Constant
R-squared
Observations
Prob >F
F statistic
Asset index
0.071***
-0.505***
0.005***
0.038**
-0.053
SE
(0.005)
(0.047)
(0.001)
(0.015)
(0.034)
0.236***
0.864***
2.390***
-1.051***
(0.015)
(0.037)
(0.093)
(0.067)
-0.141***
0.003
2.283***
0.552
19,900
0.000
190.350
(0.040)
(0.039)
(0.130)
Notes: *, **, *** denote significance at 10%, 5% and 1% levels
2.18
Child nutritional status in Malawi
We look at factors associated with child nutritional status in Malawi using the 2010 MDHS children
data set. A previous study by Ngalawa & Chirwa (2008) was based on the 1997-1998 IHS data set. As
earlier discussed, DHS surveys provide a better source of data on anthropometric measurements when
compared to IHS data (see Verduzco-Gallo et al., 2014). Furthermore, the present analysis is based on
the most recent data. Our dependent variables in the study are HAZ, WAZ and WHZ as indicators of
stunting, underweight and body wasting, respectively. Table 2.12 shows descriptive statistics of the
variables used in our three nutrition models. The proportions relate to all categorical variables just as
before.
42
Stellenbosch University https://scholar.sun.ac.za
Table 2.12: Descriptive statistics for the nutritional models
Variable
HAZ
WAZ
WHZ
Age in months
Square of age in moths
Weight at birth
Parental age difference
Birth order number
Female child
Rural area
Child is twin
Female household head
Asset index
Square of asset index
Mother's education
Incomplete primary
Complete primary
Incomplete secondary
Complete secondary
Higher
Father's education
Incomplete primary
Complete primary
Incomplete secondary
Higher
Region
Central
Southern
Obs
4,653
4,783
4,609
4,801
4,801
4,801
4,411
4,801
4,801
4,801
4,801
4,801
4,801
4,801
Mean Std. Dev.
-1.76
1.662
-0.79
1.215
0.31
1.35
28.91
16.903
11.22
10.334
5.64
3.271
5.52
4.692
3.69
2.321
50.60%
0.50
90.00%
0.299
3.20%
0.176
7.90%
0.269
-0.097
0.899
0.817
3.235
4,801
4,801
4,801
4,801
4,801
60.60%
9.40%
9.30%
3.40%
0.50%
0.489
0.292
0.291
0.182
0.068
4,717
4,717
4,717
4,717
56.30%
8.30%
23.90%
1.60%
0.496
0.275
0.426
0.127
4,801
4,801
37.10%
45.50%
0.483
0.498
Source: Own computation from MDHS 2010
Explanatory variables include child characteristics such as sex (male=1, female=2), age in months, age
squared included to account for the possible non-linearity between age and nutritional status, the weight
at birth and the status of being a twin (no=0, yes=1) and absolute birth order. We control for area of
residence (urban=1, rural=2), sex of the household head (male head=1, female head=2), mothers’ and
fathers’ education (0/5). The levels of education for the father and mother are represented by dummies
representing six categories, namely no education, incomplete primary, complete primary, incomplete
secondary, complete secondary and higher education(post-secondary education). Father’s education has
five categories only. There are no fathers with complete secondary education in the data set.
43
Stellenbosch University https://scholar.sun.ac.za
The age difference between the father and mother to captures the bargaining position of the mother in
the household. According to the bargaining literature on household decisions, bargaining status could
influence the resources that the mother may receive for herself as well as for her child, possibly leading
to adverse nutrition consequences (Linnemayr, Alderman & Ka, 2008). Finally, we include the asset
index to capture the economic status of the household to which the child belongs. In the literature,
economic status has been found to be a strong determinant of the nutritional status of the children (e.g.,
Dancer, Rammohan & Smith, 2008).
The table reveals that amongst the three measures, WHZ has the highest mean followed by WAZ. The
average age of the children is about 29 months. On average, the children were born with a weight of
around 5.64kg which is actually high and in line with our statistic suggesting that being underweight is
the least of the malnutrition problems in Malawi. The difference in age between fathers and mothers is
about 5.52 years. The average birth order of the children in our study is 3.69. Rural areas constitute a
very large proportion of children- about 90% of the total. The incidence of a twin is very low at 3.2% of
the total number of children. Only about 8% of the children come from households which are headed by
females. The majority of the mothers have incomplete primary education (60.6%) compared to 56.3%
for fathers. About 46% and 37% of the children reside in the Southern region Central region,
respectively.
2.19
Multivariate analysis of child nutrition
Table 2.13 presents results from the OLS regression analysis of HAZ, WAZ and WHZ11. The model
statistics show that the R-squared stands at 10%, 9% and 11% for HAZ, WAZ and WHZ, respectively.
Though the F-statistics indicate the hypothesis that all slope coefficients are equal to zero, it is important
to note that the WHZ model does not perform well. The results do not improve even if we regress by
the recommended WHO age categories of <24 months and >=24 months (see Table A3 in the appendix).
This finding is in line with an earlier study by Chirwa and Ngalawa (2008) which suggests that children
in Malawi seem to be ‘fatter for their age’. There could be other factors such as poor feeding practices
or genetic factors that affect their nutritional status or indeed a concern for consistent measurement error.
Child characteristics such as age, sex and twin status are statistically significant for HAZ and WAZ but
not WHZ. Birth order is significant for the WAZ and WHZ models only. The relationship between the
age of the children and nutritional status is negative and statistically significant for HAZ and WAZ only.
The square of age is included to account for possible non-linearity that exists between child nutrition
and age. The coefficient of age-squared is positive for HAZ and WAZ and statistically significant at
11
Area of residence, the weight of the child at birth, parental age difference and sex of the household head are not
significant and, therefore, not reported in the results.
44
Stellenbosch University https://scholar.sun.ac.za
1%. This implies that the relationship between age and nutritional status of a child is U-shaped
(convexity) with respect to HAZ and WHZ. When WHZ is used, the relationship is inverted U-shaped
(concavity) but insignificant. The convexity or concavity indicates that malnutrition in under-five
children worsens or improves with age but this is only up to some critical age beyond which a child’s
nutrition status improves or worsens with age.
Table 2.13: OLS regression results for child nutritional status
Variables
Age in months
Square of age
Female child
Child is twin
Birth order number
Mother's education
Incomplete primary
Complete primary
Incomplete secondary
Complete secondary
Post-secondary
Father's education
Incomplete primary
Complete primary
Incomplete secondary
Post-secondary
Asset index
Square of asset index
Rural area
Region
Central
Southern
Constant
R-squared
Prob > F
Observations
HAZ
Coeff
SE
-0.094*** (0.008)
0.128*** (0.012)
0.280*** (0.057)
-0.930*** (0.169)
0.025
(0.016)
WAZ
Coeff
SE
-0.045*** (0.007)
0.052*** (0.010)
0.102*
(0.048)
-0.901*** (0.160)
0.029*
(0.014)
WHZ
Coeff
SE
0.002
(0.007)
-0.001
(0.010)
-0.072
(0.049)
-0.286* (0.145)
0.038** (0.015)
0.082
0.110
0.186
0.060
0.439
(0.099)
(0.138)
(0.143)
(0.192)
(0.473)
0.154*
0.057
0.302**
0.243
0.236
(0.076)
(0.100)
(0.114)
(0.172)
(0.384)
0.124
0.112
0.165
0.369*
-0.049
(0.083)
(0.118)
(0.116)
(0.166)
(0.428)
-0.095
-0.178
-0.006
-0.055
0.183**
-0.011
0.127
(0.124)
(0.168)
(0.139)
(0.235)
(0.063)
(0.016)
(0.108)
-0.191*
-0.208
-0.202*
-0.169
0.195***
-0.017
0.052
(0.082)
(0.111)
(0.101)
(0.159)
(0.052)
(0.011)
(0.094)
-0.048
0.083
-0.092
-0.007
-0.002
0.008
-0.105
(0.094)
(0.130)
(0.108)
(0.211)
(0.055)
(0.014)
(0.095)
0.009
-0.081
-0.259
0.088
0.000
4,700
(0.075)
(0.072)
(0.197)
-0.102
-0.174*
0.368
0.011
0.010
4,531
(0.075)
(0.072)
(0.213)
-0.031
(0.111)
-0.006
(0.110)
-1.031*** (0.261)
0.101
0.000
4,574
Notes: *, **, *** denote significance at 10%, 5% and 1% levels
The critical ages are 37 months for HAZ, 44 months for WAZ and 54 months for WHZ. The significance
and signs of coefficients for both of age and age squared are also confirmed in Chirwa and Ngalawa
(2006) who find 30, 34 and 35 months as critical ages for HAZ, WAZ and WHZ. There exists literature
which links recovery from stunting to cognitive outcomes. For example, a panel data study by Casale
and Desmond (2015) in South Africa finds that children who were stunted at age 2 (in years) but
recovered by age 5 performed better than children who remained stunted over the study period. They
also find that children who recover from stunting by age 5 perform significantly worse on cognitive tests
45
Stellenbosch University https://scholar.sun.ac.za
than children who do not experience early malnutrition. Moreover, they find that recovery from stunting
is not uncommon among children- a similar finding in this study. However, the most important thing is
the timing of recovery as well as the need to provide the required nutritional inputs at an early stage if
the goal is to help improve children’s cognitive performance. Evidence in this research suggests it took
longer by between 7 and 19 months for children to ‘recover’ in 2010 compared to what Chirwa and
Ngalawa (2006) found using the 1997/8 data12.
The coefficient of a variable capturing the status of being one of a twin is negative in all models but
only statistically significant for HAZ and WAZ. Mussa (2014) uses HAZ as a measure of long-term
nutritional status and finds opposite signs of coefficients for age, age-squared and twin status of a child.
With respect to age and age-squared, Mussa (2014) finds concavity in which case the nutritional status
of a child improves with age but begins to worsen after some time. This is the same finding in this study
for the WHZ model although not statistically significant. We have found that the coefficient for birth
order is also positive and statistically significant for WAZ and WHZ. Mussa (2014) finds a significant
relationship for rural areas only which seems to confirm the empirical finding that the effect of childorder is cultural and must be interpreted within a given cultural context. With respect to parental agedifference, we find no significant relation for all the three models. On the other hand, Mussa (2014)
finds a negative relationship implying that households with older fathers compared to mothers perform
worse in terms of child nutritional outcomes.
The level of education of the mother matters more than that of the father in our models for child
nutritional status. For both the fathers and mothers, no education at all is used as the base category. All
levels of mothers’ education except post-secondary education play a significant role in the WAZ model.
In the HAZ model it is only incomplete secondary that matters while for WHZ, incomplete primary and
complete secondary school levels play a role. The education level of the father is only significant for
incomplete primary education and only for the WAZ model. However, this happens at the lowest levels
of education and has a negative coefficient implying that lower levels of education are associated with
poor nutritional status. Mussa (2014), who also uses “no education” as the base category, finds negative
coefficients for all levels of education but only statistically significant for some levels of educations.
Chirwa and Ngalawa (2006) find mixed results, a case also found in this study. Lower levels of education
have negative coefficients while higher levels have positive coefficients. Chirwa and Ngalawa (2006)
find that both the mothers’ and father’s matter.
12
Since we do not have panel data, we cannot conclude this is recovery but simply a finding that is consistent with
possible patterns of recovery.
46
Stellenbosch University https://scholar.sun.ac.za
Higher asset index scores are associated with better levels of child nutrition levels for two models only,
HAZ and WAZ. Mussa (2014) uses five categories of the asset index of the households, namely poorest,
poor, middle, richer and richest. Using the poorest as the base category, he finds that household wealth
seems to matter more in improving height-for-age z-scores in rural areas than in urban areas. For
example, a child born into the wealthiest quintile in rural areas has a height-for-age z-score that is 0.31
standard deviations better than that of a child from the poorest wealth quintile. We also used asset
quintiles (poorest as the base category) and found significant results (HAZ and WAZ only) for all
quintiles except the middle quintile. Of great similarity is the finding that a child from the richest quintile
is better by 0.33 and 0.32 standard deviations for HAZ and WAZ, respectively. The square of the asset
index displays convexity for HAZ and WAZ and concavity for WHZ. The turning points are asset index
scores of 5.66 (richest quintile), 7.11 (richest quintile) and -0.45(middle quintile) for HAZ, WAZ and
WHZ, respectively.
2.20
Asset index and pro-poor growth analysis
Over time, DHS surveys have been upgraded to include more information and also adjusted for
technological and demographic changes that have taken place. Our comparisons are, therefore, only
based on 9 asset variables which are common in all data sets. Table 2.14 shows that the average MCA
asset index scores have increased over time from 0.67 in 1992 to 1.04 in 2010, respectively. The spread
as measured by standard deviation has declined suggesting a change in the distribution of asset index
scores over time.
Table 2.14: Descriptive statistics for asset index scores
Survey year
Observations
Average
Standard deviation
1992
5,323
0.67
1.14
2000
14,213
0.62
0.99
2004
13,664
0.76
0.86
2010
24,825
1.04
0.92
Total
58,025
0.84
0.97
Source: Own computation from MDHS 1992, 2000, 2004 and 2010
Figure 2.7 shows the poverty incidence curves drawn using the asset index scores for 1992, 2000, 2004
and 2010. Higher curves indicate higher levels of poverty. The curves indicate that at lower levels of the
MCA asset index scores, asset poverty is lowest in 2010, followed by 2004, 2000 and finally 1992. For
higher levels of the poverty line, the curves cross making it difficult for us to conclude in which year
poverty is higher.
47
Stellenbosch University https://scholar.sun.ac.za
.6
.4
0
.2
P
overtyheadcount
.8
1
Figure 2.7: Asset poverty incidence curves by survey year
0
2
4
6
MCA asset index scores
1992
2000
2004
8
10
2010
Source: Own computation from MDHS data
We also calculate differences in the incidence of asset poverty since 1992 with the poverty line set at
the 50.7th percentile (equal to an asset score of 0.807) as the cut-off point below which households are
considered poor. The null difference is that there exists no difference in household asset ownership
between any two chosen periods against the alternative hypothesis that there exist differences.
Table 2.15 shows that there has been an improvement in living standards in Malawi as shown by the
reduction in poverty incidence from 79.5% in 1992 to 50.0% in 2010. Poverty incidence increased by 1
percentage point from 79.5% in 1992 to 80.5% in 2000 but the increase was not statistically significant.
During this period, Malawi generally experienced a down turn in economic activity although we do not
place a causal link here. This worsening of long term economic conditions was, however, offset by the
improvements made over the three survey period from 2000 and 2010. The greatest improvement was
registered between 2004 and 2010 during which was a period of economic prosperity for Malawi.in the
country driven by strong growth in the agricultural sector13.
13
As noted earlier, the agricultural sector is one of the most important economic sectors in Malawi and it
contributes at least 30% to national GDP.
48
Stellenbosch University https://scholar.sun.ac.za
Table 2.15: Differences in poverty headcount indices for household asset ownership
Description
Estimate
Std. Err.
T
P>|t|
[95% Conf. interval]
Pov. Line
1992
0.795
0.006
143.76
0.000
0.784
0.806
0.807
2000
0.805
0.003
242.46
0.000
0.799
0.812
0.807
Difference
0.010
0.006
1.56
0.118
-0.003
0.023
---
2000
0.805
0.003
242.46
0.000
0.799
0.812
0.807
2004
0.785
0.004
223.15
0.000
0.778
0.792
0.807
Difference
-0.021
0.005
-4.26
0.000
-0.030
-0.011
---
2004
0.785
0.004
223.15
0.000
0.778
0.792
0.807
2010
0.500
0.007
70.05
0.000
0.486
0.514
0.807
Difference
-0.284
0.008
-35.73
0.000
-0.300
-0.269
---
1992
0.795
0.006
143.76
0.000
0.784
0.806
0.807
2010
0.500
0.007
70.05
0.000
0.486
0.514
0.807
Difference
-0.295
0.009
-32.65
0.000
-0.313
-0.277
---
Source: Own computation from MDHS 1992, 2000, 2004 and 2010
Table 2.16 provides the growth rate in the asset index and estimates from the five different pro-poor
indices discussed in Section 2.6. The results indicate that there has been both absolute and relative propoor growth in asset ownership in Malawi over the entire period of study from 1992 and 2010.
Table 2.16: Indices of pro-poorness in child nutritional status between 1992 and 2010
Pro-poor indices
1992-2000
2000-2004
2004-2010
1992-2010
Growth rate(g)
-0.076
0.217
0.445
0.626
Ravallion and Chen (2003) index
0.017
0.654
0.545
1.217
Kakwani and Pernia (2000) index
0.565
0.788
1.098
2.240
PEGR index
-0.043
0.171
0.489
1.402
Ravallion and Chen (2003) index – g
0.093
0.437
0.100
0.591
PEGR index - g
0.033
-0.046
0.044
0.776
Source: Own computation from MDHS 1992, 2000, 2004 and 2010
Malawi experienced a negative growth rate of 0.08% between 1992 and 2000. The negative growth
suggests that not only did poverty increase but the poor households were also negatively affected in
relative terms. Between 2000 and 2004, we find evidence of absolute but not relative pro-poor growth.
However, for the period between 2004 and 2010, we are able to conclude absolute but not relative propoor growth because the PEGR index – g is negative.
49
Stellenbosch University https://scholar.sun.ac.za
In Table 2.17 we show the mean access of assets by area of residence and region based on the pooled
data set. The table helps us to determine the driving factors behind the observed rural-urban and regional
differences in the levels of asset poverty in Malawi. Furthermore, this analysis is important because as
pointed out in Section 2.7, adjusting the asset index shifts the distribution to the right which may be a
limitation for inequality comparisons over time because the mean has changed (see, e.g., Wittenberg &
Leibbrandt (2015). Therefore, we use this as robustness checks for our pro-poor growth analysis results.
With respect to private assets, large differences in ownership are noted for “Bicycle”, “Electricity”,
“Car/truck” and “Motorcycle” between urban and rural areas. Most of the households from the Northern
region own “Paraffin lamps”. Ownership of toilet facilities is low in Malawi; even in urban areas, mean
access is at a dismal 14%. Traditional pit latrines remain the dominant form of toilet facilities. Decent
floor material (tile and cement) and excellent water sources are more accessible in urban areas.
50
Stellenbosch University https://scholar.sun.ac.za
Table 2.17: Pooled mean access of assets by area and region, all periods
Description
Average household size
Private assets
Radio
Bicycle
Paraffin lamp
Electricity
Car/truck
Motorcycle/scooter
Toilet facility
Own flush toilet
Shared flush toilet
Traditional pit toilet
Ventilated improved pit latrine
No facility/bush
Other
Floor material
Mud/earth
Cement
Bricks
Wood
Tiles
Other
Water source
Piped into residence
Public tap
Piped into yard/plot
Well in residence
Public well
Protected well/borehole
Spring
River/stream
Pond/lake
Dam
Rainwater
Other
Urban
4.64
Rural
4.63
Northern
5.01
Central
4.75
Southern
4.41
Total
4.63
73.40%
32.50%
61.70%
29.50%
6.60%
2.10%
51.00%
42.90%
52.10%
2.30%
0.70%
0.90%
55.20%
34.90%
65.60%
6.70%
1.50%
0.90%
52.50%
42.00%
52.40%
5.00%
1.60%
1.00%
55.30%
43.20%
50.20%
7.30%
1.60%
1.10%
54.30%
41.40%
53.60%
6.40%
1.60%
1.10%
14.00%
0.30%
54.20%
7.00%
3.40%
21.20%
0.70%
0.10%
44.30%
2.80%
17.40%
34.80%
2.40%
0.30%
44.70%
2.60%
14.90%
35.20%
2.50%
0.00%
45.20%
3.60%
17.20%
31.50%
2.90%
0.10%
46.50%
3.60%
14.10%
32.90%
2.70%
0.10%
45.70%
3.40%
15.30%
32.80%
37.50%
8.10%
0.10%
0.30%
52.70%
1.20%
85.40%
0.70%
0.00%
0.00%
11.60%
2.40%
77.10%
3.10%
0.20%
0.00%
17.70%
1.90%
79.80%
1.50%
0.00%
0.00%
14.80%
3.80%
77.50%
1.50%
0.00%
0.10%
19.70%
1.10%
78.20%
1.80%
0.10%
0.10%
17.70%
2.20%
13.40%
26.30%
37.40%
7.50%
1.60%
2.60%
1.90%
4.90%
3.00%
0.20%
0.10%
1.20%
0.60%
2.50%
9.90%
32.60%
3.80%
7.90%
3.70%
18.60%
10.90%
0.10%
1.20%
8.20%
2.70%
10.10%
15.70%
25.20%
5.00%
4.50%
2.70%
15.00%
7.00%
0.30%
1.10%
10.70%
1.90%
4.50%
8.70%
29.80%
4.00%
9.90%
5.00%
19.70%
9.30%
0.10%
1.00%
6.20%
2.90%
5.70%
17.10%
29.50%
2.50%
6.10%
2.60%
14.80%
10.90%
0.10%
1.10%
6.70%
2.50%
6.00%
14.00%
28.90%
3.40%
7.10%
3.40%
16.50%
9.70%
0.10%
1.10%
7.20%
Source: Own computation from MDHS 1992, 2000, 2004 and 2010
We also show in Figure 2.8 access to assets by type and survey period. This analysis explains the sources
of the movements in the asset index over time. The movements themselves could be due to so many
factors including technological changes and changing tastes and preferences. Furthermore, showing
access to assets is important because adjusting the asset index shifts the distribution to the right and this
51
Stellenbosch University https://scholar.sun.ac.za
may be a limitation for poverty comparisons over time because the mean has changed. We use this as
robustness checks for our pro-poor growth analysis results.
Between 1992 and 2010, ownership of radio and bicycle has expanded by 14 and 24 percentage points,
respectively. On the other hand, ownership of paraffin lamp has dropped by 50 percentage points from
84% in 1992 to 34% in 2010. Access to tile/cement floor and electricity has marginally increased by
around 1 percentage point over the period from 1992 and 2010. Ownership of car/truck and
motorcycle/scooter has stagnated at 2% and 1%, respectively. Ownership of flush toilet and access to
piped water into the dwelling has dropped by 2 percentage points.
Figure 2.8: Access to assets by type and survey period
Source: Own computation from MDHS 1992, 2000, 2004 and 2010; data labels for 1992 and 2010
2.21
Pro-poor growth in child nutritional status
Our analysis is based on HAZ owing to the fact that it is long term as already discussed. The HAZ scores
are calculated using the new WHO (2006) child growth standards and, therefore, directly comparable
over time. Table 2.18 provides summary statistics on HAZ in terms of the mean and standard deviation.
Table 2.18: Descriptive statistics for HAZ
Survey year
Observations
Average
Standard deviation
1992
3,288
-2.01
1.57
2000
9,396
-1.93
1.79
2004
8,309
-1.94
1.80
2010
4,653
-1.76
1.66
Total
25,646
-1.91
1.75
Source: Own computation from MDHS 1992, 2000, 2004 and 2010
52
Stellenbosch University https://scholar.sun.ac.za
Just as with the asset index, there has been an improvement in the average HAZ scores from -2.01 in
1992 to -1.76 in 2010. The deviation in the HAZ scores has slightly increased.
The poverty incidence curves given in Figure 2.9 indicate that incidence of poverty headcount is
unambiguously highest in 1992. For the rest of the years, the curves are compact except for specific
lower ranges of the z-scores.
.6
.4
0
.2
Poverty headcount
.8
1
Figure 2.9: Poverty incidence curves for HAZ by DHS survey year
0
20
1992
40
60
Height for age z-scores(HAZ)
2000
2004
80
100
2010
Source: Own computation from MDHS 1992, 2000, 2004 and 2010
Results for the calculated differences in the FGT headcount index over time and for successive survey
periods are given in Table 2.19. As before, the poverty line is set at 2.3rd percentile which corresponds
to the international cut-off of -2SD proposed in the WHO (2006) standards below which children are
considered malnourished. The null hypothesis is that there exists no statistical difference in the HAZ
scores between any two chosen periods against an alternative that there exist differences.
The results show that there has been an improvement in the nutritional status of children aged below 5
years in Malawi during the period between 1992 and 2010; the incidence of stunting levels amongst
under-five aged children in Malawi has decreased from about 54% in 1992 and to about 46% in 2010.We
also find significant improvements in the levels of child nutritional status for the successive survey
periods apart from the 1992 and 2000 pair where the reduction in the incidence of malnutrition is not
statistically different from zero.
53
Stellenbosch University https://scholar.sun.ac.za
Using the asset index as a measure of welfare, results indicated that poverty increased between 1992
and 2000 but not statistically significant. Therefore, it is concluded that the results are similar and that
both asset poverty and child nutritional status remained the same over the period.
Table 2.19: Differences in the FGT poverty headcount index for HAZ
Description Estimate Std. Err.
1992
0.539
0.011
2000
0.519
0.006
Difference -0.020
0.013
2000
0.519
0.006
2004
0.496
0.007
Difference -0.023
0.009
2004
0.496
0.007
2010
0.457
0.000
Difference -0.039
0.007
1992
0.539
0.011
2010
0.457
0.000
Difference -0.082
0.011
T
49.174
82.173
-1.601
82.173
70.046
-2.430
70.046
.
-5.484
49.174
.
-7.491
P>|t|
0.000
0.000
0.111
0.000
0.000
0.016
0.000
.
0.000
0.000
.
0.000
[95% Conf. interval] Pov. line
0.517
0.561
2.300
0.506
0.531
2.300
-0.045
0.005
--0.506
0.531
2.300
0.482
0.510
2.300
-0.042
-0.004
--0.482
0.510
2.300
0.457
0.457
2.300
-0.053
-0.025
--0.517
0.561
2.300
0.457
0.457
2.300
-0.104
-0.061
---
Source: Own computation from MDHS 1992, 2000, 2004 and 2010
We also find evidence of both absolute and relative pro-poor growth in child nutritional status for 20002004, 2004-2010 and 1992-2010 as shown in Table 2.20. Not only did absolute poverty decline but the
poor also benefited more. For 1992 and 2000, the Ravallion and Chen (2003) index gives a negative
estimate of 0.203 thereby making us unable to conclude absolute pro-poorness despite the fact the
Kakwani and Pernia (2000) index and PEGR index indicate otherwise. We are also unable to conclude
relative pro-poorness using all the three relative measures, namely the Ravallion and Chen (2003) index
– g, PEGR index- g and the Kakwani and Pernia (2000)'s index.
Table 2.20: Indices of pro-poorness for HAZ
Pro-poor indices
1992-2000
2000-2004
2004-2010
1992-2010
Growth rate(g)
0.223
0.028
0.044
0.313
Ravallion and Chen (2003)
-0.203
0.360
0.946
1.042
Kakwani and Pernia (2000)
0.624
8.547
7.702
2.965
PEGR index
0.139
0.241
7.702
0.928
Ravallion and Chen (2003) – g
-0.426
0.332
0.902
0.729
PEGR - g
-0.084
0.213
0.295
0.615
Source: Own computation from MDHS 1992, 2000, 2004 and 2010
54
Stellenbosch University https://scholar.sun.ac.za
2.22
Conclusions
In this study, we have measured poverty and inequality in Malawi both spatially and temporally. Over
the period between 1992 and 2010, we find evidence of pro-poor growth in both absolute and relative
terms. We have also identified factors associated with asset poverty in Malawi. Results from the study
indicate that while household size is positively correlated with asset accumulation, age dependency ratio
has a negative association. There is no statistically significance difference in asset ownership between
male and female headed households. Households with older household heads do better in terms of asset
ownership than those headed by younger people. Shock to sickness in the household has a negative
association with asset ownership. Education attainment by the members of the household has been found
to have a positive association with the level of assets in the household. Rural areas have higher levels of
asset poverty compared to urban areas. Significant regional differences between the northern and central
regions have also been found in terms of asset poverty with the latter lagging behind.
We have also identified factors that are associated with the nutritional status of under-five children in
Malawi. Using three anthropometric measures of nutrition, namely HAZ for stunting, WAZ for
underweight and WHZ for wasting, it is shown that the incidence of stunting is the highest of the
nutritional problems amongst under 5 children in Malawi followed by underweight. On the one hand,
there no large differences between regions and areas in terms of child nutritional status. On the other
hand, when assets are used, the welfare gap between regions and areas is bigger.
2.23
Policy discussion
Non-monetary pro-poor growth analysis is important for policy because it allows us to identify the
progress made in poverty reduction over time as well as analyse the redistribution effects that arise with
growth in living standards. Specifically, as shown in the paper, the poor people in Malawi have benefited
the most from the changes in assets and child-nutritional status that have occurred over time. This means
that policy makers should not only focus on income growth when analysing progress towards goals set
in the national development agenda such as those in the Malawi Growth and Development Strategy
(MGDS) and the Sustainable Development Goals (SDGs).
Subgroup poverty comparisons are important for targeting development policies towards specific groups
that have been identified as the poorest. For example, our study has shown that poverty is higher in rural
areas than urban areas. As a response to these findings, deliberate policy measures might be put in place
to specifically target the identified poorest groups of Malawi. Similarly, policies can be developed to
address the problem of inequality as informed by the results.
A number of policy conclusions can be also be drawn from the regression analysis. Firstly, the negative
association between asset ownership and age dependency ratio suggests that increasing the income
55
Stellenbosch University https://scholar.sun.ac.za
generating opportunities for the economically active population could help in reducing asset poverty in
Malawi. Younger household members could, for example, be equipped with the necessary skills as they
wait to enter the working age group. Secondly, the significance of education attainment and the fact that
the sizes of coefficients increase with the level of education imply that post-primary education should
be emphasised in Malawi. Currently, the focus of education policy is on basic education (completion of
primary education). Thirdly, since shocks to illnesses are negatively associated with asset ownership, it
may be important to consider looking into policies that improve access to health and assist households
to withstand shocks thereby preventing them from being pushed into poverty traps.
The multivariate analysis of child malnutrition reveals that characteristics such as age, sex and twin
status matter in the outcomes of nutritional status at the child level. Child birth order, mother education
and economic status have a positive association with child nutritional status. These findings have policy
implications too. Firstly, based on the literature linking recovery from stunting to cognitive outcomes,
the study suggests that timing of nutritional inputs is critical in the cognitive development of children.
Secondly, since male children tend to have weaker nutritional status compared to their female
counterparts, nutritional feeding programmes might consider addressing this gender imbalance. Thirdly,
the importance of female education for the nutritional status of children suggests that training women in
child nutrition may require emphasis. Fourthly, related to the second implication, the significance of
higher levels of education suggests that post-primary education should be emphasised in general.
Currently, the focus of education policy is to achieve basic or primary education. Finally, since higher
levels of economic status are associated with better nutritional status, policy should continue to be geared
towards improving households’ living conditions.
56
Stellenbosch University https://scholar.sun.ac.za
Chapter 3
Externalities and returns to education in Malawi: Panel data evidence
3.1 Introduction
Malawi remains a poor country despite registering gains in poverty reduction over the past two decades.
Estimates based on Malawi’s first, second and third integrated surveys (IHS1, IHS2 and IHS3) show
that the incidence of poverty based on household percapita consumption has marginally fallen from
52.4% in 2005 to 50.7% in 2011. In 1998, the poverty rate was 65.3%. Rural areas, which make up
about 85% of the population, have significantly higher poverty rates than urban areas, although the gap
between the two is closing over time. The Gini coefficient shows that inequality increased from 0.401
in 1998 to 0.390 in 2005 to 0.452 in 2011 (National Statistical Office, 2012).
On the one hand, the labour participation rate for Malawi, defined as the share of the population aged
15 and above working or seeking employment, is very high and stands at 88%. On the other hand,
education levels are low. About 74% of the population aged 15 years and above do not have any
qualification at all and 21% reported never having attended education. Therefore, literacy remains a
challenge in Malawi given the high rate of labour force participation. As at 2011, the literacy rate
(defined as the ability to read and write with understanding in any language) amongst people aged 15
years and above stood at 65%, which is an insignificant improvement from 64% reported in 2005
(National Statistical Office, 2012).
In the Malawi Growth and Development Strategy (MGDS), education is identified as one of the themes
necessary for growth and socio-economic development. Malawi’s formal education system consists of
primary, secondary and tertiary or post-secondary education. The country’s education policy has been
largely focussed towards increasing access to primary education. In recognition of this, primary school
education was universally made free in 1994 for all government schools. According to National
Statistical Office (2012), as a response to these policy changes, the net primary enrolment rate has
increased to 86% in 2011 from 80% in 2004 while the primary dropout rate has dropped from 5% to 1%
over the same period. In addition to universal free primary education, tertiary education has been
subsidised in order to make it more affordable, although the country is yet to realise the benefits in terms
of expansion in tertiary education.
The Malawian labour market can be categorised into the formal and informal sectors. As is the case in
many other developing countries, the formal sector in Malawi only consists of a small percentage of the
labour force. Consequently, most people are involved in either self-employment activities or paid
employment in the informal sector (Chirwa & Matita, 2009). The informal sector is, therefore, important
to Malawi’s economy; and it accounts for 78% of total employment (National Statistical Office, 2014).
57
Stellenbosch University https://scholar.sun.ac.za
Education has been identified as a tool for poverty reduction in Malawi (National Statistical Office,
2012). Through education attainment, the poor are said to be empowered and equipped for better
opportunities in national development.
It is against this background that the study seeks to look at the role of education in poverty reduction in
Malawi. The link between poverty and education is identified through the labour market. A number of
studies that link education to poverty reduction have been conducted. Da Maia (2012) looks at the link
between education and poverty reduction in Mozambique. The study estimates the probability of an
individual getting employment in any of the given sectors conditioned on education and models the
relationship between education and earnings. Several researchers have looked at the role played by
education in earnings (e.g., Mincer, 1974; Psacharopoulos, 1994 & 2002; and in Malawi, Chirwa &
Matita, 2009; Chirwa & Zgovu, 2002). Clearly, the relationship between education and earnings is not
only of great interest to many scholars but also important for a policy perspective.
However, previous studies on Malawian labour markets (see Chirwa & Matita, 2009; Matita & Chirwa,
2009; Chirwa & Zgovu, 2002) have explored the link between education and earnings using crosssectional data sets due to lack of panel data. This study expands on the available literature by taking
advantage of the newly released panel data set. Use of panel data has many advantages and this is well
acknowledged in the literature. Firstly, we can control for unobservable individual heterogeneity. As
shown in the literature, failure to control for individual specific effects leads to bias in results. Secondly,
panel data contains rich information about cross-sectional variations and dynamic behaviour of the
subjects of interest. Thirdly, with panel data, we are able to identify time effects which cannot be
estimated with cross-sectional data. Baltagi (2013) provides a detailed discussion on the advantages and
limitations of panel data.
In addition to using panel data, the study distinguishes between the formal and informal sectors of the
economy to see if the role played by education in earnings differs by sector. This decomposition is
important considering the large size of the informal sector. Furthermore, the study corrects for sample
selection bias which arises as a result of selection into economic sectors as well as working with only
selected individuals who end up entering employment instead of the full working age-population. With
regard to distinguishing between sectors, we improve on previous studies (e.g. Chirwa & Matita, 2009)
by conducting coefficient comparison tests between groups by gender and sector. The test is routinely
implementable in Stata 13.1 via the suest command (see Appendix A1). The need for conducting
coefficient comparison and balancing tests has been motivated in Pischke and Schwandt (2014).
The study also explores whether earnings from non-farm household enterprises are affected by the
educational attainment of other members of the household. Basu and Foster (1998) argue that the
58
Stellenbosch University https://scholar.sun.ac.za
presence of a literate person in a given household may generate positive intra-household education
externalities. Unlike the isolated illiterate persons, illiterate individuals living in a household where at
least one member is literate enjoy some of the literate person’s capabilities. Therefore, education is said
to have some positive spill-over effects in that the literate members of the household may help in the
running of enterprises in the same manner agricultural extension workers or medical personnel help in
their communities. We explore whether there are positive education externalities in household
enterprises by the use of maximum number of years of schooling in a household. This specification is
also motivated by Jolliffe (2002) who finds that in household level income functions such as non-farm
enterprises, the maximum or average level of education in the household is a better explanatory variable
of household income. Furthermore, it is argued that in most developing countries income is largely
earned at the household rather than the level of the individual, hence our focus on household enterprises.
Moreover, according to data from IHS3, the enterprise sector is large in Malawi and affects a significant
percentage of households in Malawi. Specifically, about 20% of households are involved in non-farm
employment activities in Malawi (National Statistical Office, 2012).
The chapter addresses three main specific issues (i) to investigate the factors influencing labour force
participation and employment likelihood in Malawi; (ii) to estimate returns and externalities to education
in Malawi; and (iii) to identify and deal with inconsistencies in the IHPS data. This includes conducting
comparisons of earnings and consumption data in addition to robust treatment of outliers.
3.2 Methodology
3.2.1
Theoretical framework
There are two main competing groups of theories for explaining labour market outcomes. First is the
traditional neoclassical model of labour supply which argues for a simple competitive labour market
whereby workers are paid their marginal product. In this competitive model of the labour market,
individuals’ wages are largely determined by their productivity (Kerr & Teal, 2015). According to this
model, wage differences across sectors or occupations are competitively eliminated by workers moving
away from underpaid sectors to highly paid sectors. One of the early works explaining occupational
choice is by Roy (1951). Using a two-sector model, Roy (1951) argued that individuals choose sectors
in which they have a comparative advantage such that wage differences between sectors are simply a
reflection of the differences in unobserved ability between employees in the two sectors. However,
investments in education and training were not modelled in Roy’s model.
The simplest form of the competitive model does not distinguish between degrees of job informality or
differences in individual investments that would enhance productivity. However, by allowing for
investment in human capital, individuals are assumed to make investments in education and other
productivity enhancing abilities in such a way as to maximise their expected lifetime utility. Despite
59
Stellenbosch University https://scholar.sun.ac.za
contributions by other researchers, much of the human capital theory is based on the work of Becker
(1962) and Mincer (1974) after which the human capital theory became the standard for empirical work
in labour economics (Card, 1999). In this strand of literature, earnings are assumed to be dependent on
education or training and experience in the labour market
The alternative is the segmented labour market hypothesis which observes labour market outcomes and
wage differentials that are inconsistent with the competitive model. Specifically, the latter argues that
people’s choices of where to work and the wage received are determined by institutional factors such as
government legislation (e.g. minimum wage legislation) and other institutional factors (see (Leontaridi,
1998; Dunlop, 1957; Kerr, 1954). It is, for example, argued that labour market institutional constraints
prevent the bargaining down of wages in high wage sectors by workers in low-wage sectors.
Furthermore, it was observed that workers with similar jobs were paid different wages as a result of such
institutional constraints. A number of explanations have been given for the existence of institutional
constraints and wage setting mechanisms between sectors. One of these explanations in the developing
country context is the Harris-Todaro model of migration (Harris & Todaro, 1970). According to this
model, while the urban sector is insulated from the forces of supply and demand as a result of minimum
wage legislation, the rural sector was assumed competitive. The existence of a minimum wage in the
urban sector meant that rural workers or the unemployed were unable to bid the urban wage down to an
equilibrating level. Under these conditions, this ensured that workers who find employment in the urban
sector would therefore be paid more than if they were employed in the rural agricultural sector.
Despite these criticisms, the human capital theory remains the leading theoretical model for explaining
labour market outcomes and is used in this study. In addition to these two competing theoretical
frameworks, recently, there have also been efforts towards improved understanding and definitions of
informal markets as well as the linkages between the informal and formal economy. This has given rise
to different theories or explanations as to what constitutes or gives rise to informal labour markets (Chen,
2012). This debate is relevant in the Malawi context because of the huge size of the informal
employment, which makes up of 78% of total employment, as earlier stated.
The recognition of the existence of an informal market in the analysis of labour markets in developing
countries largely came as a result of the 1972 International Labour Organization (ILO) study in Kenya
(Singer & Jolly, 1972) where it was found that the traditional and informal sectors of the economy
included profitable and efficient enterprises as well as marginal activities. In addition to this, work by
British anthropologist Hart (1973) found that there was no wage employment among unskilled migrants
who had migrated from Northern Ghana into the capital city. Instead, these were involved in low-income
activities in the informal sector not related to the formal sector. Fields (1975) made another contribution
to the empirical and theoretical research on the informal sector by extending the Harris-Todaro
60
Stellenbosch University https://scholar.sun.ac.za
framework to include an urban informal sector. In this framework, three sectors are recognised, namely
the urban formal sector, the urban informal sector and the traditional informal sector. The urban informal
sector was made up of casual employment and was viewed as a temporary sector in which people freely
enter as they look for higher paid jobs in the formal sector. However, the presence of union activity and
minimum wage legislation prevented people from entering the formal sector and therefore kept workers
in the informal sector (Kerr & Teal, 2005).
The definition of the informal sector has changed over the years. Previously, following the ILO Kenya
report, the characteristics of informal activities included ease of entry, family ownership of enterprises
and unregulated and competitive markets (Singer & Jolly, 1972). ILO now uses the expanded statistical
criteria for distinguishing between formal and informal employment which was endorsed by the 2000
International Labour Conference and the 2003 International Conference on Labour Statisticians
(Hussmanns, 2004). The formal sector primarily includes salaried employment in the private and
government sectors as well as non-governmental organisations (NGOs). In this sector, the relationship
between the employer and employees is governed by formal labour laws including employee benefits
and income taxation. In contrast, the informal sector has two components, namely self-employment
activities and informal wage employment.
Self-employment mainly consists of employers in informal enterprises, own account workers in informal
enterprises, contributing family workers (in informal and formal enterprises) and members of informal
producers’ cooperatives. Informal self-employment includes enterprises that are not registered under
any national legislative authority and not engaged in agricultural activities.
Informal wage employment includes employees of informal enterprises, casual or day labourers,
temporary or part-time workers, paid domestic workers, contract workers, unregistered or undeclared
workers and industrial outworkers (also called homeworkers). Of great importance to the Malawian
economy is casual employment (locally known as ganyu14). Despite being largely seasonal, it is very
important in both urban and rural areas.
14
Ganyu is the dominant form of employment in the informal sector.
61
Stellenbosch University https://scholar.sun.ac.za
3.2.2
Estimating returns to education
Mincerian earning functions are the standard approach for estimating returns to education in labour
markets. The methodology started with the work of Mincer (1974) and has been widely used in the
literature (Psacharopoulos, 1994). It is based on the human capital theory which argues that investment
in education improves workers’ skills resulting in high productivity and, therefore, higher earnings
(Mincer, 1974). The basic model is summarised below:-
ln(Yit ) 1 2 S it 3 E it 4 E it2 it
(3.1)
where, for any individual i , at time t , Y is the earnings of individual, S is the number of years of
2
schooling, E is the experience, E is experience squared and is the error term. The coefficient 2 is
interpreted as the private rate of return to education (RORE) and 2 * 100 gives the percentage return
to one additional year of schooling. The above specification can be extended to include other control
variables such as gender, location, etc.
As is the practice in the literature, we improve on the classic model presented in equation (3.1) in two
main ways. Firstly, we account for the fact that returns to education may be heterogeneous rather than
homogeneous and secondly, we address the problem of selection bias which arises as a result of using
non-random data for analysis. Due to its importance, we will deal with sample selection in the section
that follows.
As stated, the basic model disregards the differences in the level of educational achieved by looking at
a single overall education level or years of schooling. This approach is called the one factor or
homogeneous model since it assumes that there are no differential trends in the returns to education for
different levels of education. There is little statistical evidence and causal empiricism for the
homogenous model. The alternative approach, called the multiple factor approach or heterogeneous
model, looks at the different levels of education as having separate effects on earnings. This involves
replacing S with an educational dummy variable to represent the different educational categories. Seven
educational qualifications are captured in the data, namely “None”, “Primary School Leaving Certificate
of Education (PSLCE)”, “Junior Certificate of Education (JCE)”, “Malawi School Certificate of
Education (MSCE)”, “Non-University Diploma”, “University Diploma” and “Post-Graduate degree”.
In some cases of our analysis, we combine the last three categories into tertiary education 15.
15
In Malawi, PSLCE, JCE and MSCE are equivalent to 8, 10 and 12 years of schooling. For tertiary education,
diplomas are usually completed within two or three years while undergraduate degrees take four years except for
62
Stellenbosch University https://scholar.sun.ac.za
Our data set does not have information on the number of years of experience. Following the standard
practice used in the literature (e.g., Matita & Chirwa, 2009; Kahyarara & Teal, 2008; Appleton, Bigsten,
& Manda, 1999), we estimate the number of potential years of experience as age less years of schooling
less preschool age. This assumes that once people complete their education, they immediately enter the
labour market. Those without education are assumed to enter the labour market at the lowest labour
market entry age of 15 years.
3.2.3
Sample selection
Sample selection16 bias arises from a number of sources. Firstly, it could arise as a result of self-selection
into different employment categories or sectors. Secondly, it could be due to non-random attrition in
panel surveys when subjects drop out of the sample for some reason. However, in our data, attrition is
not a major problem, as we will discuss in Section 3.3 on data. Consequently, we do not deal with sample
selection caused by attrition. Thirdly, it may be as a result of working with a truncated sample. This
raises the problem of sample selection because ideally we are interested in studying all individuals in
the working age population (15-64) but end up with only those that actually entered employment. Fourth
and finally, since earnings are only observed for individuals who are working at the time of the survey,
our analysis is made on the sample selected on this basis.
Failure to correct for sample selection would bias our results. The Heckman (1979) two-step procedure,
following Wooldridge (2002), has been used to correct for selection bias. In the first step, the probability
of an individual selecting into an economic sector or attriting through a probit model is estimated as:
cit xit vit
(3.2)
where: i represents an individual, c denotes the binary response variable equal to either 1 or 0 and x is
a set of regressors such as education level, age, sex, marital status, etc., including dummies for location.
The error term is given by v .
From equation (3.2), we obtain the inverse of the Mills ratio and use it as an explanatory variable in the
estimation of either the wage equation (3.1) or the likelihood of employment conditional on an
engineering, law and medicine which are completed in five years. A post-graduate degree usually takes two years
to complete.
16
Some may refer to this as self-selection. Whether we define it as sample selection or self- selection is trivial in
our study as long as we achieve our goal of addressing the non-random nature of the sample when estimating our
employment and earnings functions.
63
Stellenbosch University https://scholar.sun.ac.za
individual’s participation in the labour market (see Section 3.6.2). If the coefficient of the inverse Mills
ratio is statistically significant, then we are justified in correcting for selection bias. We then proceed to
correct for and report heteroskedasticity consistent standard errors in the first stage.
3.2.4
Modelling unobserved heterogeneity
Let us consider the following model when time T 1,2 :
yit xit uit
(3.3)
Where y it is the log of earnings for individual i at time t . Suppose that the error term is made up of
two components as follows:
u it ni vit
(3.4)
where ni is time invariant and correlated with xit , and vit is time varying and uncorrelated with xit .
If the exogeneity assumption is violated, i.e. when E ( xit ni ) 0 , the OLS estimator will be biased in
cross-section. On the other hand, panel estimators can be used to control for unobserved individual timeinvariant heterogeneity and this allows us to obtain unbiased estimates of . Even when the unobserved
correlated effect is not time invariant, using panel data techniques reduces the magnitude of the bias.
The most commonly estimated models with panel data are the fixed effects and random effects models
and several considerations will affect the choice between the two. The first consideration is the nature
of omitted variables. If we think that there are no omitted variables from the model or that the omitted
variables are uncorrelated with the explanatory variables in the model, then a random effects model is
the best (Williams, 2015). A random effects model under these assumptions will produce unbiased
estimates of the coefficients, use all the data available, and yield the smallest standard errors. However,
it is more likely that omitted variables will produce at least some bias in the estimates. If there are
omitted variables and these variables are correlated with the variables in the model, then a fixed effects
model provides a means for controlling for omitted variable bias. In a fixed-effects model, subjects serve
as their own controls. For this to work, the omitted variables must have time-invariant values with timeinvariant effects. For example, gender does not change over time and its effect on the outcome in wave
1 is the same as the effect of gender in wave 2.
Secondly, researchers consider the amount of variability within subjects: If subjects do not change much,
or not at all, across time, then fixed effects models may not work very well or even at all. There is need
to have within-subject variability in the variables to justify the use of fixed effects. When there is little
64
Stellenbosch University https://scholar.sun.ac.za
variability within subjects, the standard errors from a fixed effects model may be too large to tolerate.
Conversely, random effects models will often have smaller standard errors (Clark & Linzer, 2015).
Third are the effects are we interested in studying. In fixed effects models, we are not interested in
estimating the effects of variables that do not change or change very little over time. Rather, we control
for them or “partial them out”. On the other hand, with random effects models, we are able to estimate
the effects of time-invariant variables such as gender, although the method is no longer controlling for
omitted variables (Williams, 2015).
Given the above considerations, we choose to use the random effects model for three main reasons.
First, education, whose effect we are interested in measuring, is generally a slow changing variable
especially for a three year period over which we have data. Secondly, given that our panel is short (only
two periods) there is not much within-subject variability in most of our variables. Third, a random effects
model allows us to estimate the effects of time-invariant variables such as gender which is an important
aspect in Malawi. With fixed effects, this is not possible since the variable gets dropped off after
demeaning.
Therefore, using a random effects model naturally comes at a cost and the trade-off is that their
coefficients are more likely to be biased than the fixed effects estimates. Nevertheless, according to
Wooldridge (2002), panel data techniques reduce the magnitude of bias compared to OLS.
3.3 Description of the data
This section provides a brief description of the Malawi Integrated Household Panel Survey (IHPS) data,
a two-wave panel conducted in 2010 and 2013. The survey was implemented by the National Statistical
Office (NSO) of Malawi. The 2010 wave was part of the third nationally representative Integrated
Household Survey (IHS3)17 conducted between March 2010 and March 2011 during which 3247
households were selected as a panel subcomponent to be resurveyed in 2013. The second wave, carried
out between April and December 2013, saw the panel sample increase to 4,000 households because
split-off members who formed new households were also included.
17
IHS3 consists of four questionnaire types, namely the household, agricultural, fisheries and community
questionnaires. In the household questionnaire, individuals are asked if they were involved in agricultural and
fishing activities over the past 12 months. Those who answered “Yes” are then further administered the agriculture
and fisheries questionnaire where more details on crop, livestock and fishing practices are collected. Our analysis
is based on the household questionnaire only where the scope of our study lies.
65
Stellenbosch University https://scholar.sun.ac.za
The household formed the primary unit of analysis in the IHPS surveys. An attempt was, therefore,
made to track all baseline households as well as members that moved away from the baseline dwellings
between 2010 and 2013. Servants and guests at the time of the IHS3 were excluded and only individuals
who were expected to be at least 12 years of age and known to be residents in mainland Malawi 18 were
tracked. Of the 3,247 households initially chosen in 2010, some could not be located while others split
into new households. Our analysis is based on the 3,104 households that are available in both waves.
The rate of attrition at the household level was only 3.78 %.
The 2010 baseline had 15,597 individuals of all ages, of which 14,232 are available in both waves,
representing an overall attrition rate of 7.42 % at the individual level. Given these low rates of attrition,
which also seem random, we pursue this issue no further because we believe the representativeness of
the sample has not been affected19. For purposes of this study, our sample is restricted to economically
active available in both waves. All calculations include survey weights.
There are two sources of earnings as captured in the household questionnaire, namely wage employment
and self-employment activities. Earnings from self-employment activities are provided as profit from
non-farm enterprises over a period of 30 days while earnings from wage employment are given with an
indication of the period over which earnings are earned, i.e. day, week, two weeks or month. Ganyu
wages are given as daily earnings with an indication of the number of days worked in a week. All
earnings are converted into real monthly figures in 2013 constant prices.
3.3.1
Work and non-work activities of the employed
The surveys collected information on both work and non-work activities of individuals over the seven
days prior to the administration of the questionnaire. Non-work activities or domestic tasks are given in
(a) and (b) while the rest constitute work activities.
a. How many hours did you spend yesterday collecting water?
b. How many hours did you spend yesterday collecting firewood (or other fuel materials)?
c. How many hours in the last seven days did you spend on household agricultural activities
(including livestock and fishing-related activities) whether for sale or for household food?
d. How many hours in the last seven days did you run or do any kind of non-agricultural or nonfishing household business, big or small, for yourself?
18
Excluding Likoma district, which is an Island on Lake Malawi.
19
In the literature, the most common ways of addressing attrition are Inverse Probability Weighting (IPW) and
Heckman selection correction (e.g., Wooldridge, 2002).
66
Stellenbosch University https://scholar.sun.ac.za
e. How many hours in the last seven days did you help in any of the household's non-agricultural or
non-fishing household businesses, if any?
f.
How many hours in the last seven days did you engage in casual, part-time or ganyu labour?
g. How many hours in the last seven days did you do any work for a wage, salary, commission, or
any payment in kind, excluding ganyu?
h. How many hours in the last seven days did you engage in an unpaid apprenticeship?
All individuals and households involved in household agricultural activities identified from (c) are
administered the household questionnaire where more information related to agriculture is collected.
Questions (d) and (e) are further explored in the enterprise module where information on enterprise
owners, customers and profits is collected. Questions (f) and (g) are further explored in the module for
time use and labour where information on wages and occupations is collected.
3.3.2
Describing employment structure and hours worked
There are about nine mutually exclusive employment types that can be identified in the data set as shown
in Table 3.1. Public works programme (PWP) and church/religious organisations can be jointly referred
to as non-governmental organisations (NGO), considering the nature of their activities.
Table 3.1: Employment and occupation structures
2010
Occupations
2013
Frequency
Percent
Cumulative
Frequency
Percent
Cumulative
Private Company
270
8.9%
8.9%
303
8.0%
8.0%
Private Individual
261
8.6%
17.4%
225
6.0%
14.0%
Government
170
5.6%
23.0%
173
4.6%
18.6%
Parastatal
14
0.5%
23.5%
15
0.4%
19.0%
PWP
12
0.4%
23.9%
16
0.4%
19.4%
NGO
45
1.5%
25.3%
63
1.7%
21.1%
Other
6
0.2%
25.5%
18
0.5%
21.6%
1,620
53.2%
78.7%
2,056
54.5%
76.0%
650
21.3%
100.0%
904
24.0%
100.0%
3,048
100.0%
3,773
100.0%
Casual/Ganyu
Enterprises
Total
Source: Own computation from IPHS data
On the one hand, the structure of employment observed in the table reflects a major similarity that
Malawi shares with other less developed African countries, i.e. that casual employment and selfemployment activities (household enterprises) make up the largest share of total employment both in
67
Stellenbosch University https://scholar.sun.ac.za
and outside of agriculture. This observation holds in both 2010 and 2013. Moreover, attrition levels are
very low such that the structure observed in the panel is largely a reflection of the national figures. On
the other hand, the observed structure is a contrast to what is seen in countries like South Africa where
there is a very small informal sector (e.g., Kingdon & Knight, 2004; Kerr & Teal, 2015).
The dominance of the informal sector is reflected in the number of observations and accompanying
percentages in each of the employment types. Regular formal employment (private and government)
have registered slight declines and possibly these partly account for the growth in the informal
employment numbers. Total informal employment has increased as a result of growth in both casual
employment and self-employment activities. The IHPS report states that the percentage of households
in Malawi that operate non-agricultural enterprises has grown by 9 percentage points from 21% in 2010
to 30% in 2013.
The survey collected information on the number of hours worked20 during the 7 days prior to the survey.
We provide a breakdown of the average number of hours by occupation in Table 3.2.
The table reveals that, on average, individuals in the panel worked for about 25 hours in a week in 2013.
This is an insignificant increase from the 24 hours registered in 2010. The maximum working hours
regulated by the government is 48 per week in all jobs. Most of the occupations registered increases in
the number of hours worked. PWP and ganyu employment registered significant declines. The former
is a safety net scheme targeting the poor who mainly utilise it when there is a need. The latter employees
(ganyu) usually work for food and basic needs and their labour supply is largely dependent on meeting
daily needs. Once their daily needs are met, there is little motivation to work longer. Overall, the private
sector (companies and individuals) worked the longest number of hours in both 2010 and 2013. The
government and parastatal worked for a similar number of hours.
Table 3.2 also shows the average years of schooling by occupation. On average, number of years of
schooling is about 7.5 and 7.4 years for 2010 and 2013, respectively. This is just below the official 8
years of primary education. What we find here is reflective of what the focus of government education
policy has been in Malawi over the past two decades- the provision of free primary education.
Government employees have the highest number of years of schooling across the two years while ganyu
and public programme workers have the least. The levels of education partly explain the large earning
differentials observed earlier, consistent with the human capital theory which states that education is
positively correlated with earnings.
20
Includes agricultural work but excludes unpaid domestic activities: water and firewood collection.
68
Stellenbosch University https://scholar.sun.ac.za
Table 3.2: Changes in the average weekly hours worked and years of education by year
2010
Average weekly Years of
hours worked education
Private Company
40.20
8.80
(25.50)
(4.50)
Private Individual
32.80
7.50
(24.60)
(3.10)
Government
32.90
13.40
(23.10)
(5.10)
Parastatal
36.30
11.20
(21.0)
(4.80)
Public Works Programme
26.20
7.70
(27.70)
(3.50)
Church/Religious Organisation
30.90
10.50
(23.80)
(5.30)
Other
47.30
10.90
(24.40)
(3.60)
Casual/Ganyu
16.40
6.50
(17.60)
(2.60)
Non-agricultural enterprises
29.61
7.43
(21.40)
(3.40)
Total
24.10
7.50
(22.10)
(3.70)
Occupations
2013
Average weekly Years of
hours worked education
44.10
9.00
(23.60)
(04.50)
41.30
7.70
(28.0)
(3.30)
34.10
14.50
(23.20)
(5.60)
35.20
10.60
(29.30)
(5.20)
13.60
5.80
(15.10)
(3.30)
40.30
10.80
(27.0)
(5.50)
30.00
9.40
(17.80)
(4.20)
15.10
6.40
(18.10)
(2.70)
33.28
7.42
(22.10)
(3.40)
24.80
7.40
(23.50)
(3.80)
Source: Own computations from IHPS data; standard deviations in parenthesis.
3.3.3
Treatment of outliers, missing data and zero earnings
Missing earnings, for example, because individuals refused to answer or the respondent did not know,
are not imputed. Negative and zero earnings were dropped since their natural logs are undefined21. As
is the practice in the literature, we only look at economically active individuals aged between 15 and 64
years. We also restrict our sample to observations with non-missing values in any of the variables
reported in our tables. The resulting samples are 5,377 individuals and 1,702 households with positive
real earnings. Next, we explain how outliers have been treated, using wage earners as an example.
Observations that are substantially different from the rest can make a difference to the regression results
obtained. It is, therefore, important to not only investigate these unusual observations but also find ways
of dealing with them. Wittenberg (2014) and Burger and Yu (2007) provide a good discussion on dealing
with outliers. We discuss four approaches in the ensuing paragraphs.
21
Researchers sometimes consider an alternative of adding a very small number to earnings before taking the log,
which would allow considering zero incomes.
69
Stellenbosch University https://scholar.sun.ac.za
The first approach is to take out millionaires which are clear outliers. In our data set, there were 24
millionaires with average monthly earnings of MWK2,269 702 compared to MWK 38,471 for the rest
of the 5353 individuals. However, the choice of millionaires is arbitrary and has a potential of removing
genuine earners, especially considering that the minimum education level of these millionaires is
MSCE- above the typical education level of those in the labour force.
The second approach is to remove outliers by identifying observations with extreme regression
residuals. In linear regression, an outlier is an observation with a large residual. This is achieved by
estimating a simple Mincerian type wage regression of the log of real monthly wages on education, age,
age squared, gender and occupation. After running this regression, studentised or standardised residuals
were created. In this approach, studentised residuals with absolute values greater than five are flagged
as extreme and corresponding observations dropped. Using this approach, only five observations were
flagged as extreme. We reduced the cut-off to 4 resulting in 14 individuals being flagged as having
earnings that were too high or low for their characteristics.
The third approach is robust regression. When data is contaminated with outliers, using studentised
residuals has been found to be insufficient in identifying the ‘bad’ observations. Robust regression is an
alternative to least squares regression when this is the case, i.e. when data is contaminated with outliers
or influential observations (Wittenberg, 2014). This approach is easily handled in Stata 13.1 and
observations are given weights depending on whether they are outliers or not. Outliers are assigned zero
weights and consequently identified as not belonging in the regression. In total, robust regression
identified 35 observations as being extreme and this included all the outliers also identified through the
studentised residuals.
The final approach is to remove observations in the 100th percentile. We generated a new variable
containing percentiles of real monthly earnings. This was used to identify and then drop 56 observations
in the 100th percentile with average real monthly earnings of MWK1,394,771 compared to
MWK355,924 in the 99th percentile. However, this results in a loss of 56 observations, which we
deemed excessive. One can equally consider dropping observations in the bottom percentile but we are
more concerned with outliers in the higher percentiles. The median of earnings is low (about
MWK8,400) while the mean is high (about MWK26,253), suggesting the data distribution is positively
skewed by the presence of large outliers.
Considering sample size issues, we used the second approach with results from the other three
approaches presented as robustness checks (see Section 3.8.1.3)
.
70
Stellenbosch University https://scholar.sun.ac.za
3.3.4
Dealing with inconsistencies
We noted some inconsistencies in terms of the period over which salaries were paid. About 6% of the
total sample in the formal sector reported a payment period different from that reported in 2010.
Similarly, about 24% of those in ganyu employment reported a different number of days spent on ganyu
per week in 2013 than they previously reported in 2010. One may reason that some of these changes
may be genuine considering that people switch jobs and occupations. However, before applying the
changes, we first checked for unusually large swings in wages and earnings. Moreover, the
inconsistencies were not unique to a specific occupation but rather spread across all occupationsconfirming our argument. These inconsistencies were dealt with by creating a period that is consistent
in both waves22.
In order to reduce noise introduced by individual heterogeneity and arrive at a fairly consistent data set,
we only worked with a balanced panel and also dealt with outliers which were partly responsible for
unrealistically large increases in earnings. Some of the outliers could have been as a result of the manner
in which we generate our monthly earnings, i.e. (ganyu days in a week*daily wage*4). In some cases,
we found some few ganyu millionaires. Our analysis showed that of the 56 total outliers, 7 were ganyu
earners. We had to deal with the outliers because we believe that a ganyu earner who had a large windfall
payment on a single day will have their earnings overstated after converting the daily earnings into
monthly wages. Nevertheless, our ganyu earnings compare very well with data from FISP as explained
in Section 3.7.3.
3.4 Labour force participation
The IHS questionnaire contains four questions which allow us to compute labour force participation and
unemployment rates for Malawi. The questions are:
a. Did you do work for any number of hours in the past 7 days?
b. Even though you did not do any activities in the last seven days, do you have a job or business or
any economic activity to return to?
c. If you were offered a job, would you be willing to accept the job?
d. In the past four weeks, have you taken any action to look for any kind of work or start any kind of
business / income generating activity?
The labour force is the sum of the employed and unemployed. Two standard definitions of
unemployment exist in the literature, namely broad and narrow definitions. Broad unemployment is
where a person is without work and available for work during the reference period. In this case, it
22
See Appendix A2 for description of the Stata code.
71
Stellenbosch University https://scholar.sun.ac.za
includes responses in (a), (b) and (c). Narrow unemployment is where a person is without work and
available for work during the reference period and is seeking work. This would be captured in (a), (b),
(c) and (d). The broad definition counts all the jobless people who want jobs even though they did not
search for one. The labour force participation comprises those employed and seeking work.
3.4.1
Size of labour force and labour force participation rates
Table 3.3 shows labour force participation for the balanced panel according to some selected
characteristics. One key finding is that, regardless of the measure used, the labour force participation for
the balanced panel has increased by almost 10 percentage points over the three-year period. For the
broad measure, the labour force participation rate has increased from 81% in 2010 to 91% in 2013. The
results indicate very high labour force participation rates - a key characteristic of low-income
agricultural economies23. We also note some gender gaps in favour of males in terms of both labour
force participation and employment rates. The gap has, however, marginally declined over three years.
Based on human capital theory, education does not only improve an individual’s chances of getting
employment but also positively impacts on earnings. Based on the table results, this theory seems to
hold amongst those with some level of education, i.e. if we exclude those without formal education. The
analysis shows that the labour force participation rate increases (due to higher employment rates) with
the level of education and is highest amongst those with university education, reaching around 90% in
both 2010 and 2013. It is, however, worth noting that those without formal education have equally high
levels of labour force participation - registering rates of as high as 81% in 2010 and 91% in 2013 when
measured in broad terms. Nonetheless, most of these are probably those who immediately enter the
informal sector of the labour market because they do not have formal education.
With respect to age-groups, we note that compared to any other age groups, the (15-19) 24 age group has
the lowest participation rate, at 58% in 2010 and about 77% in 2013. Consequently, as we will see, this
age-group has the largest rates of unemployment although the situation is improving. The rise in the
labour force participation rate between 2010 and 2013 is partly due to the fact that we are working with
a balanced panel such that no new young entrants are captured in the second wave, i.e. wave 2 would
only cover the 18 and 19-year olds.
23
In Malawi as well as the rest of Sub-Saharan Africa, the traditionally high rates of labour force participation can
be linked to the existence of very large agricultural sectors.
24
Based on the IHS3 figures, about 57% of the Malawian population is aged between 0 and 19 years. In our data
set, almost 22% is in the 15-19 age-group (National Statistical Office, 2012).
72
Stellenbosch University https://scholar.sun.ac.za
Table 3.3: Labour force participation rates according to characteristics
Description
Malawi
Sex
Male
Female
Education
None
PSLCE
JCE
MSCE
Tertiary
Age groups
15-19
20-24
25-29
30-34
35-39
40-44
45-49
50-54
55-59
60-64
Regions
Northern
Central
Southern
Working age population
2010 Share 2013 Share
7,086 100% 8,106 100%
Broad labour force
2010 Share 2013 Share
5,737 100% 7,336 100%
Narrow labour force
2010 Share 2013 Share
5,033 100% 6,578 100%
3,331
3,755
47%
53%
3,869
4,237
48%
52%
2,791
2,946
49%
51%
3,526
3,810
48%
52%
2,537
2,496
50%
50%
3,241
3,337
49%
51%
5,275
754
576
386
96
74%
11%
8%
5%
1%
5,841
958
719
463
125
72%
12%
9%
6%
2%
4,278
589
449
335
86
75%
10%
8%
6%
2%
5,326
832
636
430
112
73%
11%
9%
6%
2%
3,790
498
376
288
81
75%
10%
7%
6%
2%
4,787
744
560
385
102
73%
11%
9%
6%
2%
1,487
1,209
1,102
868
688
460
466
311
255
239
21%
17%
16%
12%
10%
6%
7%
4%
4%
3%
1,796
1,311
1,203
1,023
762
604
442
407
288
269
22%
16%
15%
13%
9%
7%
5%
5%
4%
3%
869
988
954
774
625
415
427
272
220
194
15%
17%
17%
13%
11%
7%
7%
5%
4%
3%
1,387
1,179
1,142
986
734
583
420
391
271
242
19%
16%
16%
13%
10%
8%
6%
5%
4%
3%
691
829
833
681
578
374
402
259
203
182
14%
16%
17%
14%
11%
7%
8%
5%
4%
4%
1,195
1,000
1,026
894
685
541
399
362
250
224
18%
15%
16%
14%
10%
8%
6%
6%
4%
3%
924
3,054
3,109
13%
43%
44%
1,055
3,541
3,510
13%
44%
43%
740
2,467
2,530
13%
43%
44%
964
3,174
3,198
13%
43%
44%
665
2,212
2,156
13%
44%
43%
861
2,875
2,842
13%
44%
43%
Source: Own calculations from IHPS data
3.4.2
Changes in the labour force according to background characteristics
In the next few paragraphs, we examine the patterns in the working age-population and labour force
over the three years under study for individuals in the panel. The results are presented in Table 3.4 and
show that there has been an increase of 14% in the number of people making up the working age
population between 2010 and 2013 in the balanced sample. This increase is higher than the average
GDP growth rate of around 3% experienced during the same period, raising fears about the economy’s
capacity to absorb the increasing labour supply. The labour force has grown both in broad and narrow
terms by about 28% and 31%, respectively. These observed increases in the labour force could either be
genuine or simply as a result of better capturing of the labour force in the administration of the
questionnaire.
With respect to gender, there was a larger increase in the working age population for males compared
to females. However, this pattern is reversed when we look at the labour force where females now have
a higher growth rate than males for both broad and narrow definitions.
73
Stellenbosch University https://scholar.sun.ac.za
Amongst the education categories, the largest growth rate for the working age population was largest
amongst those with tertiary education although the growth is from a very small number. However, the
growth in the labour force has been the largest amongst those with PSCLE and JCE. This observation
consistently holds for both the broad and narrow definitions of the labour force.
Table 3.4: Changes in the labour force according to characteristics
Description
Malawi
Sex
Male
Female
Education
None
PSLCE
JCE
MSCE
Tertiary
Age groups
15-19
20-24
25-29
30-34
35-39
40-44
45-49
50-54
55-59
60-64
Regions
Northern
Central
Southern
Working age population
2010
2013 % change
7,086 8,106
14%
Broad labour force
2010 2013 % change
5,737 7,336
28%
Narrow labour force
2010 2013 % change
5,033 6,578
31%
3,331
3,755
3,869
4,237
16%
13%
2,791
2,946
3,526
3,810
26%
29%
2,537
2,496
3,241
3,337
28%
34%
5,275
754
576
386
96
5,841
958
719
463
125
11%
27%
25%
20%
30%
4,278
589
449
335
86
5,326
832
636
430
112
25%
41%
42%
28%
30%
3,790
498
376
288
81
4,787
744
560
385
102
26%
49%
49%
34%
26%
1,487
1,209
1,102
868
688
460
466
311
255
239
1,796
1,311
1,203
1,023
762
604
442
407
288
269
21%
8%
9%
18%
11%
31%
-5%
31%
13%
12%
869
988
954
774
625
415
427
272
220
194
1,387
1,179
1,142
986
734
583
420
391
271
242
60%
19%
20%
27%
18%
41%
-1%
44%
23%
25%
691
829
833
681
578
374
402
259
203
182
1,195
1,000
1,026
894
685
541
399
362
250
224
73%
21%
23%
31%
19%
44%
-1%
40%
23%
23%
924
3,054
3,109
1,055
3,541
3,510
14%
16%
13%
740
2,467
2,530
964
3,174
3,198
30%
29%
26%
665
2,212
2,156
861
2,875
2,842
29%
30%
32%
Source: Own computations from IHPS data.
There have been positive growth rates across all age-groups except the (45-49) group which has
registered a decline of about 5% in the working age population and a corresponding drop of 1% for both
the broad and narrow definitions of the labour force.
All the three regions have registered increases in the working- age population numbers and this is also
reflected in the broad and narrow labour force numbers. Using the broad definition, the Northern region
has registered the largest growth rate in proportion to the rest. On the other hand, when the narrow
definition is used, the largest increase is observed in the Southern region.
74
Stellenbosch University https://scholar.sun.ac.za
3.4.3
Shares in the labour force
According to results presented in Table 3.5, females make up the majority of the labour force regardless
of whether the broad or narrow definition is used. The proportion of females in the labour force ranges
between 50% and 53% and this reflects well with the population shares observed in the 2008 census
data. However, proportionately fewer females enter the labour market compared to males. We will see
this clearly when we discuss unemployment rates in Section 3.5.
Table 3.5: Shares of working-age population and labour force according to characteristics
2010
7,086
Broad labour force participation Narrow labour force participation
2010
2013
2010
2013
2013 Number Rate Number Rate Number Rate Number Rate
8,106
5,737 81%
7,336 91%
5,033 71%
6,578
81%
3,331
3,755
3,869
4,237
2,791 84%
2,946 78%
3,526
3,810
91%
90%
2,537 76%
2,496 66%
3,241
3,337
84%
79%
5,275
754
576
386
96
5,841
958
719
463
125
4,278
589
449
335
86
81%
78%
78%
87%
90%
5,326
832
636
430
112
91%
87%
88%
93%
90%
3,790
498
376
288
81
72%
66%
65%
75%
84%
4,787
744
560
385
102
82%
78%
78%
83%
82%
1,487
1,209
1,102
868
688
460
466
311
255
239
1,796
1,311
1,203
1,023
762
604
442
407
288
269
869
988
954
774
625
415
427
272
220
194
58%
82%
87%
89%
91%
90%
92%
88%
86%
81%
1,387
1,179
1,142
986
734
583
420
391
271
242
77%
90%
95%
96%
96%
97%
95%
96%
94%
90%
691
829
833
681
578
374
402
259
203
182
46%
69%
76%
78%
84%
81%
86%
84%
80%
76%
1,195
1,000
1,026
894
685
541
399
362
250
224
67%
76%
85%
87%
90%
90%
90%
89%
87%
84%
924
3,054
3,109
1,055
3,541
3,510
740 80%
2,467 81%
2,530 81%
964
3,174
3,198
91%
90%
91%
665 72%
2,212 72%
2,156 69%
861
2,875
2,842
82%
81%
81%
Working age population
Description
Malawi
Sex
Male
Female
Education
None
PSLCE
JCE
MSCE
Tertiary
Age groups
15-19
20-24
25-29
30-34
35-39
40-44
45-49
50-54
55-59
60-64
Regions
Northern
Central
Southern
Source: Own computations from IHPS data.
As far as education is concerned, we make two important observations. First, is that the labour force in
Malawi is highly uneducated with percentages as high as between 72% and 75%. This is despite two
decades of free primary education in Malawi. The second observation is that despite remaining highly
uneducated, the labour force has become slightly more educated on average. As the table shows, the
shares of those without formal education (None) have dropped by two percentage points from 74% to
72%. A similar percentage point decrease has been registered in the broad and narrow labour force
figures. We also note that the shares of individuals with some education have increased over the three75
Stellenbosch University https://scholar.sun.ac.za
year survey period. However, much of the increase in the levels of education is accounted for by PSLCE
and JCE. MSCE and Tertiary have remained constant at 8% and 2%, respectively.
The table also shows that the majority of the work force in Malawi is made up of the youth and this is
reflective of Malawi’s population pyramid. Using the International Labour Organisation (ILO)
definition, the youth (15-24) make up 38% of the total working-age-population. This proportion has
been stable for both 2010 and 2013. On the other hand, using the Southern African Development
Community (SADC) definition, the youth (15-34) consist of 66% in both 2010 and 2013. Regardless of
the definition used, the youth make up the majority of the labour force.
We observe consistent shares in the working age population among regions for both the broad and
narrow labour force definitions. The Northern region has the lowest shares across the survey periods
partly because the population is also low there.
3.4.4
Multivariate analysis of labour force participation
We analyse the likelihood of labour force participation by the use of a probit regression. The results are
presented in Table 3.6. Both models generate correctly predicted probabilities of at least 78% and this
is quite high and good. Our dependent variable is a binary variable equal to ‘1’ if an individual is in the
labour force and ‘0’, if not. For both the broad and narrow definitions, the explanatory variables are: sex
of individuals (reference group: male), five year age categories (reference group: (15-19), education
levels (reference group: none), marital status (reference group: married or non-formal union), age
dependency ratio25, region (reference group: Northern), year (reference group: 2010).
Under both the broad and narrow definitions, females are less likely to enter the labour market compared
to males. Compared to the reference category, older individuals are more likely to participate in the
labour market. We also note that labour force participation is generally concave with respect to age first increasing before reaching a turning point.
After controlling for other factors and regardless of whether we use the broad or narrow definition,
individuals with primary and junior secondary education are significantly less likely to participate in the
labour market than the reference group who has no education. This is perhaps because they choose to
continue with their education unlike those without education who might as well join the market straight
away. In the narrow definition, we pick up an interesting and expected result where those with tertiary
25
Age dependency ratio is given as the ratio of dependents (people younger than 15 or older than 64) to the
working-age population (those aged 15-64).
76
Stellenbosch University https://scholar.sun.ac.za
education are more likely to enter the labour market compared to those without education, but this is not
the case for the broad definition. As the lower labour force participation rate amongst those with some
education could be the result of longer time spent studying, it is instructive to also see whether those
without education also have a higher rate of participation if the sample is limited to those above 25 years.
Table 3.6: Probit regressions on labour force participation
Description
Female
Age categories
20-24
25-29
30-34
35-39
40-44
45-49
50-54
55-59
60-64
Education level
Primary
Junior secondary
Senior secondary
Tertiary
Not married
Age dependency ratio
Region
Central
Southern
Year=2013
Constant
Percent correctly predicted probability
R-squared
Observations
Broad
-0.215*** (0.029)
Narrow
-0.365*** (0.025)
0.496***
0.743***
0.833***
0.852***
0.793***
0.870***
0.782***
0.609***
0.459***
(0.040)
(0.050)
(0.058)
(0.067)
(0.075)
(0.080)
(0.086)
(0.091)
(0.084)
0.317***
0.571***
0.686***
0.806***
0.719***
0.906***
0.786***
0.692***
0.598***
(0.037)
(0.043)
(0.049)
(0.056)
(0.062)
(0.069)
(0.072)
(0.080)
(0.077)
-0.076*
-0.130***
0.058
0.078
-0.521***
0.051**
(0.042)
(0.044)
(0.057)
(0.098)
(0.033)
(0.021)
-0.099***
-0.198***
-0.043
0.209**
-0.398***
0.099***
(0.037)
(0.039)
(0.046)
(0.085)
(0.030)
(0.019)
-0.110***
0.005
0.532***
0.760***
86.13%
0.162
15,192
(0.037)
(0.037)
(0.027)
(0.052)
-0.026
-0.053*
0.377***
0.457***
78.08%
0.123
15,192
(0.033)
(0.032)
(0.022)
(0.047)
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
Individuals who are unmarried (separated, divorced, widowed or never married) are less likely to enter
the labour force compared to those that are married or living with a partner. Those who come from
households with more children and elderly (high age dependency ratio) are more likely to participate in
the labour market. In the broad and narrow definitions, individuals residing in the Central region and
Southern region have a significantly lower likelihood of labour force participation compared to those in
the Northern region. The conditional likelihood of labour force participation is greater in 2013, either
77
Stellenbosch University https://scholar.sun.ac.za
indicating the presence of more job opportunities in 2013 than in 2010 or simply the fact that there are
now more people looking for jobs.
3.5 Unemployment
From Table 3.4, we noted that there was a growth in the labour force regardless of whether the broad or
narrow definition is used. However, the numbers of unemployed and unemployment rates shown in
Table 3.7 have been relatively stable over the three years. Broad unemployment rates have remained
constant at 13% while narrow unemployment has only marginally increased from 2% in 2010 to 3% in
2013. This means that much of the labour force growth was absorbed, probably on account of existing
vacancies or new jobs created.
Table 3.7: Unemployment shares and rates by year
Description
Malawi
Sex
Male
Female
Education
None
PSLCE
JCE
MSCE
Tertiary
Age groups
15-19
20-24
25-29
30-34
35-39
40-44
45-49
50-54
55-59
60-64
Regions
Northern
Central
Southern
Broadly unemployed
2010
2013
Number Share Rate Number Share Rate
743 100% 13%
938 100% 13%
Narrowly unemployed
2010
2013
Number Share Rate Number Share Rate
118
100% 2%
222 100% 3%
273
470
37% 10%
63% 16%
371
566
40% 11%
60% 15%
59
59
50%
50%
2%
2%
104
118
47%
53%
3%
4%
500
95
79
60
8
67%
13%
11%
8%
1%
12%
16%
18%
18%
10%
666
104
93
59
15
71%
11%
10%
6%
2%
13%
12%
15%
14%
14%
73
12
11
17
4
62%
10%
10%
14%
4%
2%
2%
3%
6%
5%
160
18
20
18
6
72%
8%
9%
8%
3%
3%
2%
4%
5%
6%
175
183
144
97
47
37
23
10
17
10
24%
25%
19%
13%
6%
5%
3%
1%
2%
1%
20%
18%
15%
13%
8%
9%
5%
4%
8%
5%
208
218
154
124
61
52
30
43
24
23
22%
23%
16%
13%
7%
6%
3%
5%
3%
2%
15%
19%
13%
13%
8%
9%
7%
11%
9%
9%
8
36
35
15
9
3
5
2
3
2
7%
31%
29%
13%
8%
3%
5%
1%
2%
1%
1%
4%
4%
2%
2%
1%
1%
1%
1%
1%
26
43
43
38
17
15
11
18
5
7
12%
20%
20%
17%
8%
7%
5%
8%
2%
3%
2%
4%
4%
4%
3%
3%
3%
5%
2%
3%
72
268
403
10% 10%
36% 11%
54% 16%
136
355
446
15% 14%
38% 11%
48% 14%
8
47
63
7%
40%
53%
1%
2%
3%
39
70
113
17%
32%
51%
4%
2%
4%
Source: Own computations from IHPS data.
78
Stellenbosch University https://scholar.sun.ac.za
As also earlier discussed, there is an evident gender gap in favour of males as reflected in higher rates
of unemployment for females. The gap is partly due to the fact that there are more females in the
population but could also be as a result of gender discrimination. Nevertheless, the gap seems to be
closing as shown by the decline in the share of females in broad unemployment from 63% in 2010 to
60% in 2013.
With respect to education, those without education make up the largest share of unemployment numbers
and the proportions decline as the level of education increases, consistent with the human capital theory.
However, there seems to be no clear pattern between unemployment rates and levels of education since
unemployment is prevalent at all levels of education. While the unemployment rates are lower for those
without education (because they can enter the ‘labour’ market earlier), they increase for individuals with
PSLCE, JCE and MSCE, probably because these people are still in school. The level and rates of
unemployment are the lowest amongst those with tertiary education.
The younger age groups not only make up the largest shares of unemployment figures but also
experience the highest rates of unemployment, ranging between 13% and 20% when broadly measured
and between 1% and 4% using narrow measurement.
3.6 Employment trends and characteristics
This section looks at employment trends between 2010 and 2013 as well as the characteristics of those
in employment in Malawi. We also examine the work activities of the employed.
3.6.1
Employment shares and growth rates
Table 3.8 shows that the shares of employment have fairly remained stable by gender although females
dominate in 2013 on account of growth. We observe a drop in the share of those without education from
76% in 2010 to a still high 73% in 2013. The youth, regardless of how we define them, dominate the
shares of employment numbers. We note, however, that in 2013, the (15-19) age group dominates the
share of employment at 18%, overtaking the (20-24) and (25-29) age groups which previously
dominated in 2010.
On average, the number employed has grown by 28% and 29% for broad and narrow measures,
respectively. A breakdown by gender shows that there was a larger increase for females than males
regardless of the measure used. The largest growth in the number employed occurred amongst those
with either PSLCE (49%) or JCE (48%) as their highest level of education. All age-groups, except those
in the (45-49) group, have registered growth between 2010 and 2013. The largest increase (at least 70%)
has been registered in the (15-19) age group. These are likely to be individuals who drop out of the
education system in search of employment.
79
Stellenbosch University https://scholar.sun.ac.za
Table 3.8: Broad and narrow employment shares and growth by year
Description
Malawi
Sex
Male
Female
Education
None
PSLCE
JCE
MSCE
Tertiary
Age groups
15-19
20-24
25-29
30-34
35-39
40-44
45-49
50-54
55-59
60-64
Regions
Northern
Central
Southern
Broadly employed
2010
2013
Growth
Number Share Number Share (2010-2013)
4,994 100%
6,398 100%
28%
2,518
2,476
50%
50%
3,155
3,243
49%
51%
25%
31%
2,478
2,437
50%
50%
3,137
3,219
49%
51%
27%
32%
3,777
494
370
275
78
76%
10%
7%
6%
2%
4,660
728
543
371
97
73%
11%
8%
6%
2%
23%
47%
47%
35%
24%
3,717
486
364
271
77
76%
10%
7%
6%
2%
4,627
725
540
368
96
73%
11%
8%
6%
2%
24%
49%
48%
36%
25%
694
805
811
677
578
377
403
262
203
184
14%
16%
16%
14%
12%
8%
8%
5%
4%
4%
1,179
961
989
862
673
531
390
348
247
219
18%
15%
15%
13%
11%
8%
6%
5%
4%
3%
70%
19%
22%
27%
16%
41%
-3%
33%
22%
19%
683
793
798
666
569
371
397
258
200
181
14%
16%
16%
14%
12%
8%
8%
5%
4%
4%
1,170
957
983
856
668
526
389
345
245
218
18%
15%
15%
13%
11%
8%
6%
5%
4%
3%
71%
21%
23%
29%
17%
42%
-2%
34%
23%
20%
668
2,200
2,127
13%
44%
43%
828
2,819
2,751
13%
44%
43%
24%
28%
29%
657
2,165
2,093
13%
44%
43%
823
2,805
2,728
13%
44%
43%
25%
30%
30%
Source: Own computations from IHPS data
3.6.2
Narrowly employed
2010
2013
Growth
Number Share Number Share (2010-2013)
4,915 100%
6,356 100%
29%
Multivariate analysis of employment likelihood
Not all people who are in the labour force end up being employed. This raises the problem of sample
selection, which we address using the Heckman (1979) two-step procedure as discussed earlier in
Section 3.2.3. In the first step, we estimate the probability of an individual either participating in the
labour force or not. From the first step in Section 3.4.4, we obtained the inverse of the Mills ratio and
used it as an additional explanatory variable in the second step, the probit on the likelihood of
employment. Our results are presented in Table 3.9.
80
Stellenbosch University https://scholar.sun.ac.za
Table 3.9: Two-step Heckman probit results on employment likelihood
Description
Female
Age categories
20-24
25-29
30-34
35-39
40-44
45-49
50-54
55-59
60-64
Education level
PSLCE
JCE
MSCE
Tertiary
Not married
Region
Central
Southern
Year=2013
Inverse mills ratio
Constant
Percent correctly predicted probability
R-squared
Observations
Broad
-0.220*** (0.029)
Narrow
-0.076* (0.045)
-0.171**
-0.038
0.073
0.204**
0.144
0.243**
0.156
0.204**
0.200**
(0.067)
(0.083)
(0.089)
(0.092)
(0.092)
(0.10)
(0.098)
(0.098)
(0.091)
-0.070
-0.019
0.040
0.118
0.088
0.140
0.071
0.108
0.075
(0.058)
(0.086)
(0.099)
(0.109)
(0.104)
(0.118)
(0.112)
(0.112)
(0.106)
-0.051
-0.132***
-0.164***
-0.018
-0.118**
(0.037)
(0.040)
(0.044)
(0.078)
(0.048)
-0.006
-0.035
-0.085*
-0.080
-0.108**
(0.039)
(0.047)
(0.045)
(0.078)
(0.051)
0.063*
-0.072**
0.002
-1.710***
1.284***
75.63%
0.106
15,192
(0.033)
(0.031)
(0.046)
(0.235)
(0.113)
0.022
-0.034
0.026
-1.654***
1.428***
75.65%
0.106
15,192
(0.032)
(0.032)
(0.047)
(0.244)
(0.141)
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
First, we note that the coefficient of the inverse Mills ratio is statistically significant, indicating the
presence of sample selection. It is, therefore, important to correct for selection bias. We then proceed to
correct for selection bias and report heteroskedasticity consistent standard errors in the first stage. Under
the broad definition, the results show also that the following groups are less likely to enter employment:
females, those aged (20-24), those with JCE and MSCE, individuals who are not married, and people
living in the Southern region, while those aged 35 and above as well as those from the Central region
are more likely to be employed. Fewer variables are significant in the narrow definition, where we find
that the following groups are significantly less likely to be employed: females, individuals with MSCE
and those that are not married.
3.7 Earnings and changes in employment status
Table 3.10 shows that overall, the total average real monthly wages have increased by 37% from MWK
20,186 in 2010 to MWK27,704 in 2013. In the ensuing paragraphs, we attempt to breakdown and explain
the sources of the increase. The first step is to examine how the employment status of individuals has
81
Stellenbosch University https://scholar.sun.ac.za
changed between the two waves. Conditional on missing earnings, we came up with four employment
statuses, unemployed in both waves (not shown in table), employed in either 2010 or 2013 only, and
employed in both periods.
Table 3.10: Mean monthly total wages by employment status and survey period
2010
SD
2013
SD
Mean
N
Percent
Mean
Employment status
Wave 1 only
13,123
914
38%
.
.
(56,494)
Wave 2 only
.
.
.
18,552 (49,050)
.
Both waves
24,696
1,520
62%
36,119 (96,957)
(75,160)
Total
20,186 ( 68,713)
2,434
100%
27,704 (78,263)
Source: Own computation from IHPS data, earnings expressed in constant 2013 prices.
3.7.1
N
Percent
.
1,411
1,532
2,943
.
43%
57%
100%
Employed in either wave
The table shows that part of the increase in earnings is explained by the new entrants into the labour
market (n=1,411) with average real monthly earnings of MWK18,552 compared to the MWK13,123 in
2010 of those who have now exited (n=914) the labour market. Between these two groups, average
earnings have increased by 41%. Consequently, those that have exited the market have been replaced
by higher earning individuals.
3.7.2
Employed in both waves
We observe that earnings are higher amongst individuals employed in both years compared to those
only employed in either period. The gap in earnings between these two groups is stable, i.e. MWK13,123
versus MWK24,696 (1.88 times higher) in 2010 and MWK18,552 versus MWK36,119 (1.95 times
more) in 2013. Moreover, those employed in both periods also experienced an increase in earnings; their
earnings increased from MWK24,696 in 2010 to MWK36,119 in 2013, representing an increase of 46%
over three years.
Considering the importance of ganyu employment in Malawi, the analysis in Table 3.10 is repeated for
ganyu earners. The results are given in Table 3.11 where a similar pattern is observed. First, the largest
increase in earnings is observed for individuals employed in both waves. Second, those employed in
wave 2 only (new entrants) earn more compared to individuals in wave 1 only.
82
Stellenbosch University https://scholar.sun.ac.za
Table 3.11: Mean monthly ganyu wages by employment status and survey period
2010
2013
Mean
SD
N
Percent
Mean
SD
N
Percent
Wave 1 only
9,234
(32,415)
711
47%
.
.
.
.
Wave 2 only
.
.
.
.
15,733
(42,866)
1,136
58%
Both waves
10,013
(16,498)
792
53%
20,231
(54,093)
835
42%
Total
9,649
(25,222)
1,503
100%
17,718
(48,183)
1,971
100%
Employment status
Source: Own computation from IHPS data, earnings expressed in constant 2013 prices.
We further examine if there is any relationship between employment status and education attainment.
Table 3.12 shows that of those without education qualification (“None”), 47% were unemployed in both
waves and this forms the majority. On the other hand, 61% and 60% of those with university diploma
and post-graduate degree, respectively, were employed in both periods. New entrants into the labour
market (wave 2 only) have more education compared to those that have dropped out of the labour market
(wave 1 only).
Table 3.12: Employment status and education attainment of individuals
Description
None
Unemployed in both waves
5,253 960
756
350
47%
56% 58% 41%
1,616 173
122
74
15%
10%
9%
9%
2,106 268
160
126
19%
16% 12% 15%
2,143 311
256
300
19%
18% 20% 35%
11,118 1,712 1,295 849
100% 100% 100% 100%
Employed in wave 1 only
Employed in wave 2 only
Employed in both waves
Total
PSCLE JCE
MSCE
Non-University University Post-graduate
Total
Diploma
Diploma
degree
26
11
3
7,358
20%
16%
13%
48%
18
6
4
2,013
14%
10%
19%
13%
17
9
2
2,687
13%
13%
8%
18%
70
41
14
3,135
54%
61%
60%
21%
130
67
23
15,194
100%
100%
100%
100%
Source: Own computation from IHPS data, percentages in italics.
3.7.3
Identifying sources of increases in earnings
Figure 3.1 shows real monthly wages by occupation and survey year. Although we do not fully know
the reasons behind the changes, it is important to point out two observations. Firstly, as we will note in
the figure, the average earnings in the NGO sector (PWP and churches) have dropped between 2010 and
2013. These might have been negatively affected by donor aid withdrawals that Malawi faced recently.26
26
The “Other” category is made up of 7 and 19 individuals in 2010 and 2013, respectively. Perhaps we need to
add them up to the dominant category or a category with similar characteristics.
83
Stellenbosch University https://scholar.sun.ac.za
Secondly, our results compare well with Farm Input Subsidy Programme (FISP) evaluation panel data
currently being analysed by other researchers in Malawi where large increases have been observed in
real ganyu wages (nominal ganyu daily wage rates divided by maize prices) for some districts between
2012 and 2015 (see Figure A1 in the appendix).
Figure 3.1: Histogram of real monthly earnings by survey year and occupation
Source: Own computation from IHPS data, bars represent standard errors.
Given this comparison, one can argue that the increases in the earnings may be genuine although it may
be difficult to isolate the main factors from the many drivers behind this rise. Gross Domestic Product
(GDP) growth may be a contributor but the economy only grew at an average of about 3% per year. As
for ganyu wages, it is possible that with the availability of food in many parts of the country, the ganyu
workers tended to have higher bargaining power and asked for more wages.
It is also worth noting that in real terms, minimum wages were adjusted upwards twice between 2010
and 2013; first by 50% effective 1st January 2011 and second by 34% effective 1st July 2013. These
could explain the increases in earnings in the private and government sectors, although this largely
depends on effective implementation and monitoring and would only affect low-income workers.
Despite working the same number of hours, those in government received significantly more per month
than their counterparts in parastatal companies. Nevertheless, parastatal organisations more than
doubled their earnings between 2010 and 2013. There is some catch up between the two groups.
84
Stellenbosch University https://scholar.sun.ac.za
3.8 Econometric analysis of returns to education
As explained in Section 3.3, the IHS questionnaire collects information on wage employment and
household non-farm enterprises. The former is analysed at the individual and the latter at the household
level. It is, however, possible to link household to individual data by enterprise owner.
Earnings from household non-farm enterprises are analysed at a household rather than the individual
level for three reasons. Firstly, we reason that wages and enterprise earnings (profit) may not necessarily
be explained by the same factors. This is confirmed because the R-squares for the wage regressions
improve after splitting the analysis (for wages and household earnings separately). Secondly, the IHS
data have a separate module for household enterprises with background characteristics such as industry,
customers and age of enterprises which allows us to achieve this purpose. Thirdly, looking at enterprises
separately enables us to explore the existence of education externalities within the household. We
analyse externalities by using maximum education in the household. Average household education is
used for robustness checks and also works well. We begin our analysis by looking at wage employment
before moving on to enterprise earnings in Section 3.8.2.
3.8.1
Wage employment
Recall that according to the human capital theory, investment in education improves workers’ skills
resulting in high productivity and, therefore, higher earnings (Mincer, 1974). We begin our analysis with
ordinary least squares estimation (OLS) for 2010, 2013 and the pooled sample. The results are given in
Table 3.13 where the dependent variable is the log of real monthly earnings 27. The explanatory variables
are years of schooling, years of experience, square of experience, sex (base category: male), region
(base: Northern) and occupation (base: private individual).
Across the three models, the results show that there are large returns to education of between 6.8% and
7.7%. The strongly positive returns to education are consistent with other findings in Malawi (Chirwa
& Matita 2009; Chirwa & Zgovu 2002). Similar results have been found in other African countries (e.g.,
in Cameroon by Ewoudou & Vencatachellum, 2006; in Rwanda by Lassibille & Tan, 2005); Bennell,
1996, for Sub-Saharan Africa). The negative and significant gender dummy is consistent with the
general finding that females earn less than their male counterparts (Chirwa & Matita, 2009). We also
find that average earnings for private individuals are significantly lower than those in private companies.
As expected, ganyu earnings are consistently significantly lower in all the models. It is only in the 2013
27
The interpretation of the coefficients is the percentage change in the monthly earnings given a unit change in
the explanatory variable. For dummy variables the percentage effect of a change from the base category.
85
Stellenbosch University https://scholar.sun.ac.za
model that workers in public works programme earn significantly less than those in private companies,
albeit only at the 10% level of significance.
Table 3.13: OLS results for log of real monthly wages
OLS
2013
0.068***
(0.009)
0.055***
(0.006)
-0.001***
(0.000)
-0.276***
(0.042)
-0.019
(0.098)
-0.144
(0.096)
-0.555***
(0.126)
0.232
(0.162)
0.386*
(0.197)
-0.262*
(0.157)
0.171
(0.234)
0.036
(0.283)
-0.805***
(0.105)
9.080***
(0.168)
0.282
2,943
Description
2010
Years of schooling
0.077***
(0.008)
Experience
0.037***
(0.006)
Experience squared
-0.001***
(0.000)
Female
-0.260***
(0.045)
Central
-0.101
(0.081)
Southern
-0.257***
(0.074)
Private Individual
-0.210*
(0.108)
Government
0.206
(0.130)
Parastatal
0.179
(0.196)
Public Works Program
0.555
(0.376)
Church/Religious Organisation
-0.024
(0.335)
Other
0.135
(0.490)
Casual/Ganyu
-0.803***
(0.094)
Constant
8.962***
(0.153)
R-squared
0.340
Observations
2,434
Pooled
0.072***
(0.008)
0.047***
(0.005)
-0.001***
(0.000)
-0.267***
(0.032)
-0.054
(0.075)
-0.202***
(0.066)
-0.392***
(0.094)
0.206*
(0.122)
0.305**
(0.145)
0.094
(0.193)
0.107
(0.231)
0.128
(0.284)
-0.795***
(0.083)
9.015***
(0.137)
0.295
5,377
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
86
Stellenbosch University https://scholar.sun.ac.za
3.8.1.1 Homogeneous returns to education
Next, we discuss findings from three models, namely OLS, fixed effects and random effects beginning
with estimation results presented in Table 3.14. In each of these models, we assume that returns to
education are homogeneous and also ignore selectivity bias28.
Table 3.14: OLS, Fixed effects and random effects results for log of real monthly wages
Pooled OLS
1
2
3
Years of schooling
0.109*** 0.072*** 0.072***
(0.007)
(0.008)
(0.007)
Experience
0.047*** 0.047***
(0.005)
(0.005)
Experience squared
-0.001*** -0.001***
(0.000)
(0.000)
Female
-0.267*** -0.268***
(0.032)
(0.032)
Central
-0.054
-0.05
(0.075)
(0.074)
Southern
-0.202*** -0.194***
(0.066)
(0.066)
Private Individual
-0.392*** -0.369***
(0.094)
(0.096)
Government
0.206*
0.216*
(0.122)
(0.122)
Parastatal
0.305** 0.298**
(0.145)
(0.139)
Public Works Program
0.094
0.061
(0.193)
(0.210)
Church/Religious Organisation
0.107
0.091
(0.231)
(0.225)
Other
0.128
0.064
(0.284)
(0.280)
Casual/Ganyu
-0.795*** -0.805***
(0.083)
(0.085)
Year=2013
0.297***
(0.039)
Constant
8.418*** 9.015*** 8.868***
(0.057)
(0.137)
(0.133)
R-squared
0.137
0.295
0.313
Observations
5,377
5,377
5,377
Description
4
0.009
(0.008)
Fixed effects
5
0.055***
(0.010)
0.068***
(0.011)
-0.000*
(0.000)
0.567***
(0.208)
0.365
(0.265)
-0.113
(0.081)
0.085
(0.145)
0.290**
(0.126)
-0.247
(0.251)
0.172
(0.169)
0.352**
(0.164)
-0.487***
(0.089)
6
0.015
(0.015)
0.018
(0.017)
0.000
(0.000)
0.418**
(0.203)
0.281
(0.256)
-0.111
(0.079)
0.070
(0.151)
0.275**
(0.127)
-0.305
(0.241)
0.114
(0.166)
0.308**
(0.151)
-0.495***
(0.088)
0.263***
(0.053)
9.306*** 7.816*** 8.918***
(0.066)
(0.253)
(0.374)
0.638
0.672
0.685
5,377
5,377
5,377
Random effects
7
8
9
0.115*** 0.086*** 0.083***
(0.004)
(0.005)
(0.005)
0.043*** 0.041***
(0.004)
(0.004)
-0.001*** -0.001***
(0.000)
(0.000)
-0.277*** -0.280***
(0.030)
(0.030)
-0.030
-0.036
(0.037)
(0.037)
-0.179*** -0.183***
(0.037)
(0.037)
-0.302*** -0.292***
(0.056)
(0.056)
0.246*** 0.253***
(0.070)
(0.070)
0.355*** 0.371***
(0.111)
(0.109)
0.177
0.143
(0.180)
(0.186)
0.155
0.144
(0.124)
(0.124)
0.522*** 0.443**
(0.178)
(0.175)
-0.731*** -0.754***
(0.049)
(0.048)
0.240***
(0.021)
8.382*** 8.846*** 8.791***
(0.034)
(0.085)
(0.084)
0.233
0.374
0.383
5,377
5,377
5,377
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
The fixed effects model does not work well, indicating very low within variation29 in our variables
especially after the inclusion of the time dummy. The time dummy further reduces the variation in the
data and this somehow causes the results to change dramatically30. As the results show, the coefficient
for education in the OLS regression is almost 5 times as bigger as that of fixed effects- a small coefficient
28
Selection bias is explained in Section 3.2.3.
29
Low within variation refers to a situation where independent variables change very gradually over time.
30
We tested as to whether we need to be using a time dummy and the results support its inclusion.
87
Stellenbosch University https://scholar.sun.ac.za
for the fixed effect compared to OLS is a sign that there is little within variation. Moreover, in the fixed
effects model, the coefficient for education is insignificant. We, therefore, concentrate on OLS and
random effects which yield consistent results except for differences in the magnitudes in selected cases.
The Breusch Pagan LM test31 yields significant results, indicating that the random effects model is more
appropriate compared to OLS. The time dummy shows that average monthly earnings are higher in 2013
compared to 2010.
3.8.1.2 Heterogeneous returns to education
The results from the basic models discussed above disregard the differences in the level of educational
attainment by looking at a single overall education level- years of schooling. This homogeneous model
assumes that there are no differential trends in the returns to education for different levels of education.
As earlier discussed, there is little statistical evidence and causal empiricism for the homogenous model.
The heterogeneous model provides the alternative and looks at the different levels of education as having
separate effects on earnings. Using this model specification, we replace years of schooling ( S ) with an
educational dummy variable to represent the different educational categories discussed in Section 3.2.2.
The results are presented in Table 3.15 and we provide a graphical illustration in Figure 3.2 but based
on the random effects only. Regardless of gender, the returns to education increase with the level of
education supporting a convex relationship between education and earnings. Our results confirm the
finding that the returns increase with the level of schooling in Sub-Saharan Africa but are also contrary
to the literature supporting concave rates of returns such as Psacharopoulos (1994).
Also consistent with the international literature is the finding that female workers tend to have much
higher returns on education than male workers, particularly at higher levels of education. Similar
findings have been established in Malawi by Chirwa and Matita (2009). Although more men enter the
labour market, they tend to have lower returns than their female counterparts with similar education
levels suggesting that female education is more effective in generating returns in Malawi. The high rates
of return to higher education for females and tertiary education may mean that gender discrimination
favouring men is reduced at higher levels of education. It is worth noting that the education system in
Malawi has greatly emphasised on primary education especially with the introduction of free primary
education in 1994. There has been a lot of expansion in primary education attainment and this possibly
explains the low returns.
31
The test helps a researcher decide between a random effects regression and a simple OLS regression (e.g.,
Baltagi, B.H. & Li, Q., 1990). The null hypothesis is that variances across individuals are zero. Put differently,
that there is no significant difference across units (i.e. no panel effect).
88
Stellenbosch University https://scholar.sun.ac.za
Table 3.15: OLS and random effects results for log of monthly wages using education categories
Pooled OLS
Male
Female
All
PSLCE
0.168**
-0.022
0.120*
(0.076)
(0.117)
(0.067)
JCE
0.341*** 0.428** 0.356***
(0.088)
(0.208)
(0.087)
MSCE
0.670*** 1.184*** 0.753***
(0.106)
(0.166)
(0.094)
Non-University Diploma
1.430*** 1.617*** 1.498***
(0.149)
(0.206)
(0.135)
University Diploma
1.891*** 2.085*** 1.995***
(0.173)
(0.330)
(0.160)
Post-Graduate Diploma
2.773*** 3.220*** 2.819***
(0.299)
(0.295)
(0.314)
Experience
0.056*** 0.034*** 0.045***
(0.005)
(0.006)
(0.004)
Experience squared
-0.001*** -0.001*** -0.001***
(0.000)
(0.000)
(0.000)
Female
-0.273***
(0.032)
Central
-0.101
-0.07
-0.09
(0.081)
(0.087)
(0.074)
Southern
-0.214*** -0.238*** -0.227***
(0.068)
(0.083)
(0.063)
Private Individual
-0.268*** -0.339* -0.292***
(0.096)
(0.193)
(0.090)
Government
0.119
-0.004
0.113
(0.129)
(0.179)
(0.110)
Parastatal
0.379*** -0.188
0.247**
(0.139)
(0.186)
(0.122)
Public Works Program
0.052
0.162
0.139
(0.275)
(0.229)
(0.204)
Church/Religious Organisation
-0.233
0.53
0.009
(0.165)
(0.379)
(0.210)
Other
0.066
0.01
0.044
(0.293)
(0.316)
(0.238)
Casual/Ganyu
-0.637*** -0.784*** -0.698***
(0.094)
(0.160)
(0.086)
Year=2013
0.282*** 0.298*** 0.289***
(0.038)
(0.054)
(0.038)
Constant
9.100*** 9.124*** 9.234***
(0.119)
(0.183)
(0.111)
R-squared
0.308
0.316
0.337
Observations
3,244
2,133
5,377
Description
Random effects
Male
Female
All
0.162***
0.052
0.141***
(0.051)
(0.082)
(0.043)
0.411*** 0.351*** 0.401***
(0.057)
(0.103)
(0.050)
0.726*** 1.071*** 0.800***
(0.063)
(0.132)
(0.057)
1.435*** 1.612*** 1.503***
(0.093)
(0.162)
(0.082)
1.932*** 1.999*** 1.968***
(0.132)
(0.20)
(0.109)
2.590*** 3.069*** 2.673***
(0.206)
(0.290)
(0.186)
0.046*** 0.027*** 0.037***
(0.005)
(0.005)
(0.004)
-0.001*** -0.000*** -0.001***
(0.000)
(0.000)
(0.000)
-0.282***
(0.029)
-0.088*
-0.085
-0.087**
(0.046)
(0.057)
(0.036)
-0.194*** -0.234*** -0.214***
(0.045)
(0.058)
(0.036)
-0.198*** -0.254* -0.208***
(0.060)
(0.134)
(0.055)
0.154**
0.061
0.149**
(0.075)
(0.135)
(0.066)
0.279***
0.307
0.311***
(0.102)
(0.252)
(0.102)
0.136
0.211
0.204
(0.266)
(0.236)
(0.170)
0.015
0.263
0.1
(0.143)
(0.221)
(0.119)
0.349*
0.281
0.324**
(0.203)
(0.218)
(0.145)
-0.576*** -0.744*** -0.632***
(0.054)
(0.122)
(0.049)
0.237*** 0.220*** 0.234***
(0.027)
(0.035)
(0.021)
9.103*** 9.155*** 9.232***
(0.083)
(0.145)
(0.069)
0.385
0.424
0.423
3,244
2,133
5,377
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
89
Stellenbosch University https://scholar.sun.ac.za
Figure 3.2: Rates of return on education by gender
Source: Own computation from IHPS data.
3.8.1.3 Robustness checks
We conduct five main robustness checks to see if our main results are preserved under different
conditions and assumptions. These include decomposition of results by gender, analysis of returns to
education only based on individuals employed in both waves, alternative treatment of outliers, sample
selection and distinguishing the results by sector. We discuss these in the paragraphs that follow.
Firstly, we decompose our results by gender. We report OLS and random effects regression results in
Table 3.16 having split the sample by the gender of earners. A version of this segregation is already
discussed in Section 3.8.1.2 but assuming heterogeneous returns to education. The table shows that the
key results after dividing the sample into male and female sub-samples are preserved. The difference
for the coefficient comparison test (male minus female) reveals that the coefficients for females are
smaller compared to men in most cases as shown by positive and significant differences for years of
schooling, experience, government and ganyu. The difference is the largest for ganyu employment. The
only case where females have a larger coefficient is within church/religious organisations where the
magnitude is very surprisingly large.
Secondly, we consider individuals employed in both waves only. As we earlier noted in Section 3.7,
these individuals experienced the largest jump in earnings over three years and also have the highest
levels of education. The results from Table 3.16 (last two columns) do not differ much compared to
those based on the full sample (which includes those only employed in either waves 1 or 2).
90
Stellenbosch University https://scholar.sun.ac.za
Table 3.16: Regressions of log monthly wages by gender and employees in both waves
Pooled OLS
Random effects Employed in both waves
Male
Female Difference
Male
Female
OLS
Random
Years of schooling
0.078*** 0.056*** 0.022** 0.089*** 0.068*** 0.076*** 0.083***
(0.008)
(0.011)
(0.006)
(0.008)
(0.008)
(0.006)
Experience
0.056*** 0.035*** 0.021** 0.049*** 0.030*** 0.039*** 0.031***
(0.006)
(0.007)
(0.005)
(0.006)
(0.006)
(0.006)
Experience squared
-0.001*** -0.001*** 0.000** -0.001*** -0.000*** -0.001*** -0.000***
(0.000)
(0.000)
(0.000)
(0.000)
(0.000)
(0.000)
Female
-0.174*** -0.194***
(0.048)
(0.043)
Central
-0.059
-0.034
-0.025
-0.036
-0.034
-0.059
-0.023
(0.080)
(0.090)
(0.047)
(0.059)
(0.084)
(0.053)
Southern
-0.173** -0.227***
0.054
-0.158*** -0.216*** -0.166** -0.139***
(0.070)
(0.085)
(0.046)
(0.060)
(0.076)
(0.051)
Private Individual
-0.303*** -0.618***
0.315
-0.237*** -0.508*** -0.374*** -0.263***
(0.097)
(0.20)
(0.060)
(0.139)
(0.105)
(0.065)
Government
0.209
0.163
0.046** 0.246***
0.196
0.207
0.254***
(0.138)
(0.201)
(0.078)
(0.152)
(0.137)
(0.080)
Parastatal
0.430*** -0.182
0.612
0.324*** 0.414*
0.376**
0.412***
(0.155)
(0.253)
(0.117)
(0.234)
(0.153)
(0.117)
Public Works Program
0.024
-0.117
0.141
0.12
-0.056
-0.204
-0.245*
(0.280)
(0.237)
(0.294)
(0.243)
(0.186)
(0.147)
Church/Religious Organisation
-0.174
0.591
-0.765**
0.026
0.372
0.122
0.208
(0.179)
(0.392)
(0.144)
(0.245)
(0.248)
(0.139)
Other
0.029
0.036
-0.007
0.421*
0.43
-0.179
0.264
(0.337)
(0.431)
(0.242)
(0.287)
(0.270)
(0.180)
Casual/Ganyu
-0.692*** -1.108*** 0.416** -0.646*** -1.057*** -0.742*** -0.678***
(0.088)
(0.167)
(0.053)
(0.115)
(0.086)
(0.057)
Year=2013
0.286*** 0.297***
-0.011
0.240*** 0.221*** 0.328*** 0.256***
(0.039)
(0.056)
(0.027)
(0.035)
(0.042)
(0.026)
Constant
8.636*** 9.101*** -0.465** 8.566*** 9.026*** 8.923*** 8.922***
(0.135)
(0.227)
(0.103)
(0.161)
(0.145)
(0.111)
R-squared
0.287
0.278
0.348
0.376
0.348
0.411
Observations
3,244
2,133
3,244
2,133
3,052
3,052
Description
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
Thirdly, we compare results based on alternative treatment of outliers. Table 3.17 gives results based on
four alternative ways of dealing with outliers discussed in Section 3.3.3. Overall, the results do not
change much considering both magnitude and direction, implying that the method of outlier treatment
does not really matter. However, failure to account for outliers yields larger returns to education and
this would be more evident if we decompose by groups such as sector.
91
Stellenbosch University https://scholar.sun.ac.za
Table 3.17: Random effect results based on alternative treatment of outliers
Description
Outliers included
Years of schooling
0.087***
(0.005)
Experience
0.042***
(0.004)
Experience squared
-0.001***
(0.000)
Female
-0.275***
(0.030)
Central
-0.04
(0.038)
Southern
-0.181***
(0.038)
Private Individual
-0.300***
(0.057)
Government
0.228***
(0.072)
Parastatal
0.298**
(0.125)
Public Works Program
0.123
(0.186)
Church/Religious Organisation
0.114
(0.125)
Other
0.410**
(0.174)
Casual/Ganyu
-0.760***
(0.049)
Year=2013
0.246***
(0.022)
Constant
8.759***
(0.085)
R-squared
0.380
Observations
5,391
Random effects
Millionaires 100th percentile Robust regression Extreme residuals
0.081***
0.077***
0.085***
0.083***
(0.005)
(0.004)
(0.005)
(0.005)
0.041***
0.040***
0.040***
0.041***
(0.004)
(0.004)
(0.004)
(0.004)
-0.001***
-0.001***
-0.001***
-0.001***
(0.000)
(0.000)
(0.000)
(0.000)
-0.279***
-0.274***
-0.273***
-0.280***
(0.030)
(0.029)
(0.029)
(0.030)
-0.024
-0.043
-0.042
-0.036
(0.036)
(0.036)
(0.036)
(0.037)
-0.168***
-0.189***
-0.189***
-0.183***
(0.036)
(0.035)
(0.036)
(0.037)
-0.285***
-0.278***
-0.296***
-0.292***
(0.055)
(0.054)
(0.055)
(0.056)
0.180***
0.162***
0.235***
0.253***
(0.064)
(0.061)
(0.069)
(0.070)
0.362***
0.344***
0.362***
0.371***
(0.107)
(0.103)
(0.108)
(0.109)
0.134
0.055
0.139
0.143
(0.186)
(0.165)
(0.185)
(0.186)
0.095
0.045
0.066
0.144
(0.115)
(0.110)
(0.113)
(0.124)
0.435**
0.417**
0.431**
0.443**
(0.174)
(0.176)
(0.173)
(0.175)
-0.748***
-0.746***
-0.762***
-0.754***
(0.048)
(0.047)
(0.048)
(0.048)
0.242***
0.243***
0.233***
0.240***
(0.021)
(0.021)
(0.021)
(0.021)
8.786***
8.826***
8.786***
8.791***
(0.083)
(0.081)
(0.083)
(0.084)
0.370
0.370
0.397
0.383
5,367
5,335
5,356
5,377
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
Fourthly, we address sample selection which as indicated can bias results if not addressed in the earnings
regression. We, therefore, correct for sample selection using the Heckman selection model as earlier
explained. In the first step, we model the probability of labour force participation. In the second step,
we model earnings as explained in equation 2 where the inverse mills ratio is now added as an additional
explanatory variable. The sample size is reduced by two observations due to the omission of missing
values in the process of generating the mills ratio from the first stage.
The sample selection corrected results presented in Table 3.18 and do not differ much from the
uncorrected findings in Table 3.13 and Table 3.14. A simple interpretation of this is that sample selection
is not a very serious issue in our data.
92
Stellenbosch University https://scholar.sun.ac.za
Table 3.18: Wage functions corrected for sample selection
Description
Years of schooling
OLS (2010) OLS (2013)
0.067***
0.061***
(0.009)
(0.008)
Experience
0.014*
0.030***
(0.008)
(0.008)
Experience squared
-0.000*
-0.001***
(0.000)
(0.000)
Female
-0.183***
-0.257***
(0.048)
(0.042)
Central
-0.091
0.018
(0.081)
(0.099)
Southern
-0.266***
-0.139
(0.075)
(0.096)
Private Individual
-0.200*
-0.563***
(0.106)
(0.125)
Government
0.218*
0.231
(0.131)
(0.162)
Parastatal
0.184
0.377*
(0.187)
(0.201)
Public Works Programme
0.528
-0.251
(0.371)
(0.157)
Church/Religious Organisation
-0.033
0.167
(0.337)
(0.229)
Other
0.156
0.016
(0.479)
(0.277)
Casual/Ganyu
-0.788***
-0.804***
(0.095)
(0.105)
Inverse mills ratio_2010
-0.713***
(0.195)
Inverse mills ratio_2013
-1.031***
(0.330)
Inverse mills ratio
Year=2013
Constant
R-squared
Observations
9.401***
(0.187)
0.346
2,434
9.503***
(0.204)
0.287
2,941
Pooled OLS
0.065***
(0.008)
0.027***
(0.006)
-0.000***
(0.000)
-0.224***
(0.031)
-0.029
(0.075)
-0.195***
(0.066)
-0.365***
(0.094)
0.220*
(0.122)
0.295**
(0.138)
0.054
(0.205)
0.089
(0.225)
0.059
(0.275)
-0.796***
(0.084)
Random
0.077***
(0.005)
0.024***
(0.005)
-0.000***
(0.000)
-0.242***
(0.030)
-0.018
(0.037)
-0.186***
(0.037)
-0.287***
(0.056)
0.256***
(0.070)
0.377***
(0.110)
0.129
(0.181)
0.137
(0.124)
0.443**
(0.175)
-0.743***
(0.048)
-0.722***
(0.182)
0.208***
(0.044)
9.274***
(0.163)
0.317
5,375
-0.675***
(0.132)
0.162***
(0.025)
9.155***
(0.107)
0.387
5,375
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
However, it is important to correct for sample selection as we have done given that the inverse mills
ratio is significant in each of the four models. Without sample correction, the returns to education tend
to be larger by between 0.7 and 0.8 percentage points than with sample selection. Our estimates are,
93
Stellenbosch University https://scholar.sun.ac.za
therefore, upwardly biased when sample selection is not addressed. Self-selection between sectors is
addressed in the section that follows.
The final robustness check is to distinguish between formal and informal sectors. Just like in many other
developing countries, the formal sector in Malawi only absorbs a small percentage of the labour force
(Chirwa & Matita, 2009). In our data set, about 77.54% of individuals aged between 15 and 64 years
with positive earnings are employed in the informal sector. Due to the large size of the informal sector,
studying earning differentials between economic sectors becomes important for policy.
The process of accounting for self-selection is the same as before; we first run a probit on the choice of
employment sector (see Table 3.19) from which we obtain the inverse mills ratio used in the second
stage. The first stage results show that the following are less likely to enter the formal sector: females,
those not married and individuals from households with high age dependency ratio. The rest of the
explanatory variables are associated with a higher probability of entering the formal sector.
Table 3.19: Probit on choice of employment sector
Explanatory variables
Coefficient Standard error
Female
-0.439***
(0.045)
Age groups
20-24
0.278***
(0.098)
25-29
0.607***
(0.096)
30-34
0.911***
(0.099)
35-39
0.913***
(0.103)
40-44
1.079***
(0.110)
45-49
1.040***
(0.113)
50-54
1.093***
(0.124)
55-59
0.906***
(0.140)
60-64
0.737***
(0.141)
PSLCE
0.512***
(0.062)
JCE
0.990***
(0.067)
MSCE
1.867***
(0.077)
Non-University Diploma
2.343***
(0.166)
University Diploma
2.691***
(0.267)
Post Graduate degree
2.420***
(0.412)
Not married
-0.096*
(0.051)
Age dependency ratio
-0.212***
(0.033)
Central
0.160***
(0.056)
Southern
0.296***
(0.055)
Year=2013
-0.169***
(0.041)
Constant
-1.276***
(0.104)
Percent correctly predicted 80.10%
Pseudo R-squared
0.315
Observations
5,416
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
94
Stellenbosch University https://scholar.sun.ac.za
For each of the two sectors, we present two sets of results, namely without and with self-selection
correction. To achieve this, we distinguish between the formal and informal sectors. Sector
decomposition analysis of earnings sheds some light on the importance of distinguishing between the
different types of employment sector in estimating returns to education in developing countries. Studies
that fail to take this into account tend to overstate the returns to education by assuming that returns are
the same in both the formal and informal sectors. We with uncorrected results (without sample selection)
in Table 3.20 where we also conduct the coefficient comparison test.
Table 3.20: Regression results on log of monthly wages by sector without sample selection
Description
Formal
Years of schooling
0.099***
(0.011)
Experience
0.041***
(0.008)
Experience squared
-0.001***
(0.000)
Female
-0.079
(0.076)
Central
0.030
(0.136)
Southern
-0.072
(0.132)
Private Individual
-0.328***
(0.092)
Government
0.082
(0.128)
Parastatal
0.262**
(0.130)
Public Works Programme
0.054
(0.216)
Church/Religious Organisation
0.031
(0.215)
Other
0.025
(0.276)
Year=2013
0.204***
(0.060)
Constant
8.576***
(0.221)
R-squared
0.238
Observations
1,909
Pooled OLS
Random effects
Informal Difference Formal Informal
0.037*** 0.062*** 0.086*** 0.047***
(0.009)
(0.006)
(0.007)
0.048***
-0.007
0.027*** 0.048***
(0.006)
(0.006)
(0.004)
-0.001***
0.000
-0.000** -0.001***
(0.000)
(0.000)
(0.000)
-0.345*** 0.266*** -0.117* -0.368***
(0.033)
(0.066)
(0.032)
-0.113
0.143
0.085 -0.122***
(0.079)
(0.073)
(0.040)
-0.264***
0.192
-0.052 -0.263***
(0.070)
(0.071)
(0.040)
-0.243***
(0.056)
0.171**
(0.068)
0.347***
(0.098)
0.113
(0.182)
0.158
(0.126)
0.353**
(0.139)
0.322***
-0.118
0.164*** 0.280***
(0.047)
(0.030)
(0.028)
8.348***
0.228
8.767*** 8.308***
(0.115)
(0.135)
(0.078)
0.120
0.297
0.132
3,468
1,909
3,468
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
The returns to education are positive in both sectors but with larger magnitudes in the formal sector for
both OLS and random effects. It is not surprising for the informal sector to have positive returns to
95
Stellenbosch University https://scholar.sun.ac.za
education. Previous research in Malawi by Chirwa and Zgovu (2002) has also shown that casual
employment (informal sector wage employment) on peak labour tasks may be well paid, above the daily
minimum wage, but short-lived.
The convex relationship between earnings and experience is also maintained in both sectors except that
the coefficient is bigger in the informal sector than the formal sector. This implies that experience
matters more in the informal sector. The time dummy coefficient is also positive in both models but
larger in the informal sector, an indication that earnings grew much stronger within the informal sector.
Next, we look at results corrected for sample selection in Table 3.21.
Table 3.21: Regression results on log of monthly wages by sector with sample selection
Pooled OLS
Formal Informal
Years of schooling
0.058***
0.012
(0.014)
(0.011)
Experience
0.029*** 0.030***
(0.010)
(0.008)
Experience squared
-0.001*** -0.001***
(0.000)
(0.000)
Female
0.070 -0.207***
(0.087)
(0.048)
Central
-0.005
-0.154*
(0.137)
(0.082)
Southern
-0.153 -0.329***
(0.130)
(0.071)
Private Individual
-0.275***
(0.092)
Government
0.060
(0.126)
Parastatal
0.266*
(0.137)
Public Works Programme
0.116
(0.218)
Church/Religious Organisation
0.015
(0.215)
Other
0.062
(0.269)
Inverse mills ratio
-0.459*** -0.323***
(0.146)
(0.082)
Year=2013
0.249*** 0.371***
(0.064)
(0.045)
Constant
9.472*** 9.144***
(0.330)
(0.256)
R-squared
0.251
0.127
Observations
1,909
3,466
Description
Random effects
Formal Informal
0.044*** 0.015*
(0.007)
(0.008)
0.012* 0.027***
(0.006)
(0.005)
0.000 -0.000***
(0.000)
(0.000)
0.041 -0.210***
(0.065)
(0.041)
0.023 -0.169***
(0.073)
(0.040)
-0.145** -0.337***
(0.072)
(0.041)
-0.178***
(0.057)
0.111*
(0.066)
0.324***
(0.099)
0.209
(0.187)
0.148
(0.126)
0.366***
(0.131)
-0.570*** -0.363***
(0.071)
(0.059)
0.215*** 0.331***
(0.029)
(0.029)
9.799*** 9.224***
(0.174)
(0.163)
0.311
0.142
1,909
3,466
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
96
Stellenbosch University https://scholar.sun.ac.za
We note a number of things that we did not observe without sample selection. Although the key results
are preserved in terms of direction and significance after correcting for self-selection, the magnitudes of
the coefficients significantly become smaller. This suggests that failure to account for sample selection
upwardly biases the results reported in studies. For example, after controlling for sample selection, we
find a smaller statistically significant coefficient for gender in the formal sector compared to the formal
sector. In addition, while the coefficients for regions are insignificant in the formal sector, the differences
are not only significant but also larger in the informal sector.
3.8.2
Household non-farm enterprise earnings
In this section, we explore two main ideas. Firstly, we would like to establish if there are positive
‘externalities’ in education in the running of household enterprises. Secondly, we establish if the
observed externalities have a turning point beyond which they either increase or decrease.
Table 3.22 presents the results where we establish that strong education externalities exist within the
households in non-farm enterprises. After controlling for other factors, we find returns to education of
up to 12.8% and 15.8% for OLS and random effects, respectively. The control factors are experience,
the gender of the owner, location of the enterprise, enterprise industry, the age of the enterprise, type of
enterprise customers, enterprise registration status with an association and time dummy.
We add a square term of maximum education to establish if education externalities have a turning point
beyond which they either increase or decrease. Some studies have established that education
externalities have been found to be monotonic. For example, Mussa (2014) uses quartiles of average
household education and finds that the marginal effects do not switch signs with respect to efficiency
and production uncertainty of maize in Malawi. This finding is confirmed in this study; the coefficient
of the square of maximum education is not significant, implying that education externalities are
monotonic, i.e. they do not increase with levels of education.
Our study finds contrary evidence to that by Matita and Chirwa (2009), who find that enterprises owned
by females tend to be less profitable by about 73% compared to those owned by males. We find no
statistically significant earning differences between males and females. Enterprises in rural areas earn
almost 60% lower than their counterparts operating in urban areas. This is probably due to the fact that
the demand for goods and services is higher in urban areas than in rural areas. Matita and Chirwa (2009)
find similar results.
With respect to industry, the following enterprises are more profitable when compared to agricultural
enterprises: mining and quarrying (OLS only), construction and transport, storage and communication
sector. Older enterprises tend to generate more profit probably because they are more likely to have a
97
Stellenbosch University https://scholar.sun.ac.za
loyal customer base accumulated over time and have accumulated the necessary experience. As
expected, enterprises in the informal sector are less profitable compared to those in the formal sector.
Informal sector enterprises are typically micro and small enterprises with limited access to capital and
entrepreneurial skills.
Table 3.22: Regressions for monthly household non-farm enterprise earnings
Explanatory variables
Maximum years of schooling
Maximum years of schooling squared
Owner experience
Owner experience squared
Enterprise owner is female
Rural
Enterprise industry
Mining and Quarrying
Manufacturing
Construction
Wholesale, Retail and Trade
Transport, Storage and Communication
Financing, Insurance and Business
Community, Social and Personnel
Informal sector
Age of household enterprise
Main enterprise customers
Traders
Other small businesses
Large established businesses
Marketing board (ADMARC)
Other
Not registered with enterprise association
Year=2013
Constant
R-squared
Observations
OLS
1
2
3
4
0.144*** 0.189*** 0.128**
0.128**
(0.018)
(0.059)
(0.053)
(0.053)
-0.002
-0.002
-0.002
(0.002)
(0.002)
(0.002)
0.008**
0.008**
(0.003)
(0.003)
-0.000*** -0.000***
(0.000)
(0.000)
0.055
0.053
(0.069)
(0.068)
-0.567*** -0.567***
(0.117)
(0.118)
Random effects
5
6
7
0.143*** 0.229*** 0.163***
(0.012)
(0.046)
(0.045)
-0.003*
-0.003*
(0.002)
(0.002)
0.008***
(0.003)
-0.000***
(0.000)
-0.008
(0.064)
-0.526***
(0.078)
8
0.158***
(0.045)
-0.003
(0.002)
0.008**
(0.003)
-0.000***
(0.000)
-0.016
(0.063)
-0.531***
(0.078)
0.806*
0.804*
(0.454)
(0.463)
-0.299*
-0.279
(0.171)
(0.179)
1.316*** 1.338***
(0.402)
(0.40)
-0.078
-0.061
(0.184)
(0.193)
0.334
0.346
(0.274)
(0.279)
0.32
0.374
(0.917)
(0.912)
0.029
0.049
(0.236)
(0.242)
-1.182*** -1.183***
(0.123)
(0.124)
0.018*** 0.017***
(0.005)
(0.005)
0.896
(0.565)
0.052
(0.284)
1.675***
(0.476)
0.269
(0.284)
0.808**
(0.347)
0.753
(0.825)
0.173
(0.316)
-1.097***
(0.113)
0.020***
(0.004)
0.908
(0.565)
0.078
(0.285)
1.699***
(0.474)
0.291
(0.286)
0.821**
(0.347)
0.793
(0.817)
0.201
(0.317)
-1.096***
(0.113)
0.019***
(0.004)
0.228
(0.172)
0.645***
(0.216)
1.486***
(0.355)
4.467***
(0.240)
0.102
(0.177)
-0.376*
(0.202)
0.301**
(0.152)
0.840***
(0.237)
0.905*
(0.547)
4.031***
(0.152)
0.197
(0.210)
-0.23
(0.215)
0.230
(0.173)
0.661***
(0.218)
1.502***
(0.354)
4.550***
(0.263)
0.138
(0.177)
-0.377*
(0.203)
0.108
(0.079)
7.312*** 7.059*** 9.455*** 9.382***
(0.189)
(0.336)
(0.470)
(0.476)
0.101
0.101
0.256
0.257
1,702
1,702
1,702
1,702
0.304**
(0.153)
0.855***
(0.238)
0.939*
(0.539)
4.120***
(0.160)
0.231
(0.206)
-0.234
(0.215)
0.119**
(0.060)
7.446*** 6.943*** 8.709*** 8.664***
(0.120)
(0.276)
(0.465)
(0.467)
0.130
0.131
0.288
0.290
1,702
1,702
1,702
1,702
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; robust standard errors in parenthesis
98
Stellenbosch University https://scholar.sun.ac.za
The data also allows us to control for the type of enterprise customers. We use final consumers as the
base category since we expect enterprises selling to these customers to be less profitable compared to
institutional customers such as other small businesses, large established businesses and the national
marketing board (ADMARC). This is confirmed in our study for both OLS and random effects models.
Enterprises that are not registered with any enterprise association make less profit compared to those
that belong to an association according to OLS but in the random effects model, the result is not
significant.
3.9 Measurement error using panel data
In panel data, measurement error, just like non-random attrition bias, is of concern and an attempt is
usually made in the literature to arrive at results that are robust to these concerns (Deaton, 1997). We
discuss two types of measurement error as discussed in Wooldridge (2002), namely measurement in the
dependent variable and measurement in the independent variable.
The assumptions we make about the measurement error are important. First is the usual assumption that
the measurement error has zero mean. However, if this is not the case, then the estimation of the intercept
is affected. The second assumption relates to the relationship between the measurement error and the
explanatory variables included in the model. If the measurement error in the earnings is statistically
independent or uncorrelated with each explanatory variable, then the OLS estimators from equations are
consistent (and possibly unbiased as well). Consequently, measurement error does not bias the
coefficients but only leads to larger standard errors than when the dependent variable is not measured
with error, i.e. it leads to loss of efficiency.
In our model, one can reasonably argue that measurement error is correlated with education since people
with more education tend to report their earnings more accurately. However, in the absence of additional
information, it is difficult to establish if measurement error in earnings is related to any of the
explanatory variables. One solution to measurement error in earnings is to collect more data because
more observations imply a better estimator of variance and consequently reduces the errors in inferences.
This solution is beyond the researcher’s control considering that the data is secondary data32.
Measurement error in the independent variables is considered a more serious problem than measurement
error in the dependent variable. In panel data, the most common method of dealing with measurement
error is first differencing (short and longer differencing). However, our data comes from a panel of two
32
Alternatively, we can use the Malawi Labour Force (2013) cross-sectional data section for comparison purposes.
Nevertheless, there is no guarantee that this is ‘better’ quality data.
99
Stellenbosch University https://scholar.sun.ac.za
periods such that longer differencing is not possible. Moreover, first differencing when T=2 yields the
same results as fixed effects presented in Table 3.14. We are thus unable to correct for the possibility of
such measurement error.
3.10
Comparing income and consumption
There is a good discussion in the literature as to which is a better measure between income and
consumption in the money-metric approach to the measurement of living standards. In this section, we
want to find out if our consumption and income data tell a consistent story as pointed out in Section 3.1.
In general, consumption is considered a better measure of long-term living standards than income
because it is less volatile on an annual basis for most households. Unlike income, consumption is said
to fluctuate much less dramatically. This argument particularly holds for low-income countries
(including Malawi) where there is a large informal sector. Generally, it is difficult and less
straightforward to estimate income from self-employment activities and casual employment in the
informal sector (Haughton & Khandker, 2009). Furthermore, Malawi is largely agricultural based and
incomes are largely vulnerable to seasonal and weather patterns. On the other hand, income is relatively
easy to measure in many high-income countries where salaries and wages form the largest sources of
income. Another advantage of using consumption expenditure over income is that while income usually
fluctuates over an individual’s lifetime, consumption remains relatively stable as a result of smoothing
by saving and borrowing (Blundell & Preston, 1998); McKay, 2000; Duclos & Araar, 2006; Haughton
& Khandker, 2009) as explained in the permanent income hypothesis.
In the IHPS survey, consumption seems to have been collected in a greater detail than income; only two
modules of the household questionnaire are dedicated to collecting information on earnings (Modules E
and N) compared to five modules aimed at collecting information about consumption (Modules G
through K). In order to reduce recall associated with consumption and aspects of agricultural activities,
it was planned that households be visited twice in 201333. The visits were only approximately three
months apart which we believe does not adversely affect recall bias.
The reference period for consumption depends on the items (e.g., the reference period is one week for
food consumption and either one week, one month, three months and 12 months for non-food
expenditures). However, the reference period for earnings is over the past twelve months. According to
Deaton (1997), large increases in prices (inflation) will tend to overstate consumption relative to income
due to the differences in the periods of reporting as noted in the questionnaire. Following this line of
33
About 92% of the households in the survey were visited twice in 2013.
100
Stellenbosch University https://scholar.sun.ac.za
argument, consumption is more likely to be reported in more recent and now higher prices than is
income. However, inflation has not increased a lot over the period under study. Moreover, this inflation
bias (if any) is partly reduced due to the fact that our consumption estimates are both spatially and
temporally deflated.
We compare income and consumption using kernel densities as presented in Figure 3.3. Our analysis
excludes households that split to form new households although the picture does not change even with
the inclusion of the split-off households. Based on the figure, we can see that both consumption and
earnings data seem to tell a consistent story. The data shows that, on average, both percapita income and
percapita consumption have remained the same amongst households that were available in both waves.
Nevertheless, we note that on the lower tail, consumption is generally larger than income meaning that
some households to consume more than their income. This could be due to dissaving or receipts of grants
and grants.
However, using household total income and consumption and, therefore, ignoring household size, the
data shows that income and consumption have increased by 6.3% and 3.0%, respectively over three
years. This shows that adjusting for household size is important especially when there are fewer earners
in the household. Household percapita income (consumption) is calculated by diving total annual
household income (consumption) by household size. We loosely define income as the sum of wages and
earnings from self-employment activities. Household income is the sum of the incomes of all the earners
in the household. The income was then converted into annual figures and adjusted into percapita figures
by dividing by household size.
Figure 3.3: Kernel densities for annual household percapita consumption and income by year
Source: Own computation from IHPS data; the vertical line is the annual poverty line
101
Stellenbosch University https://scholar.sun.ac.za
3.11
Dynamics of household percapita consumption
Table 3.23 gives a breakdown of household percapita consumption by survey year. Despite consumption
remaining unchanged, we note that some consumption components such as housing have declined by
almost 4 percentage points. An analysis of housing expenditure by urban-rural areas shows that the
largest decline came from urban areas (34%) compared to (12%) in rural areas. This disparity is perhaps
due to the fact that most households in rural areas live in their own houses. Surprisingly, households
spend equal proportions on alcohol and education.
Table 3.23: Average household percapita consumption by components and year
2010
Description
Food
Amount
2013
Percent of total
Amount
Percent of total
84,962
54%
88,939
56%
Alcohol
2,949
2%
4,508
3%
Clothing
4,340
3%
4,978
3%
Housing
34,001
22%
27,853
18%
Furnishings
5,493
3%
5,400
3%
Health
2,420
2%
1,958
1%
Transport
7,730
5%
9,865
6%
Communication
5,890
4%
3,972
3%
Recreation
1,484
1%
1,070
1%
Education
2,459
2%
2,707
2%
Hotels and restaurants
1,706
1%
2,478
2%
Miscellaneous
3,855
2%
4,180
3%
Total
157,287
100%
157,907
100%
Source: Own computation from IHPS data
In order to analyse the household consumption poverty dynamics, we categorise households into three
groups with reference to the poverty line34. The first group identifies households below the food poverty
line to capture the extent of extreme poverty (0- K53,262). The second group gives households with
percapita consumption between food and absolute poverty line (K53,262-K85,852). The second group
captures movements with poverty. Finally, the third group consists of non-poor households with
consumption above absolute poverty (>K85,852).
34
Households whose total percapita household consumption is below K53,262 (food poverty line)are considered
ultra-poor while those with consumption below K85,852 are poor. The absolute poverty line is the sum of the food
and non-food poverty lines.
102
Stellenbosch University https://scholar.sun.ac.za
The resulting consumption transition matrix presented in Table 3.24 shows that in 2010, the proportion
below the food poverty line was about 12%. This proportion has declined to about 9% in 2013.
Conversely, the proportion above the food poverty line has increased from about 88% in 2010 to about
91% in 2013. This change is reflected in two parts. First, there is a slight increase in the proportion
above the absolute poverty from 66% in 2010 to about 67% in 2013. Second, there has also been
movement with poverty as reflected in the increase in the proportion between the food and absolute
poverty line from 22% in 2010 to 24% in 2013.
Despite the improvement in household welfare between 2010 and 2013, there are some considerable
transitions, into, within and out of poverty. Out of the 12% that was below the food poverty line, about
4.7% moved upwards but still within absolute poverty while 4.4% moved above the absolute poverty
line. About 2.7% households have been stuck in extreme poverty over the three years.
Table 3.24: Consumption transition matrix
2013
Below food poverty line
(K53,262)
Between food and
2010
absolute poverty line
Above absolute poverty
line (K85,852)
Total
Above
Below food
Between food
poverty line
and absolute
(K53,262)
poverty line
2.7%
4.7%
4.4%
11.8%
(0.004)
(0.006)
(0.005)
(0.011)
3.6%
8.0%
10.6%
22.2%
(0.004)
(0.008)
(0.007)
(0.011)
3.0%
11.3%
51.8%
66.0%
(0.004)
(0.008)
(0.017)
(0.016)
9.3%
24.0%
66.8%
100.0%
(0.008)
(0.013)
(0.016)
absolute
poverty line
Total
(K85,852)
Source: Own computations from IHPS data; figures in parenthesis are standard errors
Amongst those within poverty (22.2%), about 8% remained within poverty while 10.6% moved out of
poverty and a further 3.6% slipped below the food poverty line into extreme poverty. Of those that were
above absolute poverty line (66.0%) in 2010, 3% ended up in extreme poverty, 11.3% became absolutely
poor while 51.8% of the households remained out of poverty.
103
Stellenbosch University https://scholar.sun.ac.za
In the IHS panel survey, respondents were asked about the shocks that affected their households and
how they responded to the shocks. According to the results, the major shocks were increases in the prices
of food and agricultural inputs. In order to smooth their consumption, households embarked on a number
of mitigation measures against the shocks. In 2013, nearly more than one-third (about 35%) of
households reported having used their own savings to cope. This is an increase compared to only 21%
of the households in 2010. There has also been an increase in the number of households using other
mitigation measures between 2010 and 2013. These measures include seeking help from relatives and
friends, help from the government, changes in dietary patterns, selling of assets, use of credit and
spiritual help.
3.12
Conclusion and policy implications
The study sought to examine the returns and externalities to education in Malawi using the IHS3 panel
data set. The random effects results based on the standard Mincerian earnings functions show that the
average rate of return to years of schooling in Malawi is around 8.3% for the full sample. Consistent
with the existing literature, we obtain returns to education that increase with the level of education and
are also higher for females than males. Decomposition of the sample by economic sector reveals that
the rate of return on education is lower in the informal sector at 8.6% in the formal sector and 4.7% in
the informal sector. The rates of return found in this study compare favourably with those observed in
other studies in Malawi and other Sub-Saharan African countries including Ghana, Cameroon and
Rwanda. Sectoral analysis of earnings gives us some more interesting results different than when the
sample is not split. First, we observe that there is a smaller statistically significantly difference in
earnings between males and females in the formal sector than in the informal sector. Second, we find
that returns to education are positive in both sectors but lower in the informal sector. These two findings
highlight the importance of distinguishing between the different types of employment sectors in
estimating rates of return on education in developing countries. The assumption that returns are the same
in both sectors of the economy is not realistic and suggests that studies that fail to take this into account
tend to overstate the returns to education. The coefficient comparison tests help us arrive at a better way
of comparing estimates across models. The results are robust to different model specifications and
sample selection. After accounting for sample selection, the returns to education drop by between 0.7
and 0.8 percentage points. Our results show that education externalities play a significant role in nonfarm enterprises. Education in Malawi has wider social benefits which should not be underestimated.
We also conduct data consistency checks to ensure data quality. This is important if we are to construct
stable and meaningful comparable data for further analysis. Specifically, the chapter highlights issues
related to data quality and inconsistencies in measuring returns to education. Unless the sources of
inconsistencies are explained, it is difficult to conduct further meaningful analysis with the data.
104
Stellenbosch University https://scholar.sun.ac.za
The results have a number of implications with respect to policy. First, the positive and large returns
from schooling suggest that education is a good investment. This is particularly supported by the
presence of education externalities. Access to education, therefore, is important for all. Second, since
returns increase with the level of education, it may be important to invest more resources into higher
level education while not neglecting primary education. The current policy makes primary education
universally free. This is good given the positive social returns of education established in the operation
of non-farm enterprises. We, however, note with concern that the government has recently reduced its
subsidy contribution being offered for tertiary education yet the findings from this study show that
returns to education are highest at the tertiary level. Perhaps, the government should simply redirect
these resources to other areas within tertiary education rather than reduce the contribution. Third,
education policy should encourage female education considering that females with similar skills to
males tend to have higher returns. This may be both economically efficient and equitable although
currently there are fewer females entering the labour market in Malawi. Therefore, the focus of policy
should not only be primary education but also tertiary education where the returns are the highest.
105
Stellenbosch University https://scholar.sun.ac.za
Chapter 4
Patterns of migration and employment in Malawi: Spatial data analysis
4.1 Introduction
There is growing interest in spatial analysis in the literature and it is now widely recognised that the
standard non-spatial econometric techniques produce biased results in the presence of spatial
autocorrelation and heterogeneity. Therefore, spatial analysis is at the research frontier and has recently
become more central in both theoretical and applied econometrics including the fields of labour and
agricultural economics (Anselin, 2003). This growing interest has allowed the development of
econometric techniques that can handle spatial data. There has also been a rapid spread of geographic
information systems (GIS) and the associated availability of data sets containing the location of
observations.
Taking advantage of these developments as well specific literature gaps in Malawi, this study pursues
three main aims: (i) to understand the spatial and temporal patterns of employment and migration in
Malawi through the application of spatial data econometric techniques; (ii) to analyse how long-term
changes in age structure affect labour force participation; and; (iii) to explore the effects of land reform
policy on migration and employment through the use of a difference-in-difference estimation strategy.
The study contributes to the literature in three main ways. First, we add a gender dimension to our
analysis and this has implications for development policy in a poor country such as Malawi. In the
literature, it is argued that economic growth and structural change help remove barriers faced by women
in development and, therefore, improve their labour force participation. However, little attention has
been given to the empirical analysis of such assertions especially in the context of developing countries
(Gaddis & Klasen, 2014). The study contributes to this strand of literature. The second contribution of
this study relates to the application of spatial panel data econometric methods to the study of
employment and migration in Malawi. This is important because the economic phenomena we are
studying are essentially spatial in nature in the sense that we expect their values in a given location to
be determined by the values in neighbouring locations. To our knowledge, such a study has not yet been
done in Malawi. With respect to this objective, we compare results obtained with the application of
spatial econometric models with estimation results obtained from non-spatial panel data models. Finally,
by matching census data GIS codes that are consistent over time at the level of small geographical places,
it is now possible to integrate census data with other data for similar spatial analysis. This is relevant
because all nationally representative data sets collected by Malawi’s National Statistical Office (NSO)
use censuses as the sampling frame. For example, one can now merge census data with Demographic
Health Survey (DHS) data aggregated at the district or small geographical area. This kind of data
integration could be an area for further study.
106
Stellenbosch University https://scholar.sun.ac.za
4.2 Data and methods
4.2.1
Data
Using data from the Integrated Public Use Microdata series (IPUMS)-International, Minnesota
Population Centre, we matched district and small geographical area level GIS data codes to create
boundaries that are consistent over time, namely for 1987, 1998 and 2008. This matching gives us a
panel of districts and small geographical areas (also called traditional authorities) over which we conduct
our spatial analysis while also exploring the advantages of working with panel data. Based on previous
work, most of the indicators are heterogeneous at small geographical areas rather than regions where
there seems to be uniformity.
At this stage, it is important to provide a brief background on Malawi’s administrative units. The country
has three regions, namely Northern, Central and Southern regions. The regions are divided into districts
which are further subdivided into small geographical areas called traditional areas. Figure 4.1 shows the
administrative boundaries for Malawi.
Malawi has four cities: Lilongwe, located in the Central region, Mzuzu in the Northern region, and
Zomba and Blantyre in the Southern region. Since Lilongwe, Zomba and Blantyre are named after
districts, the word “city” is added to the district name to signify the city. Consequently, “Lilongwe”
refers to Lilongwe district while “Lilongwe City” to the city (National Statistical Office, 2008).
While the regions have remained stable over time, the number of districts and small geographical places
in Malawi has been changing over time35. Through the changes, Machinga district was split into
Machinga and Balaka, Mulanje was split into Mulanje and Phalombe, and Mwanza was split into
Mwanza and Neno. Nkhata-Bay and Likoma are combined in the data (National Statistical Office,
2008). As expected, these changes create a problem of inconsistent comparability of the spatial units
over time because we may not be working with the same observations as desired.
In order to allow for consistent comparison over time, we dissolved the new districts or traditional
authorities and merged them into what they were as at 1987. The end result is data for a total of 24
districts and 178 traditional authorities instead of the 28 districts and 224 traditional authorities as at
2008. These form our units of analysis. Consequently, our variables of interest are collapsed by
traditional authority as proportions by year and gender, depending on the level of analysis. A detailed
list consisting of regions, districts and traditional areas is provided in Table A4 in the appendix.
35
According to National Statistical Office (2008), there were 23 districts in 1966; 24 districts in 1977; 24
districts in 1987; 28 districts and 4 cities in 1998 and 27 districts and 4 cities in 2008.
107
Stellenbosch University https://scholar.sun.ac.za
Figure 4.1: Map of Malawi showing administrative boundaries
Source: Own computation from census data
4.2.2
Theoretical framework
Migration involves the movement of people from their usual places of residence to another area over a
given period of time. Theories of migration try to explain why people migrate and where they come
from and where they are going. All models of migration assume that migration is voluntary, although
some forms of migrations are involuntary. The theoretical perspectives of migration from the field of
economics and other disciplines can be broadly grouped into either of two disciplines, namely the
disequilibrium and equilibrium perspectives (Greenwood, 2005). We now discuss these two alternative
perspectives.
The disequilibrium framework has been the underlying model for understanding and the study of
migration among economists dating as far as before the late 1970s. In this framework, migration was
thought to be driven by differences in wages across regions or sectors. Consequently, people would
migrate from one area to another in search of market opportunities in the form of wages and earnings.
This framework has been criticised by the advocates of the equilibrium hypothesis who argue that the
geographical differences in wages or incomes are compensating and do not reflect opportunities from
108
Stellenbosch University https://scholar.sun.ac.za
which people can benefit. Instead, the equilibrium models argue that migration depends on imbalances
in amenities. According to this theory, the patterns of migration are such that people move from amenitypoor areas into amenity-rich areas. However, as people migrate into amenity-rich areas, wages tend to
decline and the prices of locally produced goods and services are driven up. Consequently, as wages
and prices continue to diverge across regions until they compensate households for the differences in
the amenity bundles supplied by the various regions. The common feature in both the disequilibrium
and equilibrium models is that migration is motivated by regional imbalances, with the former dependent
on wages while the latter on amenities. Furthermore, regardless of the model used, migration is
essentially spatial (Cooke, 2013; Greenwood, 2005).
Within the equilibrium approach are the gravity and modified gravity models, which have foundations
in the gravity law of spatial interaction. In the basic gravity model, migration is a function of the distance
between any two places as well as the population sizes of the origin and destination. It has been argued
that the basic gravity model does not fully capture all the factors behind migration. The modified gravity
models, therefore, include several additional factors including income, unemployment rates, the degree
of urbanisation, amenities and taxes, among other factors. Since the 1960s, the modified gravity models
hold a prominent place in the literature and combine elements of both the disequilibrium and equilibrium
hypotheses of migration (Greenwood, 2005).
Human capital theory is another theory used to explain migration. This model rests on the disequilibrium
approach in which individuals are assumed to be economic agents seeking to maximise utility from
leisure and income. Implicitly, this assumes that an individual's supply of labour is dependent on the
wage rate. Individuals will, therefore, migrate only when the expected return exceeds the costs of
migration among other costs. The human capital theory has been central to the understanding of
migration and faced no competing theory for almost 20 years. Extensions of the human capital model
are the ‘spatial job-search models’. Just like the human capital model, individuals migrate to other
locations in search of opportunities but have reservation wages and would reject opportunities that give
wages below the reservation wage. The reservation wage is itself dependent on several factors, including
the rate of unemployment. Presumably, for unemployed people the reservation wage may fall over time
(Willekens, 2008; Greenwood, 2005).
4.2.3
Spatial autocorrelation
The concept of spatial dependence36 is based on the first law of geography by Waldo Tobler which states
that “Everything is related to everything else, but near things are more related than distant things"
36
We use spatial dependence and spatial autocorrelation interchangeably.
109
Stellenbosch University https://scholar.sun.ac.za
(Tobler, 1970, p. 236). Based on this law, two or more objects that are spatially close tend to be more
similar to each other with respect to a given attribute Y than are spatially distant objects. In this case, we
expect positive spatial autocorrelation among geographically close regions. Measures of spatial
autocorrelation can be grouped into two broad categories, namely global and local indices.
Global indices express the overall level of similarity between spatially close locations in a given study
area with respect to a variable Y. A global measure summarises the variable of interest in a single value
and, therefore, only shows the average degree of the spatial distribution of the phenomena of interest.
Local spatial autocorrelation measures overcome this limitation by detecting spatial clusters which
provide greater detail than the global measures. Thus, local measures may help identify the locations
that contribute most to the overall pattern of spatial clustering which some have referred to as hotspots
(Sokal, Oden, and Thomson, 1998).
The most commonly used tests for spatial autocorrelation are the Moran’s
and Geary’s c statistics
respectively named after their developers, Moran (1948) and Geary (1954). It is the practice in the
literature to report both statistics as a means of robustness checks for each other in the detection of
spatial dependence. A detailed discussion of these indices is provided in Pisati (2001) and Sokal et al.
(1998). We begin our discussion with the global measures.
Moran’s is given by
I=
Where: w
∑
∑
(4.1)
denotes the elements of the spatial weights matrix W corresponding to the location pair
(i, j), Z = Y − , Y represents the value taken by a variable of interest Y at location i;
mean variable of ;
=∑ ∑
and
=∑
denotes the
⁄ .
Under the null hypothesis of no global spatial autocorrelation, the expected value of I is given by
(4.2)
E( ) = −1/( − 1)
On the one hand, if I is larger than its expected value, then the overall distribution of variable Y can be
seen as characterised by positive spatial autocorrelation, meaning that the value taken on by Y at each
location i tends to be similar to the values taken on by Y at spatially contiguous locations. On the other
hand, if I is smaller than its expected value, then the overall distribution of variable Y can be seen as
characterised by negative spatial autocorrelation, meaning that the value taken on by Y at each location
i tends to be different from the values taken on by Y at spatially contiguous locations.
110
Stellenbosch University https://scholar.sun.ac.za
Inference is based on z-values, computed by subtracting E( ) from I and dividing the result by the
standard deviation of I as follows:
z =
( )
( )
(4.3)
Geary’s c is calculated as
c = ( − 1)
∑
∑
(4.4)
Under the null hypothesis of no global spatial autocorrelation, the expected value of c equals 1. If c is
larger than 1, then the overall distribution of variable Y can be seen as characterized by negative spatial
autocorrelation; on the other hand, if c is smaller than 1, then the overall distribution of variable Y can
be seen as characterized by positive spatial autocorrelation. As in the case of Moran’s I, inference is
based on z-values, computed by subtracting 1 from c and dividing the result by the standard deviation
of c:
z =
(4.5)
( )
Both z and z are assumed to follow a normal distribution (asymptotically) such that their significance
can be evaluated through the use of a standard normal table (Pisati, 2001).
So far we have looked at global measures. We denote the local measures of autocorrelation by subscript
to denote location. Therefore, Moran’s I and Geary’s c detect positive and negative spatial
autocorrelation around a given location. Positive I z-values and negative c
z-values indicate
clustering of similar values of Y around location i, that is, positive local spatial autocorrelation, while
negative I z-values and positive c z-values indicate clustering of dissimilar values of Y around
location , that is negative local spatial autocorrelation.
4.2.4
Spatial panel data models
Spatial regression models estimate the relationship between a dependent variable
explanatory variables
and one or more
while taking into account the spatial dependence among observations. In the
literature, it is argued that when the observations are spatial units or locations, the standard regression
models are usually misspecified because of the presence of spatial dependence (Pisati, 2001).
111
Stellenbosch University https://scholar.sun.ac.za
Several Stata tools exist for fitting spatial regression models and their application depends on whether
one is working with cross-sectional37 or panel data. In this study, we fit spatial panel data regressions
through the use of xsmle by Belotti, Hughes, and Mortari (2013). This approach fits the models through
the maximum likelihood (ML) estimation, which is the widely used method for spatial models. Anselin
(2003) provides a good discussion on the different approaches for estimating spatial regression models
including ML estimation, the method of moments estimators and spatial two-stage least squares.
Our discussion begins by considering the following general specification of spatial panel models from
which specific types of spatial panel models are derived:
Where:
=
+
denotes an
+
+
;
+
other;
denotes an
rowdenotes
denotes the spatially lagged
dependent variable (DLAG) which assumes that the value taken by
affected by the values taken by
denotes an
denotes the number of locations;
matrix of observations on the explanatory variables;
The term
(4.6)
1 vector of observations on the outcome variable;
standardised and inverse distance38 spatial weights matrix;
and
+
=
in each geographical area is
in the neighbouring areas.
1 vector of normally distributed homoscedastic errors uncorrelated to each
the spatial matrix for the spatially lagged independent variables; Z the spatially lagged
regressors; and
the spatial matrix for the idiosyncratic error. The component
or random effects and
is the time effect. The terms
and
is the individual fixed
denote the spatial autoregressive
parameters. From the general model specification given in (4.6), we derive and discuss five spatial panel
models which are normally considered in the literature.
First is the Spatial Auto-regressive model (SAR) with lagged dependent variable ( =
given by:
=
+
+
(4.7)
+
where the standard SAR model is obtained by setting
37
For example, see Stata’s
= 0) and it is
by Pisati (2001) and
= 0.
by Jenty (2010) for a discussion on spatial
cross-sectional analysis.
38
Row standardised means that all nonzero weights are rescaled so that they sum up to 1 within each row and
inverse distance means that locations that are closer are given higher weights than those farther apart.
112
Stellenbosch University https://scholar.sun.ac.za
Second is the Spatial Durbin model (SDM) with time-lagged dependent variable ( = 0) and it is
specified as:
=
+
+
+
(4.8)
+
where the standard SDM model is obtained by setting τ = 0. The package xsmle allows the use of a
different weighting matrix for the spatially lagged dependent variable W and the spatially lagged
regressors D together with a different set of explanatory and spatially lagged regressors Z . The default
is to use W = D and X = Z .
Third is the Spatial Autocorrelation Model (SAC) ( =
weighting matrix for the spatially lagged dependent variable
=
+
with
+
+
with
=
and the error term . It is given as:
+
=
The fourth model is the Spatial Error Model (SEM) ( =
=
= 0) and it allows the use of a different
+
(4.8)
=
= 0) and it is given as:
Finally, we have the Generalised Spatial Random Effects (GSPRE) model ( =
specification is as follows:
=
+
with
=∅
+
and v =
+
=
(4.9)
= 0) whose
(4.10)
Table 4.1 provides the summary of the spatial available fixed and random effects spatial models which
are the most commonly estimators with panel data39. The SAR and SDM allow for a time-lagged
dependent variable which may not be feasible for a panel with
39
= 2.
Williams (2015), Clark and Linzer (2015); Wooldridge (2002) provide a detailed discussion on panel data
methods and the choice between random and fixed effects.
113
Stellenbosch University https://scholar.sun.ac.za
Table 4.1: Summary of spatial panel model options
Spatial
Generalised Spatial
Spatial
Spatial Durbin Spatial Error
Autocorrelation
Panel Random
Description
Autoregressive
Model (SDM) Model (SEM)
(SAC)
Effects (GSPRE)
Model (SAR)
Random effects (RE)
No time lagged
Yes
Yes
Yes
N/A
Yes
dependent variable
With time lagged
Yes, DLAG
Yes, DLAG
N/A
N/A
N/A
dependent variable
Fixed effects (FE)
No time lagged
Yes
Yes
Yes
Yes
N/A
dependent variable
With time lagged
Yes, DLAG
Yes, DLAG
N/A
N/A
N/A
dependent variable
Source: Compiled from Belotti et al. (2013
We use fixed effects models because we are able to control for the effects of time-invariant variables
with time-invariant effects. Our main results are based on SAR because it combines coefficients for both
the weighting matrix and the matrix for the idiosyncratic error term, namely
and . SEM and SAC
are used as robustness checks. Since SDM uses a time-lagged dependent variable, it is not possible to
use for a panel with two periods of data. The non-spatial results are obtained with the help of the standard
pooled ordinary least squares (OLS) and fixed effects estimation models.
4.3
Descriptive analysis
In this section, we show the spatial distribution of some of the variables of interest on the map of Malawi.
This presents a first step in the spatial analysis of the data. Trends in fertility, population age structure,
labour force participation, occupational mobility and migration are also discussed.
4.3.1
Spatial distribution of employment, education, assets and fertility
From Figure 4.2 to 4.5 we show choropleth maps for some of our indicators of interest drawn using two
Stata tools, namely
ℎ 2
and
40.
These maps reveal some spatial interaction and
heterogeneity in Malawi in terms of proportions of employment status, type of employment, education,
asset index, years of schooling and fertility rates. The long white strip on the map with no (or missing)
data is Lake Malawi. The two dots near the Lake are Likoma and Chizumula Islands.
From the maps, we get an overall idea of the spatial distribution of the aggregated economic indicators
under study. For example, most of the clustering in terms of the inactive population occurs in the
40
ℎ 2
converts shape boundary files to Stata datasets and
114
is for visualisation of spatial data.
Stellenbosch University https://scholar.sun.ac.za
Northern Region. The two main reasons for being inactive cited in the data are schooling (52%) and
involvement in housework (33%). We also observe some clusters of self-employment activity in some
parts of the Central and Southern regions of the country. Education levels are lowest in the Southern
region where there are also high levels of fertility and low asset ownership. The statistics generally agree
with what is known about Malawi in terms of poverty from Chapter 2.
Figure 4.2: Spatial distribution of status in employment for 2008
Source: Own computation from census data
Figure 4.3: Spatial distribution of employment type for 2008
Source: Own computation from census data
115
Stellenbosch University https://scholar.sun.ac.za
Figure 4.4: Spatial distribution of educational attainment for 2008
Source: Own computation from census data
Figure 4.5: Spatial distribution of asset index, schooling and fertility for 2008
Source: Own computation from census data
4.3.2
Fertility trends
In Table 4.2, we present patterns of age-specific fertility rates41 for three years of census data, namely
1987, 1998 and 2008. The unadjusted fertility first declined from 5.53 children per woman in 1987 to
4.75 in 1998 before slightly increasing to 5.05 in 2008. Overall, there has been a decline in total fertility
rate over the two decades under study. With respect to the child-bearing ages, the patterns in the age41
The age-specific fertility rate is calculated as the number of births in the last twelve months divided by the
number of women for each age group. The result is then multiplied by 5 (the number in each of the age group).
116
Stellenbosch University https://scholar.sun.ac.za
specific fertility rates have not been uniform. They consistently remained high in the 20-24, 25-29 and
30-34 age groups but particularly increased for women in the 20-24 age group.
Table 4.2: Age-specific fertility rates (ASFR) and total fertility rates (TRF)
Year
1987
Age group
No. of women
No. of births in last 12 months
Age-specific fertility rate
15 – 19
20 – 24
25 – 29
30 – 34
35 – 39
40 – 44
45 – 49
Total
37,732
34,103
28,211
21,072
20,147
14,128
12,538
167,931
4,477
8,174
6,497
4,230
3,314
1,375
677
28,744
0.59
1.20
1.15
1.00
0.82
0.49
0.27
5.53
15 – 19
20 – 24
25 – 29
30 – 34
35 – 39
40 – 44
45 – 49
Total
55,640
54,149
39,260
29,465
24,550
17,782
16,637
237,483
5,537
12,029
7,981
5,049
3,251
1,415
699
35,961
0.50
1.11
1.02
0.86
0.66
0.40
0.21
4.75
7,558
16,863
12,797
7,614
4,252
1,495
513
51,092
0.55
1.22
1.12
0.94
0.72
0.34
0.15
5.05
1998
2008
15 – 19
68,134
20 – 24
69,177
25 – 29
57,291
30 – 34
40,575
35 – 39
29,428
40 – 44
21,712
45 – 49
17,057
Total
303,374
Source: Own computation from census data
Studies have linked patterns of female employment to fertility rates. For example, the literature on
feminisation U-hypothesis42 suggests that the female labour force participation rate exhibits a U-shaped
trend with respect to fertility rates as countries develop. At the early stages of economic development,
when gross domestic product (GDP) is also still very low, most women tend to participate in the labour
force even when fertility rates are still high. The women are said to be able to combine economic activity
with child-bearing (Gaddis & Klasen, 2014). In the context of Malawi, much of the work is on household
42
The feminisation hypothesis states that economic growth first lowers the participation of women in the labour
market and increases it at later stages of economic development.
117
Stellenbosch University https://scholar.sun.ac.za
enterprises and small-scale household farms. However as the countries become richer, the shift in the
structure of the economy towards manufacturing combined with the emergence of a formal sector results
in lower levels of female labour force participation and employment while fertility is high.
Despite the improvement in economic activity and favourable structural change, the low levels of female
education and the difficulty of combining wage employment with child-bearing contribute to the low
participation of females in the emerging opportunities in industry and formal sector expansion. This may
as well be influenced by social restrictions against females working in industries as well as allowing
married women with children to get employment outside of the home.
The subsequent increase in female participation rate could be as a result of a combination of several
factors. One possible explanation could be the expansion of education among females. However, it has
generally been the case that labour force participation is also quite high amongst those without
education. Therefore, improvement in education does not explain everything. Another explanation could
be the emergence of attractive employment opportunities for women in the white-collar service
industries which give women a chance to enter the labour force without many restrictions. In this case,
a decline in fertility rates possibly frees up the time women spend on child-bearing which can be in the
end reallocated to other work. Improvements and increased access to child-care facilities may also have
allowed women to combine work outside the home with raising children.
4.3.3
Changes in population age structure and labour supply
Data for this section are drawn from the October 2011 International Labour Organisation (ILO)
Economically Active Population, Estimates and Projections (EAPEP) (6th ed., October 2011). ILO’s
EAPEP database contains country estimates and projections of the total population, the activity rates
and the economically active population by sex and age groups. In Table 4.3, we use the EAPEP data to
show changes in population and labour force between 1987 and 2008.
118
Stellenbosch University https://scholar.sun.ac.za
Table 4.3: Changes in economically active population and labour force by sex (1987-2008)
Population
Age group
1987
1998
2008
Change
(1987-2008)
Labour force
1987
1998
2008
Change
(1987-2008)
Male
15-19
20-24
25-29
30-34
35-39
40-44
45-49
50-54
55-59
60-64
65+
Total
Female
15-19
20-24
25-29
30-34
35-39
40-44
45-49
50-54
55-59
60-64
65+
Total
407,691
558,786
782,638
343,770
493,273
629,291
284,125
372,529
533,372
223,757
298,164
449,713
192,948
246,465
314,596
157,269
195,192
236,352
120,458
167,955
192,212
97,986
147,204
155,313
77,201
110,054
137,155
57,122
87,802
119,605
92,185
153,561
199,020
2,054,512 2,830,985 3,749,267
420,190
557,397
775,215
363,609
495,638
626,106
306,338
386,990
522,124
247,318
318,279
426,374
219,557
267,212
299,434
178,186
214,836
241,890
135,505
181,539
216,010
110,125
155,631
185,938
88,609
113,442
162,105
68,982
90,758
137,274
116,213
173,485
229,905
2,254,632 2,955,207 3,822,375
92%
83%
88%
101%
63%
50%
60%
59%
78%
109%
116%
82%
84%
72%
70%
72%
36%
36%
59%
69%
83%
99%
98%
70%
139,838
165,401
378,598
249,233
350,717
369,706
262,531
347,942
515,089
215,254
290,114
438,683
187,160
241,536
311,360
152,708
191,483
233,980
116,965
165,100
190,379
94,850
144,407
153,884
74,808
107,743
135,824
54,387
85,344
118,235
81,330
144,040
193,049
1,629,065 2,233,826 3,038,788
171%
208,414
242,468
485,026
267,253
367,763
439,787
241,088
315,784
510,079
203,295
268,627
419,522
185,745
231,138
295,232
155,022
189,915
239,644
120,193
163,204
214,815
98,562
140,846
178,502
79,659
103,119
155,785
59,351
81,682
131,613
87,863
145,727
202,316
1,706,446 2,250,274 3,272,321
133%
48%
96%
104%
66%
53%
63%
62%
82%
117%
137%
87%
65%
112%
106%
59%
55%
79%
81%
96%
122%
130%
92%
Source: Own computation from EAPEP data
Firstly, we observe that the labour force growth is higher than that of the population. The magnitude is
larger for females than for males. For males (females), the labour force has grown by 87% (92%)
compared to 82% (70%) for the population. Secondly, the percentage of the youth has been on the
increase. Using the International Labour Organisation (ILO) definition, the youth (15-24) made up
37.7% of the total working-age population in 2008 compared to 36.6% in 198743. This explains some of
the large increases that have taken place within the youth groups in terms of both the economically
active population and labour force participation.
These observed changes in the age structure of the population have significant implications on the supply
of labour and growth. As we see in Figure 4.6 the population pyramid for 2008 is different from the
43
Estimates from the 2010 IHS3 put the figures at 38% and 66% for the ILO and SADC definitions, respectively.
119
Stellenbosch University https://scholar.sun.ac.za
other two in terms of labour force participation rates especially for the youth (ages 15-29) where the
numbers have expanded, suggesting that a majority of the youths now participate in the labour market.
Figure 4.6 : Pyramids showing populations and labour force participation by sex (1987)
Figure 4.7: Pyramids showing populations and labour force participation by sex (1998)
120
Stellenbosch University https://scholar.sun.ac.za
Figure 4.8: Pyramids showing populations and labour force participation by sex (2008)
Source: Own computation from census data
Changes in population depend on migration, fertility and mortality rates. The picture is, however, not
complete because education levels are also increasing and this has implications on fertility and the time
people spend in education (and therefore entry into the labour market). Fertility rates have not fallen
significantly in Malawi between 1987 and 2008 as earlier shown. The mortality rate (measured by crude
death rate) has dropped over time from 20.39 deaths per 1000 in 1987 to about 12.13 per 1000 in 2008.
It is not straightforward to know whether the combined changes of fertility, mortality and migration
translate into significant differentials in the size, growth and structure of the labour force. In Table 4.4
we show the proportions of the population and labour force belonging to each of the age groups. With
the exception of the (20-24) age group, the overall changes in the population proportions seem to mirror
those in the proportions in the labour force.
121
Stellenbosch University https://scholar.sun.ac.za
Table 4.4: Proportions of population and labour force by age group
Age groups 15-19 20-24 25-29 30-34 35-39
Population
1987
19.2 16.4 13.7 10.9 9.6
1998
19.3 17.1 13.1 10.7 8.9
2008
20.6 16.6 13.9 11.6 8.1
Labour force
1987
10.4 15.5 15.1 12.5 11.2
1998
9.1 16.0 14.8 12.5 10.5
2008
13.7 12.8 16.2 13.6 9.6
Source: Own computation from EAPEP data
40-44 45-49 50-54 55-59 60-64 65+ Total
7.8
7.1
6.3
5.9
6.0
5.4
4.8
5.2
4.5
3.8
3.9
4.0
2.9
3.1
3.4
4.8 100
5.7 100
5.7 100
9.2
8.5
7.5
7.1
7.3
6.4
5.8
6.4
5.3
4.6
4.7
4.6
3.4
3.7
4.0
5.1 100
6.5 100
6.3 100
We also examine trends in the age dependency ratio and the proportions of the total labour force by age
groups. We define dependency ratio as the ratio of dependents (people younger than 15 or older than
64) to the working-age population (those aged 15-64). Based on this definition, the dependency ratio
was 98% in 1987, declined to 94% in 1998 before increasing again to 96%.
4.3.4
Spatial and temporal patterns of migration
Migration data is only available for 1987 and 2008 and these can be consistently compared over time.
The censuses collected migration data at the district level and geographical mobility was only defined
where an individual’s district of residence one year ago was different from the district of residence at
the time of the census. Movements within a district or city are not considered as migration. However,
for districts which share boundaries with cities, if a person moves from rural areas into a city (e.g. from
Lilongwe rural to Lilongwe city) and vice versa is recorded as a movement.
Migration is thought to be responsible for changes in spatial distribution of populations and labour force
in countries. For example, in Malawi, the urban population amongst individuals aged between 15 and
64 years has steadily increased from 11.7% in 1987 to 16.0% in 1998 to 16.6% in 2008. Although these
are not very large changes, the trend is likely to increase and this may have implications for a wide range
of social, economic and political dynamics in the country, including the general employment situation.
The expansion of the labour force must be matched by the creation of new productive jobs into which
migrants can be absorbed. While the Northern and Central regions have registered increases, the
proportion of people living in the Southern region has dropped from 49.3% in 1987 to 44.5% in 2008.
Our analysis assumes, for the sake of simplicity, that Malawi is a closed system. We, therefore, only
concentrate on internal migration. Moreover, the census data provide information only on immigrants
into the country and not on Malawians now abroad but who lived in the country a year earlier. Two main
patterns of migration can be identified as taking place in the country, namely regional and district level
migration.
122
Stellenbosch University https://scholar.sun.ac.za
4.3.4.1 Regional movements
The first level of analysis involves movements across Malawi’s three administrative regions. The
transition frequencies are given in Table 4.5 and accompanying probabilities in Table 4.6.
As one might expect, the majority of the people remained in their home region with only a small
percentage moving across. Approximately, 96.73% of the males and 97.71% of females resident in the
Northern region in 1986 were still there in 1987. A slightly higher percentage than in the North remained
in the Centre and South. The patterns largely remained the same in 2008 although slightly more people
had migrated from their original region of residence. Much of the people are migrating towards the
Centre and in the long-run, it might be expected to have the majority of the people. The reasons for
migration are not captured in the census data. However, most migration seems to be undertaken as a
way of improving incomes and economic status. Other reasons might include education, marriage and
retirement.
Table 4.5: Inter-regional migration transition frequencies by gender
Year
North
1987 North 204,270
Centre
4,990
South
4,270
Total
213,530
2008
North 365,230
Centre 24,830
South
17,580
Total
407,640
Male
Centre
South
4,030
2,870
720,780
12,790
16,460
876,540
741,270
892,200
22,780
1,312,940
93,200
1,428,920
15,510
54,470
1,389,170
1,459,150
Total
211,170
738,560
897,270
1,847,000
North
231,620
3,470
2,700
237,790
403,520
1,392,240
1,499,950
3,295,710
404,750
22,870
13,130
440,750
Female
Centre
South
3,280
2,150
778,960
9,390
9,800 1,027,740
792,040 1,039,280
22,810
1,381,350
70,840
1,475,000
14,470
52,400
1,537,130
1,604,000
Total
237,050
791,820
1,040,240
2,069,110
442,030
1,456,620
1,621,100
3,519,750
Source: Own computation from census data
Table 4.6: Inter-regional migration transition probabilities by gender
Year
1987 North
Centre
South
Total
2008
North
Centre
South
Total
North
96.73
0.68
0.48
11.56
Centre
1.91
97.59
1.83
40.13
Male
South
1.36
1.73
97.69
48.31
90.51
1.78
1.17
12.37
5.65
94.30
6.21
43.36
3.84
3.91
92.61
44.27
Source: Own computation from census data
Total
100
100
100
100
North
97.71
0.44
0.26
11.49
100
100
100
100
91.57
1.57
0.81
12.52
123
Female
Centre
South
1.38
0.91
98.38
1.19
0.94
98.80
38.28
50.23
5.16
94.83
4.37
41.91
3.27
3.60
94.82
45.57
Total
100
100
100
100
100
100
100
100
Stellenbosch University https://scholar.sun.ac.za
4.3.4.2 District level migration
As we will see, the district level migration patterns reveal another important aspect of migration in
Malawi namely, rural-rural migration driven by people searching for agricultural land, considering the
country’s dominantly agricultural base and high fertility rates. Just like with regional movements, we
capture district movements as a transitional probability matrix obtained by a cross tabulation of the
current and previous district of residence. The population proportions that migrated in each district by
gender and year are shown in Figure 4.9. We also provide a detailed analysis of population figures for
each district by gender and year, the number of people that moved between districts and the
accompanying proportions in the appendix in tables A6, A7 and A8.
Figure 4.9: Proportions that moved between districts
Source: Own computation from census data
124
Stellenbosch University https://scholar.sun.ac.za
Malawi has had policies that can potentially affect migration patterns such as those shown in the figure.
We discuss two of these government policies. Firstly, the government has been supporting the
establishment of rural growth centres which act as focal points for development within rural areas. This
project falls under the Integrated Rural Development Programme outlined in the MGDS II. This might
have the effect of keeping people in rural areas. Secondly, the Malawi Community Based Rural Land
Development Project (CBRLDP), which was launched in 2004 and ran through 2009, enabled landless
or land-poor households to voluntarily purchase plots from fallow estates and resettle there. The
historical problem of land shortage is evident in Thyolo and Mulanje districts, which are traditionally
home to tea estate farms since the early 1900s. The land reform policy was aimed at addressing emergent
social conflicts related to unequal access to land. As a pilot programme, about 15,000 households from
these rural districts were allowed to resettle in other rural districts of Machinga and Mangochi, but later
it was expanded to Ntcheu and Balaka to ease pressure on land prices (World Bank, 2012). The policy
has implications for migration patterns we observe post-2004 in the Southern region districts affected
by CBRLDP and adjacent areas. The observed patterns of migration due to this policy are largely ruralrural migration since the programme targeted Malawi’s rural poor. We do not have data on these specific
15,000 households but in the sections that follow we simply take advantage of the natural experiment
by looking at conditions of the affected districts (as a whole) before and after the policy.
4.3.5
Spatial autocorrelation in variables
Table 4.7 extends the spatial comparisons in the maps to show the calculated Moran’s and Geary’s
statistics for our variables of interest. For lack of space, we only report statistics based on the pooled
sample, but we have also computed them on year by year basis where we find that they are generally
stable over time as shown in the appendix in Tables A3, A4 and A5. The table shows that there is
similarity between spatially close traditional authorities as shown by positive statistically significant
Moran’s and Geary’s
values. Technically, spatial dependencies justify spatial regression analysis.
Table 4.7: Results showing spatial dependencies in variables (pooled 1987, 1998 and 2008)
Variables
Moran's I
SD
p-value
Geary's c
SD
p-value
Migration
0.397
0.009
0.000***
0.612
0.011
0.000***
Employed
0.063
0.009
0.000***
0.930
0.013
0.000***
Years of schooling
0.167
0.009
0.000***
0.793
0.013
0.000***
Sex
0.153
0.009
0.000***
0.857
0.013
0.000***
Age
0.056
0.009
0.000***
0.954
0.015
0.002***
Married
0.127
0.009
0.000***
0.863
0.012
0.000***
Dependency ratio
0.075
0.009
0.000***
0.927
0.013
0.000***
Note: *, **, *** denote significance at 10%, 5% and 1% levels
125
Stellenbosch University https://scholar.sun.ac.za
We compute and plot Moran’s
spatial correlograms for employment and migration based on
cumulative distance bands, namely (0; 100,000 metres], (0; 200,000 metres],…, (0; 700,000 metres]
presented in Figure 4.10 and Figure 4.11, respectively.
Figure 4.10: Spatial autocorrelation for employment
Moran's I spatial correlogram
(mean) employed
I
0.10
0.05
0.00
0-100000
0-200000
0-300000
0-400000
0-500000
Distance bands
0-600000 0-700000
Figure 4.11: Spatial autocorrelation for migration
Moran's I spatial correlogram
(mean) migration
0.80
I
0.60
0.40
0.20
0.00
0-100000
0-200000
0-300000
0-400000
0-500000
Distance bands
Source: Own computation from census data
126
0-600000 0-700000
Stellenbosch University https://scholar.sun.ac.za
The correlograms help us to draw two conclusions. Firstly, spatial dependence is larger for shorter
distance bands but diminishes with larger cumulative bands and even becomes negative for employment
at large distances. Secondly, spatial dependence is higher for migration than for employment.
Employment has lower dependence partly because of the fact that most of Malawi is agriculturally based
and there seems to be little variation.
Figure 4.12 shows pairs of choropleth maps of two measures of local spatial autocorrelation (Moran’s
and Geary’s ) for employment and migration in Malawi. For each small geographical area, the maps
confirm some degree of similarity between neighbouring areas with respect to both employment and
migration. Generally, the two measures yield consistent results in spatial patterns.
Figure 4.12: Measures of local spatial autocorrelation for employment and migration
Source: Own computation from census data
4.4
Spatial panel regression results
The spatial descriptive results and spatial autocorrelation statistics presented in Section 4.3.1 indicate
that there is some clustering in our variables of interest at different locations. However, according to
Anselin (1992b), this does not explain why the clustering occurs. We, therefore, extend our previous
analysis through spatial panel regression analysis.
Our analysis is segregated by gender and the 178 small geographical areas form our spatial units. We
begin our analysis with the standard non-spatial ordinary least squares and fixed effects regressions.
However, as discussed in Section 4.2.3, these are inadequate in providing answers to our question due
127
Stellenbosch University https://scholar.sun.ac.za
to the problem of misspecification when spatial locations form our unit of analysis and in the presence
of spatial dependence (Pisati, 2001). The rest of the models are spatial fixed effects models whose results
we compare with OLS and non-spatial fixed effects (FE).
In this section, we specifically explore the effects of land reform policy, explained earlier, on migration
and employment in Malawi through a difference-in-difference (DID) estimation strategy. Considering
that migration and employment are inherently spatial, we expect the programme to have spill-over
effects affecting surrounding districts.
We now provide a brief discussion on the set-up of the DID. Firstly, we created a dummy variable called
time to indicate the time when the treatment started. The land reform started in 2004 and we assign a
value of “0” to the period before 2004 and “1” post-2004. Secondly, we create a dummy variable
treatment to identify the group exposed to the treatment. Small geographical areas in Thyolo and
Mulanje are assigned a value of “1” and “0” is assigned for the rest of the country. Finally, we create an
interaction term between time and treatment as follows (DID = time ∗ treatment). The coefficient for
DID is the difference-in-difference estimator in which we are interested.
On the basis of the foregoing discussion, the regression implementation of DID for estimating the impact
of treatment on the outcome variable (y) is as follows:
y = time ∗
+
+
+
(4.13)
It is possible to combine fixed effects with propensity score matching (PSM) on the baseline data to
ensure that the comparison and treatment groups are similar before applying double differences to the
matched sample. Alternatively, balancing tests can be used to check whether the mean of the observables
are the same in the base period. The tests are necessary for obtaining valid results because the treatment
and comparison groups must be balanced using observed characteristics.
4.4.1
Effects of land reform policy on migration
Our econometric results are presented in Table 4.8 for the male and female subsamples, respectively.
The key explanatory variables are DID, schooling and the spatially lagged migration (W*migration).
We begin with results based on pooled OLS where the results show a positive significant impact of the
land reform policy on migration while holding other factors constant.
Next, we use the fixed-effects regression model (without spatial effects), which allows us to control for
unobserved and time-invariant characteristics that may influence migration. The results again show a
positive significant effect but only for the female subsample. The third regression (SAR) is also a fixed
128
Stellenbosch University https://scholar.sun.ac.za
effects regression model but takes into account spatial dependencies between variables. Although the
positive effect is retained, the sizes of coefficient reduce after controlling for spatial effects. There is
also clear evidence of spatial spill-overs in migration between areas as shown by large and statistically
significant coefficients for the spatial lag term, namely
∗
. Our results are also robust to
the choice of spatial models because the coefficients also remain stable for the SEM and SAC models
(see the results in Table A 11 in the appendix).
Overall, three issues are worth noting from the results. Firstly, allowing for spatial effects further reduces
the size of coefficients for our variable of interest (DID). Secondly, higher migration in one area is
associated with higher migration in neighbouring areas, as indicated by statistically significant
coefficients for
and
∗
(shown in the appendix). Thirdly, there seems to be a
gender dimension to our results. We observe that compared to males, the coefficients for females are
not only larger but statistically significant even after controlling for spatial effects.
Table 4.8: Effects of land reform policy on migration
Time
Treated
DID
OLS
-0.180***
(0.010)
-0.039***
(0.014)
0.035*
(0.020)
-0.004
(0.004)
0.056
(0.053)
-0.001
(0.001)
0.229**
(0.095)
-0.036
(0.039)
0.039***
(0.005)
Male
Non spatial
-0.188***
(0.028)
SAR
-0.089***
(0.029)
0.029
(0.020)
-0.004
(0.014)
0.106
(0.083)
-0.002
(0.001)
0.427**
(0.177)
-0.08
(0.057)
0.034**
(0.014)
OLS
-0.153***
(0.009)
-0.036***
(0.012)
0.036**
(0.017)
0.000
(0.003)
-0.045
(0.036)
0.001
(0.001)
0.083
(0.061)
-0.112***
(0.022)
0.020***
(0.005)
Female
Non spatial
-0.076**
(0.031)
0.019
0.038*
(0.014)
(0.020)
Schooling
-0.009
-0.022
(0.009)
(0.014)
Age
0.097*
-0.123
(0.056)
(0.078)
Age squared
-0.002**
0.002
(0.001)
(0.001)
Married
0.389***
0.174
(0.120)
(0.152)
Employed
-0.081**
-0.116***
(0.038)
(0.035)
Assets
0.034***
-0.002
(0.010)
(0.015)
W*migration
0.726***
(0.160)
Observations
356
356
356
356
356
Spatial units
178
178
178
178
178
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; standard errors in parenthesis
129
SAR
0.001
(0.025)
0.027**
(0.013)
-0.023**
(0.009)
-0.130**
(0.052)
0.002**
(0.001)
0.108
(0.102)
-0.114***
(0.023)
-0.005
(0.010)
0.783***
(0.131)
356
178
Stellenbosch University https://scholar.sun.ac.za
We further conducted balancing tests, which is a t-test of the difference in the means of the covariates
between the control and treated groups in the base period. This test was performed Stata’s user defined
command, namely
. The balancing tests presented in Table 4.9 show that there are no differences
in observed mean outcomes between the treatment and control samples. This further confirms that our
results are robust. Moreover, the World Bank (2012) has rated the land reform policy as satisfactory and
exceeding expectation. According to their findings, 15,142 households were successfully relocated
against a target of 15,000. Therefore, our finding that the people successfully migrated is substantiated.
Table 4.9: Balancing tests for base period
Description Weighted Variable(s)
Mean Control Mean Treated
Male
Migration
0.150
0.106
Schooling
2.565
2.472
Age
20.736
20.622
Married
0.333
0.326
Employed
0.646
0.636
Assets
-0.410
-0.500
Female
Migration
0.148
0.093
Schooling
1.275
1.233
Age
21.681
21.758
Married
0.364
0.361
Employed
0.623
0.627
Assets
-0.463
-0.595
Note: *, **, *** denote significance at 10%, 5% and 1% levels
4.4.2
Diff.
t
P-value
-0.044
-0.093
-0.115
-0.007
-0.01
-0.09
3.12 0.0021***
0.74
0.4600
0.72
0.4720
1.48
0.1400
0.2410
1.18
0.56
0.5770
-0.054
-0.042
0.077
-0.003
0.004
-0.132
3.69 0.0003***
0.4
0.6888
0.7268
0.35
0.82
0.4132
0.22
0.8271
0.9
0.3707
Effects of land reform policy on employment
Regression results based on agricultural and government employment are presented in Table 4.10 and
Table 4.11 respectively for both the male and female subsamples. It is important to note that our
treatment dummy in this section is for the areas where the people were resettled. Effectively, our results
should be understood in terms of the regions where the land reform was implemented.
DID, schooling and the spatially lagged employment (W*employed) are the three explanatory variables
of focus. The results show that the land reform had positive significant effects on agricultural
employment, more especially for females where the coefficients for DID (non-spatial fixed effects and
SAR) are not only significant but also generally larger than for males. Similar findings are established
for the SEM and SAC as reported in Table A 12 in the appendix. The programme, therefore, might have
improved the opportunities of women in agricultural employment.
130
Stellenbosch University https://scholar.sun.ac.za
The gender dimension of our results has implications for policy and poverty reduction because the
participation of females in employment contributes to economic growth. For example, in a recent study,
the Food and Agriculture Organisation (2011) found that by closing the resource gap between men and
women in developing countries, agricultural yields for females could potentially improve by between
20% and 30%. As a result, total agricultural output of developing countries would grow by between
2.5% and 4.0%. In turn, the levels of malnutrition would decline by between 12% and 17%, which would
globally imply that 100 to 150 million people would be lifted out of hunger.
Table 4.10: Effects of land reform policy on agricultural employment
Time
Treated
DID
OLS
-0.289***
(0.035)
0.080***
(0.027)
0.036
(0.037)
-0.064***
(0.010)
2.989***
(0.549)
-0.048***
(0.009)
0.109
(0.282)
0.677***
(0.188)
Male
Non spatial
-0.390***
(0.045)
SAR
-0.306***
(0.106)
0.025
(0.029)
0.077**
(0.033)
0.225
(0.777)
-0.004
(0.012)
-0.071
(0.425)
-0.262
(0.202)
OLS
-0.316***
(0.028)
0.032
(0.021)
0.042
(0.029)
-0.042***
(0.006)
1.840***
(0.237)
-0.030***
(0.004)
0.217
(0.146)
0.199*
(0.118)
Female
Non spatial
-0.443***
(0.056)
SAR
-0.252***
(0.087)
0.025
0.048*
0.045**
(0.020)
(0.028)
(0.019)
Schooling
0.074***
0.060
0.057**
(0.023)
(0.040)
(0.027)
Age
0.237
0.986*
1.020***
(0.538)
(0.507)
(0.348)
Age squared
-0.004
-0.016*
-0.016***
(0.009)
(0.008)
(0.006)
Married
-0.103
-0.100
-0.103
(0.297)
(0.378)
(0.260)
Dependency
-0.257*
0.105
0.111
(0.140)
(0.155)
(0.106)
W*employed
0.268
0.569**
(0.324)
(0.233)
Observations
356
356
356
356
356
356
Spatial units
178
178
178
178
178
178
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; standard errors in parenthesis
The results further show that government employment increased for both males and females but the
magnitude was larger for males (see Table 4.11 and Table A13). The analysis was repeated for other
types of employment such as self-employment and private employment (see results in Table A 14 in the
appendix) and we found a negative coefficient for DID, which is significant for both types of
employment for males and also in the case of private employment for women. The opposite sign of these
coefficients suggests some trade-off between wage sectors, particularly when contrasted with
agricultural and government employment where the coefficient is positive for both males and females.
131
Stellenbosch University https://scholar.sun.ac.za
We find positive and significant effects of schooling for all models, except for OLS where the coefficient
is negative in agricultural employment. More generally, it is interesting to note that the coefficient for
schooling reduces in size after taking into account spatial dependence. The reduction demonstrates the
importance of taking into account spatial effects. The statistically significant spatially lagged term W ∗
employed also confirms evidence of spatial spill-overs which implies that the employment status in
each small geographical areas affects the likelihood of employment in other neighbouring regions.
Table 4.11: Effects of land reform policy on government employment
Time
Treated
DID
OLS
0.042***
(0.007)
-0.011**
(0.005)
0.029***
(0.008)
0.015***
(0.002)
-0.335***
(0.113)
0.005***
(0.002)
-0.004
(0.058)
0.039
(0.039)
Male
Non spatial
0.063***
(0.015)
SAR
-0.010
(0.011)
0.027***
(0.009)
-0.005
(0.011)
-0.415
(0.252)
0.007*
(0.004)
-0.008
(0.138)
0.061
(0.065)
OLS
0.017***
(0.006)
-0.011**
(0.004)
0.015**
(0.006)
0.008***
(0.001)
-0.116**
(0.049)
0.002**
(0.001)
-0.019
(0.030)
0.006
(0.024)
Female
Non spatial
-0.006
(0.014)
SAR
-0.032**
(0.013)
0.022***
0.013*
0.012**
(0.006)
(0.007)
(0.005)
0.027***
0.029***
Schooling
0.004
(0.007)
(0.010)
(0.007)
Age
-0.344**
0.455***
0.453***
(0.162)
(0.130)
(0.089)
Age squared
0.006**
-0.007***
-0.007***
(0.003)
(0.002)
(0.001)
Married
-0.058
0.161*
0.134**
(0.089)
(0.097)
(0.067)
Dependency
0.04
0.044
0.041
(0.042)
(0.040)
(0.027)
W*employed
0.913***
0.648***
(0.060)
(0.195)
Observations
356
356
356
356
356
356
Spatial units
178
178
178
178
178
178
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; standard errors in parenthesis
The World Bank (2012) also cites some positive outcomes of the land reform policy which we briefly
discuss. The positive effects include an expansion in the average land holding size from less than 0.5
hectares before to approximately 2.2 hectares after the project. In addition, households that participated
reported an improvement in maize production (maize stocks after relocation lasted 10.7 months
compared to about 3.6 months before); incomes of relocated households grew by 6 times when compared
to the control group; maize and tobacco yields were respectively 4 and 2.6 times higher compared to the
base period; yields also reached an average level of 50 to 60% higher as compared to control groups in
the surrounding areas.
132
Stellenbosch University https://scholar.sun.ac.za
Due to the positive results and favourable experience, the government and the World Bank are willing
to scale up the programme to cover the whole country, where at least 100,000 households would be
resettled (World Bank, 2012). However, while the positive benefits of the reform are widely
acknowledged, Chinsinga (2008) raises a number of issues that need to be addressed to ensure its
sustainability. The main concern surrounds access to functioning services such as health, water,
transport and markets. For example, it was reported that due to lack of markets for produce some farmers
ended up selling at very low prices. Related to this, an outbreak of cholera almost led to a collapse of
one of the settlement trusts. Specifically, about 20 out of 35 households belonging to Kalungu trust
immigrated back to Thyolo for lack of access to portable water facilities.
4.5 Conclusions
In this chapter, we created panel data at the level of traditional authorities from national censuses to
examine spatial and temporal patterns of employment and migration in Malawi. We make a distinction
between males and females because this has implications for development policy.
The results showed that over the long-term changes in the population structure have effects on labour
force participation rates. We also found clear evidence of spatial spill-overs in our phenomena of
interest. Further, allowing for spatial effects reduces the sizes of coefficients and this confirms
arguments in the literature that failure to take into account spatial dependencies tends to overstate the
results. The results show that Malawi’s community-based land reform policy launched in 2004
influenced migration patterns which were largely rural-rural. Also as a result of the land reform, there
was a trade-off in the employment sectors with respect to employment. On the one hand, agricultural
and government employment improved, more particularly for the females. On the other hand, selfemployment and private employment declined. Finally, by matching GIS codes at the level of traditional
authority areas, the work undertaken for this study makes it possible to integrate census data with other
Malawian data sets for similar spatial analysis.
In terms of policy, since most indicators are heterogeneous within the country, analysis at low level
geographical units as we have done has benefits of allowing for a detailed understanding of phenomena
and designing interventions adapted for specific areas. Spatial visualisation of data is also particularly
useful for policy makers because it is intuitive.
Our study uses aggregated data at the traditional authority level rather than individual level data. It has
been argued in the literature that aggregated data ignores the rich cross-sectional evidence of the
individual observations and fails to say anything about the industries, firms and individuals themselves.
Moreover, geographical aggregation is arbitrary with respect to the definitions of the spatial
observational units. One of the solutions suggested in the literature is to estimate models at a micro level
133
Stellenbosch University https://scholar.sun.ac.za
(Arbia, 2016). However, the main challenge with this approach is the size of the spatial weighting
matrix, considering that in censuses we are typically working with millions of observations. The current
packages available to the researcher can only handle up to 11,000 x 11,000 for Stata 14.1 SE or MP
(e.g., Belotti et al., 2013). This may be an area for further study as we explore other available options.
134
Stellenbosch University https://scholar.sun.ac.za
Chapter 5
Conclusion
5.1
Introduction
This thesis has used multiple data sources to investigate trends in non-monetary dimensions of welfare,
labour market outcomes and migration in Malawi. As noted in Chapter 1, the issues are investigated in
three separate chapters, namely Chapters 2, 3 and 4. Section 5.2 provides the summary of findings from
each of the chapters. In Section 5.3, we discuss the conclusion of the thesis and implications for policy.
Research implications are discussed in Section 5.4. Contributions made to the literature are summarised
in Section 5.5. Finally, in Section 5.6, we provide suggestions for future research.
5.2
Summary of findings
Chapter 2 was dedicated to the analysis of poverty and inequality over time using two non-monetary
measures of welfare, namely child nutritional status and household asset ownership. Data for this study
was drawn from nationally representative DHS data sets for 1992, 2000, 2004 and 2010. On the one
hand, the temporal aspect of the study provided an understanding of long-term economic well-being in
Malawi. On the other hand, the spatial aspect gave the profile of welfare across population groups such
as regions and urban-rural areas. The study showed that poverty in Malawi has declined tremendously
from as high as 80% in 1992 to about 50% in 2010. Having established that welfare has improved over
time, the study set out to investigate who benefited from these gains in welfare. This was investigated
through pro-poor growth analysis which showed that all population groups benefited but the greatest
gains accrued to the poor as opposed to the rich. This finding implies that inequality in Malawi has
declined between 1992 and 2010.
The study also conducted rankings of welfare that are robust to the choices of the poverty line and the
dimension of well-being. For example, we found that poverty and inequality are higher in rural areas
than urban areas regardless of whether we use assets or child-nutritional status. Our results show that
while welfare differs among regions, the regional differences are not as large as the difference between
rural and urban areas. For example, using the asset index, only about 7% of the population is poor in
urban areas compared to 53% in rural areas. The gap is huge when contrasted with regional poverty
rates of 32.4%, 51.3% and 44.5% for the Northern, Central and Southern regions, respectively.
Nevertheless, we conclude that poverty in Malawi is both a rural area and regional problem. Therefore,
policy interventions should focus more on reducing poverty in rural areas while not neglecting the
regional imbalances.
135
Stellenbosch University https://scholar.sun.ac.za
One of the advantages of the measures of poverty and inequality employed in Chapter 2 is that they are
decomposable. Poverty decomposition shows that urban areas and households headed by females
contribute less to poverty compared to rural areas and households headed by males, respectively. We
also decomposed total inequality into between-group (for example, rural and urban) and within-group
(for example, rural-urban asset gap) inequality. According to Haughton & Khandker (2009), withingroup inequality is typically at least 75% of total inequality in a given country. Our results confirmed
that inequality is indeed higher within groups as opposed to between groups; we find magnitudes of
between 65% and 100% depending on the measure of welfare used as well as population group pair.
Chapter 3 used the two waves of the IHS3 panel data, namely 2010 and 2013, to analyse returns to
education in the wage sector and externalities to education in household enterprises. Three main issues
were considered. Firstly, we discussed the various issues that could affect the reliability of the returns
and externalities to education. The conclusion was that a robust treatment of outliers and inconsistencies
in the data makes it possible to obtain trends in the labour market that are both consistent and reliable.
Secondly, we tested the argument that educational attainment is positively associated with poverty
reduction through the labour market. This assertion was investigated through an econometric analysis
of the role played by education in the determination of the likelihood of labour force participation and
employment. After controlling for other factors, we found that individuals with tertiary education are
more likely to participate in the labour market than those without education. However, the results
showed that those with primary and secondary are less likely to participate in the labour market than
those without education. We reason that those without formal education drop out of schooling to
immediately enter the labour market, unlike those with primary and secondary education who choose to
continue with their education and, therefore, only enter the labour market at a later stage. Moreover, the
largest proportion of people without education is engaged in informal sector jobs where earnings are
low. We follow a similar line of thinking with the multivariate analysis of employment where we find
that those with JCE and MSCE are less likely to get employed compared to those without education.
Thirdly, the study investigated if higher levels of schooling or education are associated with higher
incomes. Our findings, which are robust to sample selection, yielded large positive returns to years of
schooling in Malawi for both wage employment and self-employment in household enterprises. The
analysis was repeated with education categories instead of years of schooling. The results did not
change, implying that the definition of education does not matter. However, using education categories
gives us more information in the sense that we are able to show that the returns to education are
heterogeneous rather than homogeneous. Not only do the returns increase by the level of education but
are also higher amongst females compared to males. Through these findings, the study demonstrates the
136
Stellenbosch University https://scholar.sun.ac.za
importance of distinguishing between groups such as economic sectors and gender when analysing
labour market outcomes.
In Chapter 4, trends in employment and migration were examined using the 1987, 1998 and 2008 census
data which are the only three publicly available census data sets in the country. The study has shown
that both geography and time matter in the understanding of long-term patterns of employment and
migration. While time is important, we demonstrated that it is not everything we should be controlling
for when analysing economic outcomes that are essentially spatial such as employment and migration.
The importance of geography is proven by the fact that the magnitude of explanatory coefficients
dramatically decline once we take spatial dependencies into account in our regressions.
Our analysis was done in two stages. In the first stage, we confirmed that spatial dependencies exist in
our data using the two most commonly used tests for spatial autocorrelation, namely the Moran’s I and
Geary’s c statistics. Spatial distributions on the map of Malawi showed some heterogeneity and
clustering in the variables of interest. The detection of spatial dependence only provided the first step in
spatial data analysis since it is possible that such clustering could occur randomly. Therefore, as a second
step, we conducted some diagnostic tests in order to understand the type of clustering. This further
informed the specification of the spatial regression model. The diagnostic tests showed that it is possible
to fit both the spatial lag and error models or their variants with our data. Therefore, our regression
results are robust to the choice of model specification from a whole range of the available model options
commonly used in spatial panel data analysis.
The study also explored the effects of the Malawi Community Based Rural Land Development Project
(CBRLDP) on migration and employment. The results show that the land reform policy positively
influenced migration patterns and also resulted in increased employment opportunities for the
individuals who migrated. These results have implications for policy. By showing that it is possible for
people to migrate in the pilot phase, the programme can potentially be scaled up to cover the whole
country given the positive benefits of the programme in terms of gains in employment and incomes for
the participating households. The World Bank and Government of Malawi already have plans to expand
the programme to cover the whole of the country for similar settlements. The results also revealed an
interesting gender dimension in the sense that employment effects differ between males and females;
not only did agriculture employment improve by larger magnitudes for the female subsample when
compared to males, but was also significant, unlike the male subsample where we obtained nonsignificant results. Therefore, we expect the programme to have improved the incomes of women
involved in agricultural employment for a living. This finding is important because the Food and
Agriculture Organisation (2011) states that if resource gap in developing countries between men and
women were to be closed, the latter could potentially increase yields by between 20% and 30% and lift
137
Stellenbosch University https://scholar.sun.ac.za
many out of food poverty. Although government employment increased for both the male and female
subsamples, the magnitude was larger for males.
5.3
Conclusions
The poverty and inequality profiles derived in Chapter 2 suggest that although the welfare changes over
time have pro-poor, Malawi remains a poor country in terms of both assets and child-nutritional status.
Malawi’s poverty is, therefore, multidimensional. A robust ranking of welfare shows that poverty and
inequality levels are highest in rural areas and amongst households headed by females. Rankings for the
three regions of Malawi depend on the welfare measure used. While asset poverty is the highest in the
Central region, child-nutritional status is highest in the Southern region. With respect to inequality, while
no robust ranking is determined with respect to assets, height-for-age z-scores (HAZ) and weight-forage z-scores (WAZ), we are able to conclude that inequality is the lowest in the Northern region when
weight-for-height z-scores (WHZ) are used as a measure of welfare. Amongst the three measures of
child-nutritional status, HAZ yields the highest levels measures of poverty and inequality.
Regardless of the welfare measure used, poverty decompositions seem to suggest that poverty is the
highest in groups where the population shares are the largest. For example, rural areas which constitute
about 85% of the population, contribute to about 91.4% of the poverty using WHZ as a measure of childnutritional status. Similarly, the Central region and households headed by males contribute more to
poverty compared to other population groups.
Using multivariate analysis, we have identified factors associated with asset poverty and child
nutritional status in Malawi. With respect to child-nutritional status, the study has shown that childindividual characteristics such as age, sex and birth order play an important role in addition to household
welfare and the levels of education by both the father and mother. It was also established in Chapter 2
that household characteristics such as household size, dependency ratio, levels of education in the
household and sex of the household head are important correlates to asset poverty. Furthermore, asset
poverty is geographical and depends on the area of residence.
Chapter 3 confirms the human capital theory. First, we show that education indeed improves chances of
employment, particularly for those with tertiary education. Secondly, we show that returns to education
increase with levels of education. We also find that the returns are higher for females than males.
Furthermore, given that wage employment is a small proportion of total employment, we segregate our
results by economic sector. Our results show that modelling returns to education based on the formal
sector only is misleading as it assumes a homogeneous labour market which is not true for developing
countries such as Malawi. Finally, our results show that we should be correcting for sample selection
when analysing labour market outcomes in Malawi.
138
Stellenbosch University https://scholar.sun.ac.za
In Chapter 4, we confirmed that employment and migration are spatial economic phenomena.
Econometric analysis of these variables of interest should ideally, therefore, control for both time and
geography. We also showed that, between 1987 and 2008, there has been an increase in labour force
participation of the youth in Malawi, particularly those aged between 15 and 19 years. This might have
implications for youth unemployment in the country. The analysis of long-term occupational mobility
showed that a larger proportion of women remains trapped in agricultural employment compared to
males. The results also show that the land reform policy in Malawi caused people to migrate to other
areas in search of agricultural land. In terms of the effects on employment, agricultural employment
increased for the female subsample while government employment particularly increased for the males.
5.4
Implications of the research
The main aim of this research was to contribute to the empirical understanding of poverty, inequality
and the labour market in Malawi. Using multiple sources of data which stretch over time, the study
shows interesting dynamics of the aforesaid economic phenomena in a country which has been largely
static in terms of economic growth and urbanisation. One of the main conclusions one can draw from
the study is that Malawi’s poverty profile is a ‘bad picture’ given that almost 50% of the population was
still poor in 2010, but a ‘good movie’ in that the incidence of poverty had fallen from as high as 80% in
1992. This holds for both monetary and non-monetary indicators of welfare. A number of implications
can be drawn from this research and these may have broader policy applications. Following the structure
of the thesis, the implications are discussed separately for each of the three main chapters.
Chapter 2 provided research evidence on poverty and inequality in Malawi using two non-monetary
dimensions of welfare, namely an asset index and child-nutritional status. The first implication arising
from this chapter emanates from the finding that Malawi has very high levels of poverty regardless of
whether assets or child-nutritional status are used as the measure of welfare. Other research based on
consumption (e.g. Mussa, 2013) derives similar implications with respect to this finding. As shown in
Chapter 2, despite improvements over time, the data shows very low levels of asset ownership in
Malawi. There is also poor access to sanitation, and poor quality of dwellings as reflected in the floor
material, as well as poor water services. These have either remained the same or declined over time.
Therefore, despite gains in poverty reduction, Malawi remains a very poor country; addressing this
problem remains extremely important.
A second implication relates to the finding that there is not much regional variation in terms of both
poverty and inequality, as shown by the narrow bands in the poverty and inequality mapping shown in
Chapter 2. This suggests that the population subgroups are largely undifferentiated. This homogeneous
feature of welfare seems to be a feature that is quite unique to Malawi, as few other developing countries
show such small differences in poverty measures across geographic space. Similarly, we find that there
139
Stellenbosch University https://scholar.sun.ac.za
are no large differences between male and female-headed households, a finding not so common in most
developing countries, including most of Malawi’s neighbours. Perhaps these findings relate to the
largely undiversified nature of the livelihoods of the Malawian population.
The research further established that rural areas dominate urban areas in terms of poverty. This finding
is not surprising, considering that 85% of the population in Malawi lives in rural areas and that urban
areas usually offer at least some improved formal sector wage earning opportunities. Nevertheless, it is
important to note that despite the rural dominance, poverty is also quite high in urban areas. For example,
incidences of stunting in urban and rural areas stand at 40.8% and 48.3%, respectively. Thus, while
government policy should focus on rural areas, it is important not to neglect urban areas.
In stark contrast with consumption-based inequality estimates normally used in official reports, this
study interestingly shows that non-money metric inequality is higher in rural areas compared to urban
areas. This is a new finding and responds to the call by some, such as Grosse et al. (2008), who have
highlighted the need to extend pro-poor growth analysis to non-monetary dimensions. Clearly, this has
implications for how government considers its policy on inequality. It may be necessary to rethink the
policy approach to reducing inequality in Malawi. Specifically, policy makers need to explicitly take
into account non-income dimensions of welfare when formulating public policy.
Chapter 3 used panel data to estimate externalities and returns to education in Malawi. Although the
chapter shows that returns to education are positive in both the formal and informal sectors, they are
lower in the latter where the majority of the population is employed. Therefore, unless workers move
from the informal sector to the formal sector, the benefits from education remain low for the majority
of the workforce. A second feature evident from the analysis in Chapter 3 relates to the structure of
employment in Malawi. Casual employment and self-employment activities (household enterprises)
continue to make up the largest share of total employment both in and outside of agriculture. They jointly
constitute at least 75% of total employment.
The chapter also shows that Malawi has very high labour force participation rates, ranging between 71%
and 91%, depending on whether the broad or narrow definition is used. These traditionally high rates of
labour participation can be linked to the existence of a large agricultural sector which in turn makes it
difficult to clearly identify actual labour force participation rates, as subsistence activities in the
agricultural sector are categorised as labour force participation.
140
Stellenbosch University https://scholar.sun.ac.za
The chapter also shows that the labour force in Malawi remains poorly educated; only 17% of
individuals in the working age population has more than primary education. The low levels of education
imply low productivity levels that partly account for the large earning differentials between the formal
and informal sector. Consequently, there is a need to continue investing in education which has been
found to yield positive returns for those employed.
The fact that returns to education are higher for females than males with similar skills (more particularly
at high levels of education) shows that the continued expansion of education for girls should translate
into gains in the fight against poverty and the economic empowerment of women. It is further shown in
the study that the returns to education in non-farm enterprises are higher when maximum household
education is used (signifying externalities) as opposed to the level of education of the owner of the
enterprise. This suggests that provision of free primary education and subsidies in higher education can
be justified on these grounds. In this case, government’s investment in education is not only the right
thing to do but also have much wider benefits for society in terms of poverty reduction and
improvements in food security. Furthermore, the presence of education externalities raises issues for
further study. For example, motivated by findings emergent from this study, one may want to further
investigate how the allocation of education to different household members of different ages or gender
involved in various activities would affect household income. Simply put, whose education matters in
the determination of household income?
The use of panel data has enabled an examination of changes in occupational status as well as
movements into and out of the labour market. This is not only important for monitoring trends but also
for identifying the sources of changes over time. For example, from the data, it is possible to identify
individuals who are employed or unemployed in both waves. This is not possible with cross-sectional
data. Consequently, panel data not only provided transparent ways of dealing with data issues but also
allow a better understanding of the economic phenomena of interest, including how earnings have
changed between surveys. The data shows that compared to those only employed in either wave, the
initial earnings amongst individuals employed in both years were not only higher (almost double as
much) but also increased by a greater magnitude. Specifically, the earnings of those employed in both
waves increased by 46% between 2010 and 2013. While most of the individuals employed in both waves
were educated, the majority of those unemployed had no education. This seems to suggest that education
not only improves chances of finding employment but also that higher education is associated with
higher earnings. Furthermore, with panel data, it was possible to control for individual heterogeneity
since the individuals served as their own controls.
Chapter 4 looked at migration and employment in Malawi using spatial data. The government has
generally regarded the movement of people within Malawi in the light of rural-urban migration which
141
Stellenbosch University https://scholar.sun.ac.za
is seen as a natural and inevitable process. Policies have, therefore, tended to focus on developing rural
centres to prevent massive movements of people into urban areas, particularly the four main urban
centres, namely Mzuzu, Lilongwe, Blantyre and Zomba. However, given the low levels of urbanisation,
rural-urban migration is not as massive as perceived. On the contrary, the observed patterns of migration
in Malawi are largely rural-rural. Specifically, while most people stay in the home regions there are
significant movements across districts. This points to the continued dominance of rural livelihoods, even
for those in search of better opportunities.
Furthermore, the study has shown that the land reform policy enabled people to move from the rural
areas of sending districts into the rural areas of receiving districts. This policy has had large implications
for development because of its effects on employment. Consequently, government policy needs to go
beyond simply providing access to land because this may potentially create as many problems as it seeks
to solve. On a more substantive level, deliberate policies need to be put in place to support agricultural
production in addition to the provision of markets and other necessary amenities.
In Chapter 2, it was shown that there are no major differences between regions in terms of economic
welfare. However, there exists some heterogeneity in terms of economic indicators when small-level
geographical data is used as shown in the spatial maps in Chapter 4. Specifically, while there seems to
be broad uniformity across the main regions, a descriptive spatial analysis of employment, schooling,
fertility and assets shows some heterogeneity at the smallest geographical units, namely traditional
authorities. One implication of this finding is that policy should not be based on the average picture but
instead needs to be tailored to the small geographical areas. In addition, visualisation of spatial is a
powerful tool and can be used to aid policies that are specifically adapted for small geographical units.
Some studies (e.g. Gaddis & Klasen, 2014) link fertility rates to economic development. Chapter 2 of
this thesis has shown that fertility rates in Malawi have dropped between 1987 and 2008. Although this
is the broad trend for the population, fertility rates have increased by almost 2% amongst women aged
between 20 and 24 years, contrary to what is desired; this may call for policy action. High fertility rates
are known to have effects on education; education expenditures per child tend to be lower in large
families. Furthermore, high fertility rates may also negatively affect the supply of labour of the parents,
especially when the children are still young. This may, in turn, have a negative effect on the ability of
families to save for the children.
The analysis of the long-term patterns of labour force and population in Malawi shows that the
percentage of the youth participating in the labour force has been on the increase, particularly for
females. These observed changes in the age structure of the population and of labour force participation
have significant implications for the supply of labour and therefore also the need for economic growth.
142
Stellenbosch University https://scholar.sun.ac.za
There is a need for the corresponding improvement in education and skills development. The youth,
particularly those aged between 20 and 24 years, constitutes the largest percentage of the unemployed.
Youth unemployment has implications for development. For example, as education continues to expand,
many of those educated may fail to find employment due to increased competition for scarce formal
sector jobs. Even those who eventually enter into employment may actually end up in informal
employment and subsistence agriculture where the earnings are low.
Another important implication of the research is that it provides a platform for further spatial analysis
in Malawi. Specifically, the study has demonstrated the feasibility of creating GIS codes that are
consistent over time. This is an important step towards linking census data with other national data sets
for similar analysis at the lowest geographical unit.
This thesis has thus shown that Malawi, one of the poorest countries in the world, has many of the
features of poor countries, but also some features that make it quite unique. These include the relatively
undifferentiated economic activity in rural areas, which lies at the heart of the limited economic and
spatial differentiation across the country as a whole. This thesis has also shown that there has been
considerable pro-poor growth in Malawi, something that is to be commended and holds promises that
future growth too may benefit Malawi’s many poor people. As the country develops, however, it is
likely that differentiation would increase and that heterogeneity would become the norm in the economic
field as in the labour market. The data used in this study offer a useful first attempt at evaluating the
features and consequences of economic progress. Further work on new datasets that become available
over time would be beneficial for following the trajectory of growth and development.
5.5
Summary of contributions
Previous studies on poverty and inequality in Malawi have been static in nature, due to data limited
availability. This study was able to conduct more inter-temporal analysis that was made possible by
increased data availability. Chapter 2 has contributed to the literature by examining trends in poverty
through pro-poor growth analysis which allows us to understand the distribution of the gains from
reductions in poverty and inequality over time. More generally, this is the first attempt to apply an asset
index to the measurement of poverty in Malawi. While it is generally understood that poverty is multidimensional, there has been a dearth of literature dedicated to the empirical understanding of asset
poverty and how the results compare with what is already known with respect to other dimensions such
as consumption, education and nutritional status.
Until the release of the IHS3 panel data (IHSP), there has been a lack of panel data in Malawi. Therefore,
previous evidence on education and earnings has been based on cross-sectional data. In Chapter 3, we
take advantage of the recently made available panel survey results set to explore the interesting dynamic
143
Stellenbosch University https://scholar.sun.ac.za
aspects of the observations as well as control for unobservable individual heterogeneity. Furthermore,
we conduct robust treatment of outliers and bring to light some of the inconsistencies in the data. These
are some of the issues that new users of the data will have to look out for in order to ensure quality data
before further analysis. This more generally applies to most datasets in less developed countries with a
large informal sector.
The main contribution in Chapter 4 is the application of spatial econometric methods to the
understanding of employment and migration in Malawi. For example, by comparing results based on
spatial effects to those based on non-spatial regression analysis, we demonstrate that studies that do not
take into account spatial effects tend to overestimate other coefficients. Potentially, this study provides
an exploratory platform for further spatial analysis in Malawi. In order to achieve this, we have matched
GIS codes that are consistent over time with the view to link census data with other data sets for similar
spatial analysis. This kind of data integration is an area for further study, as noted in the ensuing section.
5.6
Future research
Chapter 2 dealt with the dynamic aspects of poverty and inequality in Malawi by examining trends over
time. However, since we are not tracking the same households, our results only show the overall
population changes in welfare that have occurred over time. Another important aspect of the dynamic
nature of poverty considered in the literature is the identification of the proportions of households or
individuals that are either chronically or transiently poor. This kind of analysis, which can only be
achieved with panel data, has not been done in this study. Future research can take advantage of the
IHSP data which has been recently released by the National Statistical Office.
Furthermore, in Chapter 3, we identified measurement error in the independent variables as one of the
possible sources of bias when estimating returns and externalities to education. The most commonly
suggested method of dealing with this kind of measurement error in the literature is differencing (short
and longer differencing). Unfortunately, this approach only works for panels with at least three waves
of data panel, which is not available in Malawi at the moment. However, as more waves are added to
the IHSP in the future, there would be the possibility of accounting for measurement error in the
independent variables.
Another area for future research relates to the treatment of individuals with zero earnings. Our approach
has been to exclude from the analysis observations with zero incomes. One of the criticisms of this
approach is that it may affect the reliability of the estimates (Yu, 2012), especially if the observations
are many. Therefore, researchers sometimes consider an alternative of adding a very small number to
earnings before taking the log, which would allow considering zero incomes. This option is not
considered in this study because zero incomes constitute about 10% percentage in our data which we
144
Stellenbosch University https://scholar.sun.ac.za
believe is large and may, therefore, eventually end up distorting the distribution of the data. As a further
study, one may set out to investigate the alternative treatment of zero earnings not only for the panel
data set but also for other data sets in the country including consumption figures.
The spatiotemporal analysis carried out in Chapter 4 can be generally applied to economic phenomena
by matching GIS census data with other data in Malawi. This constitutes an area for future research.
With the development of geosystems, it can be expected that geo-coded data will become more readily
available in Malawi. This will make spatial analysis possible on the new datasets, but the richness of
such data could perhaps be further enhanced by adding information from the census for smaller
geographical areas. Thus, this study acts as a platform for further spatial analysis. Another extension of
this study would be to conduct spatial analysis using individual level data instead of aggregated data as
is the case in the present study. However, this approach requires working with very large matrices
involving millions of observations, as is typically the case with census data. The current Stata packages
only handle smaller matrix sizes as explained in Section 4.5. Consequently, use of individual data for
spatial analysis is dependent on the identification of methodologies and tools that overcome this
limitation.
145
Stellenbosch University https://scholar.sun.ac.za
List of references
Alkire, S. & Santos, M.E. (2010). Acute Multidimensional Poverty: A New Index for Developing
Countries (OPHI Working Paper No. 38). Retrieved from https://www.ophi.org.uk/wpcontent/uploads/ophi-wp38.pdf.
Anselin, L. (1992b). Spatial Data Analysis with GIS: An Introduction to Application in the Social
Sciences. National Center for Geographic Information and Analysis, University of California,
Santa Barbara.
Anselin, L. (2003). Spatial Econometrics, in Baltagi, B.H. (Ed.), A Companion to Theoretical
Econometrics. Malden, Massachusetts, USA: Blackwell Publishing Ltd.
doi: 10.1002/9780470996249.ch15.
Appleton, S., Bigsten, A. & Manda, D.K. (1999). Educational Expansion and Economic Decline:
Returns to Education in Kenya, 1978-95 (Working Paper WPS/99-6). Centre for the Study of
African Economies, University of Oxford.
Araar, A. & Duclos, J.Y. (2013). DASP: Distributive Analysis Stata Package [Computer Software].
Laval University, World Bank, PEP and CIRPÉE.
Arbia, G. (2016, June). Spatial econometrics: a broad view. Paper presented at the X World
Conference Spatial Econometrics Association (SEA), Rome, Italy.
Baltagi, B. H. & Li, Q. (1990). A Lagrange multiplier test for the error components model with
incomplete panels. Econometric Reviews 9 (1), 103–7.
Baltagi, B.H. (2013). Panel Data Methods. Handbook of Applied Economic Statistics. New York.
Basu, K. & Foster, J. E. (1998). On measuring literacy, Economic Journal, 108(451), 1733-49.
Becker, G. S. (1962). Investment in human capital: A theoretical analysis. The journal of political
economy, 9-49.
Belotti, F., Hughes,G. & Mortari, A.P.(2014). “XSMLE: Stata module for spatial panel data models
estimation," Statistical Software Components S457610, Boston College Department of
Economics.
Bennell, P. (1996). Rates of Return on Education: Does the Conventional Pattern Prevail in SubSaharan Africa. World Development, 24 (1), 183-99.
Blundell, R. & Preston, I. (1998). Consumption inequality and income uncertainty. Quarterly Journal
of Economics. 113(2), 604-40.
146
Stellenbosch University https://scholar.sun.ac.za
Bokosi, F.K. (2006). Household poverty dynamics in Malawi: A bivariate probit analysis. Journal of
applied sciences, 7 (2), 258-62.
Booysen, F., Van der Berg, S., Burger, R., Von Maltiz, M., & Du Rand, G. (2008). Using an Asset
Index to Assess Trends in Poverty in Seven Sub-Saharan African Countries. World
Development, 36 (6), 1113-1130.
Bourguignon, F. & Chakravarty, S.R. (2003). The measurement of multidimensional poverty.
Journal of Economic Inequality, 1 (1), 25-49.
Burger, R. & Yu, D. (2007). Wage trends in post -Apartheid South Africa: Constructing an earnings
series from household survey data (Working Paper 07/117). Development Policy Research
Unit, University of Cape Town.
Card, D. (1999). The causal effect of education on earnings. Handbook of labor economics, 3, 18011863.
Casale, D. & Desmond, C. (2015). Recovery from stunting and cognitive outcomes in young
children: Evidence from the South African Birth to Twenty Cohort Study. Journal of
developmental origins of health and disease, available on CJO2015.
doi:10.1017/S2040174415007175.
Chen, M. A. (2012). The informal economy: Definitions, theories and policies. Women in informal
economy globalizing and organizing: WIEGO Working Paper, 1.
Chirwa, E. & Matita, M. (2009). The Rate of Return on Education in Malawi (Department of
Economics Working Paper No. 2009/01). University of Malawi, Chancellor College.
Chirwa, E.W. & Ngalawa, H.P.E. (2008). Determinants of Child Nutrition in Malawi. South African
Journal of Economics, 76(4), 628-40.
Chirwa, E.W. & Zgovu, E.K. (2002). Does the Return to Schooling Depend on the Type of
Employment? Evidence from the Rural Labour Market in Malawi.
Chinsinga, B. (2008). Exploring the politics of land reforms in Malawi: A case study of the
Community Based Rural Land Development Programme (CBRLDP). University of
Manchester: Manchester, UK.
Clark, T.S. & Linzer, D. A. (2015). Should I Use Fixed or Random Effects? Political Science
Research and Methods, 3(2), 399–408.
Crow, K. (2015). SHP2DTA: Stata module to converts shape boundary files to Stata datasets.
Cooke, T. J. (2013). Internal migration in decline. The Professional Geographer, 65(4), 664-675.
Da Maia, C. (2012). Understanding Poverty and Inequality in Mozambique. The Role of
Education and Labour Market Status (Doctoral thesis, Stellenbosch University, South Africa).
Retrieved from http://scholar.sun.ac.za/handle/10019.1/71857.
Dancer, D., Rammohan, A. & Smith, M. D. (2008). Infant mortality and child nutrition in
Bangladesh. Health Economics, 17(9), 1015–35. doi:10.1002/hec.1379.
147
Stellenbosch University https://scholar.sun.ac.za
Davidson, R., & Duclos, J. Y. (2000). Statistical inference for stochastic dominance and for the
measurement of poverty and inequality. Econometrica, 68(6), 1435-1464.
Deaton, A. (1997). The analysis of household surveys: A microeconomic approach to
development policy. Baltimore: The John Hopkins University Press.
Duclos, J. & Araar, A. (2006). Poverty and equity: Measurement, policy and estimation with
DAD (1st ed.). Ottawa: Springer.
Duclos, J.Y. & Verdier-Chouchane, A. (2010). Analyzing Pro-Poor Growth in Southern
Africa: Lessons from Mauritius and South Africa (Working Papers Series No. 115). African
Development Bank, Tunis, Tunisia.
Ewoudou, J. & Vencatachellum, D. (2006). An Empirical Analysis of the Rates of Returns to
Education in Cameroon. Paper presented at the Centre for the Study of African Economies
Conference, University of Oxford, Oxford.
Food and Agriculture Organisation. (2011). The state of food and agriculture: Women in
agriculture- closing the gender gap for development. Rome, Italy: FAO. Retrieved from
http://www.fao.org/docrep/013/i2050e/i2050e.pdf.
Foster, J., Greer, J., & Thorbecke, E. (1984). A Class of Decomposable Poverty Measures.
Econometrica, 52(3), 761-66. doi:1. Retrieved from http://www.jstor.org/stable/1913475 doi:1
Gaddis, I. & Klasen, S. (2014). "Economic development, structural change, and women’s
labour force participation:," Journal of Population Economics, Springer, 27(3), 639-81.
Geary, R. (1954). The contiguity ratio and statistical mapping. The Incorporated Statistician,
5 (3), 115–45.
Government of Malawi. (2007). Malawi Growth and Development Strategy I: From Poverty
to Prosperity (2006-2011). Lilongwe, Ministry of Economic Planning and Development.
Retrieved from https://www.imf.org/external/pubs/ft/scr/2007/cr0755.pdf.
Government of Malawi. (2012). Malawi Growth and Development Strategy II (2011-2016).
Lilongwe, Ministry of Finance and Development Planning. Retrieved from
https://www.imf.org/external/pubs/ft/scr/2012/cr12222.pdf.
Gondwe, A.S. (2011). Spatial Comparisons of Multidimensional Poverty in Malawi
(Unpublished Master’s thesis). Chancellor College, University of Malawi, Malawi.
Greenacre, M. (2007). Correspondence analysis in practice. CRC press.
Greenwood, M. J. (2005). Modeling migration. Encyclopedia of social measurement, 2, 725-734.
Grosse, M., Harttgen, K., & Klasen, S. (2008). Measuring pro-poor growth in non-income
dimensions. World Development, 36(6), 1021-1047.
148
Stellenbosch University https://scholar.sun.ac.za
Harris, J. R., & Todaro, M. P. (1970). Migration, unemployment and development: a two-sector
analysis. The American economic review, 126-142.
Haughton, J. & Khandker, S.R. (2009). Handbook on poverty and inequality. Washington:
The World Bank.
Heckman, J. (1979). Sample Selection as a Specification Error. Econometrica, 47 (1), 15361.
Human Development Report. (2015. Work For Human Development: Retrieved from
http://hdr.undp.org/sites/default/files/2015_human_development_report.pdf.
Hussmanns, R. (2004, February). Statistical definition of informal employment: Guidelines endorsed
by the Seventeenth International Conference of Labour Statisticians (2003). In 7th Meeting of
the Expert Group on Informal Sector Statistics (Delhi Group) (pp. 2-4).
International Labour Organization. (2011). ILO estimates and projections of the
economically active population: 1990–2020 (6th ed.). Retrieved from
http://laborsta.ilo.org/applv8/data/EAPEP/eapep_E.html.
Jeanty, P.W. (2010). spmlreg: Stata module to estimate the Spatial lag, the Spatial error, the Spatial
Durbin, and the General Spatial models.
Jolliffe, D. (2002). Whose education matters in the determination of household income? Evidence
from a developing country. Economic Development and Cultural Change, 50(2), 287-312.
Kahyarara, G. & Teal, F. (2008).The Returns to Vocational Training and Academic Education:
Evidence from Tanzania. World Development, Elsevier, 36(11), 2223-42.
Kakwani, N. (1980). On a class of poverty measures. Econometrica: Journal of the Econometric
Society, 437-446.
Kakwani, N., & Pernia, E. M. (2000). What is pro-poor growth? Asian development review, 18(1), 116.
Kerr, A. & Teal, F.(2015). The Determinants of Earnings Inequalities: Panel data evidence from
KwaZulu-Natal, South Africa. Journal of African Economies, 24(4), 530-58.
Khandker, S.R., Koolwal, G.B. & Samad, H.A. (2010) Handbook on impact evaluation: quantitative
methods and practices. Washington, DC: The World Bank.
Kingdon, G. & Knight, J. (2004). Unemployment in South Africa: The Nature of the Beast. World
Development, 32 (3), 391-408.
Lassibille, G. & Tan, J. (2005).The Returns to Education in Rwanda. Journal of African Economies,
14 (1), 92-116.
Linnemayr, S., Alderman, H. & Ka, A. (2008). Determinants of malnutrition in Senegal: Individual,
household, community variables, and their interaction. Economics & Human Biology, 6 (2),
252-63.
149
Stellenbosch University https://scholar.sun.ac.za
Matita, M. & Chirwa, E. (2009). The Impact of Education on Self-Employment, Farm Activities and
Household Incomes in Malawi. (Department of Economics Working Paper No. 2009/02).
University of Malawi, Chancellor College.
McKay, A. (2000). Should the survey measure total household income? In Grosh, M. and Glewwe,
P. (Eds.), Designing household survey questionnaire for developing countries: Lessons from
15 years of the living standards measurement study (pp. 83 – 104). Washington: The World
Bank.
Mincer, J. (1974). Schooling, Experience and Earnings. New York: National Bureau of Economic
Research. Retrieved from http://papers.nber.org/books/minc74-1.
Minnesota Population Center. (2015). Integrated Public Use Microdata Series, International:
Version 6.4 [Machine-readable database]. Minneapolis: University of Minnesota. Retrieved
from https://international.ipums.org/international-action/variables/group.
Moran, P. (1948). The interpretation of statistical maps. Journal of the Royal Statistical Society,
Series B (Methodological), 10 (2), 243–51.
Murkherjee, S. & Benson, T. (2003). The determinants of poverty in Malawi, 1998. World
Development, 31(2), 339-358.
Mussa, R. (2013). Spatial Comparisons of Poverty and Inequality in Living Standards in Malawi.
South African Journal of Economics, 81(2), 192-210.
Mussa, R. (2014). A matching decomposition of the rural–urban difference in malnutrition in
Malawi. Health Economics Review, Springer, 4(1), 1-10.
Mussa, R. (2014). Externalities of Education on Efficiency and Production Uncertainty of Maize in
Rural Malawi (MPRA Paper 54628). University Library of Munich, Germany.
National Statistical Office. (2008). Population and Housing Census Report. Zomba, Malawi.
Retrieved from
http://www.nsomalawi.mw/images/stories/data_on_line/demography/census_2008/Main%2
0Report/ThematicReports/Migration%20Report.pdf.
National Statistical Office. (2012). Integrated Household Survey 2010-2011: Household SocioEconomic Report. Zomba, Malawi. Retrieved from
http://www.nsomalawi.mw/images/stories/data_on_line/economics/ihs/IHS3/IHS3_Report.
pdf.
150
Stellenbosch University https://scholar.sun.ac.za
National Statistical Office. (2012). Malawi Labour Force Survey Report 2013. Zomba, Malawi.
Retrieved from http://www.nsomalawi.mw/component/content/article/8/209-malawilabour-force-survey-2013.html.
National Statistical Office. (2014). Integrated Household panel Survey 2010-13: Household SocioEconomic Characteristics Report. Zomba, Malawi. Retrieved from
http://www.nsomalawi.mw/images/stories/data_on_line/economics/ihs/IHPS%202013/IHP
S%20Report.pdf.
National Statistical Office. (2015). Welfare Monitoring Survey 2014. Retried from
http://www.nsomalawi.mw/images/stories/data_on_line/agriculture/wms_2014/Welfare%20
Monitoring%20Survey%202014.pdf.
National Statistical Office & ICF Macro. (2011). Malawi Demographic and Health Survey 2010.
Zomba, Malawi, and Calverton, Maryland, USA. Retrieved from
http://www.dhsprogram.com/pubs/pdf/FR247/FR247.pdf.
Organisation for Economic Co-operation and Development. (2008). 2008 Survey on Monitoring the
Paris Declaration: Making Aid More Effective by 2010. OECD Publishing, Paris. doi:
10.1787/9789264050839-en.
Pisati, M. (2001). Tools for Spatial Data Analysis. Stata Technical Bulletin. March 2001. STB60:21-36. Retrieved from http://www.stata.com/products/stb/journals/stb60.pdf.
Pisati, M. (2012). Spatial Data Analysis in Stata: An Overview. Department of Sociology and Social
Research. University of Milano-Bicocca (Italy). Retrieved from
http://www.stata.com/meeting/italy12/abstracts/materials/it12_pisati.pdf.
Pischke, J. S., & Schwandt, H. (2014). Poorly Measured Confounders are More Useful on the Left
Than on the Right (Working Paper). London School of Economics.
Psacharopoulos, G. (1994). Returns to Investment in Education: A Global Update, World
Development, 22 (9), 1325-43.
Ravallion, M. (2011). On multidimensional indices of poverty. The Journal of Economic Inequality,
9(2), 235-248.
Ravallion, M., & Chen, S. (2003). Measuring pro-poor growth. Economics letters, 78(1), 93-99.
Reserve Bank of Malawi. (2015). Financial and Economic Review, 49 (4). Retrieved from
https://www.rbm.mw/Publications/EconomicReviews/.
Roy, A. D. (1951). Some thoughts on the distribution of earnings. Oxford economic papers, 3(2),
135-146.
Rutstein, S.O. & Johnson, K. (2004). The DHS Wealth Index (DHS Comparative Reports No. 6).
Calverton, Maryland: ORC Macro.
151
Stellenbosch University https://scholar.sun.ac.za
Sahn, D.E. & Stifel, D.C., 2003. Urban-rural Inequality in Living Standards in Africa. Journal of
African Economies, 12 (4), 564-97.
Sen, A. (1976). Poverty: an ordinal approach to measurement. Econometrica: Journal of the
Econometric Society, 219-231.
Sen, A. (1985). Commodities and Capabilities. North Holland, Amsterdam.
Sen, A. (1987). The Standard of living. In McMurrin, S. (Ed), Tanner lectures on Human Values,
Cambridge: Cambridge University Press. Retrieved from
http://tannerlectures.utah.edu/_documents/a-to-z/s/sen86.pdf.
Singer, H., & Jolly, R. (1972). Employment, Incomes and Equality: A Strategy for Increasing Productive
Employment in Kenya. Geneva: ILO.
Sokal, R. R., Oden, N. L. & Thomson, B. A. (1998), Local Spatial Autocorrelation in a Biological
Model. Geographical Analysis, 30: 331–54. doi: 10.1111/j.1538-4632.1998.tb00406.x
StataCorp. (Version 13.1). Stata Statistical Software [Computer Software]. College Station, Texas:
Stata Corporation.
StataCorp. (Version 14.1). Stata Statistical Software [Computer Software]. College Station, Texas:
Stata Corporation.
Tobler, W. (1970). A Computer Movie Simulating Urban Growth in the Detroit Region. Economic
Geography, 46, 234-40. doi:1. Retrieved from http://www.jstor.org/stable/143141 doi:1
Verduzco-Gallo, I., Ecker, O., & Pauw, K. (2014). Changes in Food and Nutrition Security in Malawi.
IFPRI Working Paper 06.
Villa, J.M. (2011). DIFF: Stata Module to Perform Differences in Differences Estimation. Statistical
Software Components. Boston College Department of Economics.
Willekens, F. (2008). Models of migration: Observations and judgement. International migration in
Europe: Data, models and estimates, 117-147.
Williams, R. (2015). Panel Data 4: Fixed Effects vs. Random Effects Models. University of Notre
Dame. Retrieved from https://www3.nd.edu/~rwilliam/stats3/Panel04-FixedVsRandom.pdf.
Wittenberg, M. (2014). Wages and wage inequality in South Africa 1994-2011: The evidence from
household survey data. (SALDRU working paper No. 135). Southern Africa Labour and
Development Research Unit, University of Cape Town.
Wittenberg, M., Leibbrandt, M. (2015). Measuring Inequality by Asset Indices: A general approach
with application to South Africa. (SALDRU Working Paper No. 141). Southern Africa Labour
and Development Research Unit, University of Cape Town.
Wooldridge, J. (2002). Econometric Analysis of Cross Section and Panel Data. Cambridge: The MIT
Press.
152
Stellenbosch University https://scholar.sun.ac.za
World Bank. (2004). Poverty reduction strategies papers sourcebook. Washington, D.C.: World Bank.
World Bank. (2012). Malawi - Community-Based Rural Land Development Project. Retrieved from
http://documents.worldbank.org/curated/en/2012/03/16215843/malawi-community-based-ruralland-development-project.
World Health Organisation. (2006). WHO Child Growth Standards: Methods and development.
Retrieved from http://www.who.int/childgrowth/standards/Technical_report.pdf.
Yu, K.C.D. (2012). Using household surveys for deriving labour market, poverty and inequality
trends in South Africa (Doctoral thesis, Stellenbosch University, South Africa). Retrieved from
http://scholar.sun.ac.za/handle/10019.1/71638.
153
Stellenbosch University https://scholar.sun.ac.za
Appendices
Appendix A1: Coefficient comparison test
The coefficient comparison was achieved by seemingly unrelated regression (SURE) which takes the
following code form in Stata 13.1:
.svy: regress lwage edu x if Male
.estimates store Male
.svy: regress lwage edu x if! Male
.estimates store Female
.suest Male Female
.test_b [Male:edu]=_b[Female: edu]
.test_b [Male:x]=_b[Female: edu]
Since suest (and its alternatives) do not work with xtreg , we conduct the coefficient comparison tests
for pooled Ordinary Least Squares (OLS) only.
Appendix A2: Dealing with data inconsistencies
In order to ensure data quality, we created a salary payment period that is consistent across the waves
using the following Stata code (in Stata 13.1) for the formal and informal sectors, respectively:
.bys PID: egen period1=max (salary_period)
.bys PID: egen period2=max (ganyu_dys)
It does not matter whether we use maximum (max) or minimum (min). The most important thing is to
come up with a period that is consistent across waves. Otherwise, the data will show that wages
increased while in actual fact it is as a result of changes in the period of payment.
154
Stellenbosch University https://scholar.sun.ac.za
Figure A 1: FISP evaluation panel data on real ganyu wages
Source: Farm Input Subsidy Programme (FISP) evaluation panel data
155
Stellenbosch University https://scholar.sun.ac.za
Table A 1: Table showing the distribution of households by year and population subgroups
2010
2004
2000
1992
Description
Number Percent Number Percent Number Percent Number Percent
Areas
Urban
4,116
16.6
2,262
16.6
1,949
13.7
603
11.3
Rural
20,709
83.4
11,402
83.5
12,264
86.3
4,720
88.7
Region
Northern
2,716
10.9
1,584
11.6
1,496
10.5
589
11.1
Central
10,627
42.8
5,589
40.9
5,744
40.4
2,043
38.4
Southern
11,482
46.3
6,491
47.5
6,973
49.1
2,691
50.6
Residence
Rural North
2,476
10.0
1,325
9.7
1,215
8.6
529
9.9
Rural Centre
8,876
35.8
4,615
33.8
5,049
35.5
1,822
34.2
Rural South
9,356
37.7
5,462
40.0
6,000
42.2
2,369
44.5
Urban North
240
1.0
259
1.9
281
2.0
60
1.1
Urban Centre
1,751
7.1
974
7.1
695
4.9
221
4.2
Urban South
2,126
8.6
1,029
7.5
973
6.9
322
6.1
Sex
Male head
17,857
71.9
10,295
75.3
10,431
73.4
4,013
75.4
Female head
6,968
28.1
3,369
24.7
3,782
26.6
1,310
24.6
Total
24,825
100
13,664
100
14,213
100
5,323
100
Table A 2: Table showing the distribution of children by year and population subgroups
2010
2004
2000
1992
Description
Number Percent Number Percent Number Percent Number Percent
Age
0-23
2,050
42.7
4,083
46.9
4,497 46.11
1,591
47.45
2,751
57.3
4,624
53.1
5,256 53.89
1,762
52.55
24-59
Sex
2,331
48.6
4,355
50.0
4,821
49.4
1,676
50.0
Male
Female
2,470
51.4
4,352
50.0
4,932
50.6
1,677
50.0
Areas
Urban
721
15.0
1,129
13.0
1,272
13.0
360
10.7
Rural
4,080
85.0
7,578
87.0
8,481
87.0
2,993
89.3
Region
Northern
514
10.7
1,174
13.5
1,064
10.9
399
11.9
Central
2,226
46.4
3,417
39.3
4,229
43.4
1,387
41.4
Southern
2,061
42.9
4,116
47.3
4,460
45.7
1,567
46.7
Residence
Rural North
460
9.6
1,001
11.5
855
8.8
354
10.6
Rural Centre
1,903
39.6
3,024
34.7
3,726
38.2
1,235
36.8
Rural South
1,717
35.8
3,553
40.8
3,899
40.0
1,404
41.9
Urban North
54
1.1
173
2.0
209
2.1
45
1.3
Urban Centre
323
6.7
393
4.5
502
5.2
152
4.5
Urban South
344
7.2
563
6.5
561
5.8
163
4.9
Total
4,801
100
8,707
100
9,753
100
3,353
100
156
Stellenbosch University https://scholar.sun.ac.za
Table A 3: WHZ regression results by age category
Age<24 months
Variables
Age>=24 months
Coeff
SE
Coeff
SE
Age in months
-0.080**
(0.029)
0.038
(0.025)
Square of age
0.286*
(0.114)
-0.050
(0.029)
Female child
0.036
(0.087)
-0.162**
(0.054)
Child is twin
-0.417
(0.250)
-0.101
(0.163)
Birth order number
0.045
(0.027)
0.032*
(0.013)
Incomplete primary
0.200
(0.146)
0.059
(0.093)
Complete primary
-0.009
(0.209)
0.186
(0.126)
Incomplete secondary
0.173
(0.201)
0.144
(0.131)
Complete secondary
0.480
(0.252)
0.257
(0.201)
Post-secondary
-0.089
(0.763)
0.012
(0.370)
Incomplete primary
-0.052
(0.164)
-0.028
(0.099)
Complete primary
0.084
(0.218)
0.096
(0.135)
Incomplete secondary
-0.141
(0.190)
-0.028
(0.113)
Post-secondary
0.476
(0.317)
-0.331
(0.288)
Asset index
-0.002
(0.086)
0.000
(0.060)
Square of asset index
0.007
(0.023)
0.009
(0.016)
Rural area
-0.163
(0.142)
-0.042
(0.109)
Central
-0.162
(0.131)
-0.044
(0.094)
Southern
-0.215
(0.124)
-0.142
(0.094)
0.652
(0.358)
-0.127
(0.517)
Mother's education
Father's education
Region
Constant
R-squared
0.026
0.017
Prob > F
0.004
0.097
Observations
1,863
2,668
Notes: *, **, *** denote significance at 10%, 5% and 1% levels
157
Stellenbosch University https://scholar.sun.ac.za
Table A 4: List of regions, districts and traditional areas consistent over time
Region
District
Traditional area
Northern Region
Chitipa
Chitipa Boma, Kameme, Mwabulambya
Mwenemisuku
Nyika National Park, Mwenewenya
Karonga
SC Mwakaboko, Silupula
Karonga, Kyungu
Wasambo
Karonga Boma
SC Mwirang'ombe
Nkhata Bay,
Likoma
Nkhata Bay Boma, Mkumpa, SC Mkumbira, Mankhambira
Malenga Mzoma, SC Fukamalaza, SC Malanda
Kabunduli
Timbiri
Fukamapiri, SC Zilakoma
Boghoyo, SC Mkondowe, SC Nyaluwanga, Musisya
Rumphi
Rumphi Boma, SC Mwahenga, Katumbi, SC Zolokere, Vwaza Marsh
Game, Chikulamayembe
SC Chapinduka, Mwamlowe, SC Mwankhunikira
SC Kachulu, SC Mwalweni
Mzimba
SC Jaravikuba Munthali, Mtwalo
Chindi
M'mbelwa
Mzikubola
Mabulabo
SC Kampingo Sibande
Vwaza Marsh Game Reserve, Mpherembe
SC Khosolo Gwaza
Mzukuzuku
Mzimba Boma, Mzuzu city
Central Region
Kasungu
Wimbe
Santhe
Chulu
SC Chisikwa, Kaluluma
SC Kawamba
158
Stellenbosch University https://scholar.sun.ac.za
Kaomba and Kasungu national park
Kasungu Township
SC Njombwa
SC Chilowamatambe
Kapelula
SC M'nyanja
SC Lukwa
SC Simlemba
Nkhota Kota
SC Kafulazira, Kanyenda
Mwadzama
Nkhota Kota Game Reserve, Malenga Chanzi
SC Mphonde
Nkhota Kota Boma
Mwansambo
Ntchisi
Ntchisi Boma, Kalumo
SC Chilooko
Kasakula, SC Nthondo
Chikho
Dowa
SC Chakhaza
SC Kayembe
Chiwere
Dowa Boma, Mkukula
Dzoole
Msakambewa
Mponela Urban, SC Mponela
Salima
SC Msosa, Khombedza, Kuluunda
Chipoka Urban, SC Kambalame, Ndindi
Maganga
Salima Township
SC Kambwiri
SC Mwanza
Pemba
Lilongwe
Chiseka
Kalolo
Kabudula
Chadza
159
Stellenbosch University https://scholar.sun.ac.za
Khongoni
Mazengera
Malili
Chimutu
Kalumbu
SC Mtema
SC Chitekwele
SC Njewa
Tsabango
Chitukula
Kalumba
Lilongwe city
Mchinji
Mchinji Boma, Zulu
SC Mavwere
Mkanda
SC Mduwa
SC Dambe
Mlonyeni
Dedza
Kaphuka
Pemba
Kachindamoto
Kasumbu
Tambala
SC Chilikumwendo
SC Kamenya Gwaza
Dedza Township
SC Chauma
Ntcheu
Goodson Ganya
Champiti, Makwangwala
Ntcheu Boma, Kwataine
Phambala
Njolomole
Mpando
Chakhumbira
Masasa
Southern Region
Mangochi
Palm Beach forest, Monkey Bay urban
160
Stellenbosch University https://scholar.sun.ac.za
Chimwala
Mponda
Chowe
SC Mbwana Nyambi
Jalasi
Makanjila
Katuli
Mangochi Boma
SC Namabvi
Machinga
Kawinga
Liwonde National Park, Liwonde, Lake Malawi national park
SC Chikweo
Nyambi and SC Chiwalo
SC Mlomba
Machinga Boma, Sitola, Liwonde Township
SC Ngokwe
SC Mposa
Chamba
Zomba
SC Mbiza
Mlumbe
Mwambo
Kuntumanji, Mkumbira
Malemia
Chikowi
Zomba municipality
Chiradzulu
Chiradzulu Boma, Chitera, Mpama
Kadewere
Likoswe
Nkalo
Mchema
Blantyre
Kapeni
Kuntaja
Somba
Chigeru
Makata, Machinjili
Kunthembwe
161
Stellenbosch University https://scholar.sun.ac.za
Lundu
Blantyre city
Mwanza
Mwanza Boma, Nthache
Kanduku
Symon
Mlauli
Dambe
Ngozi
Thyolo
Luchenza Township, Chimaliro
Bvumbwe
Thyolo Boma, Nchilamwela
Kapichi
SC Thukuta, Nsabwe
SC Mphuka
SC Kwethemule
SC Mbawela
Changata
Thomas
Mulanje
Mabuka
Nkanda, Mulanje Mountain
Mulanje Boma, Chikumbu
Juma
Laston Njema
Nthiramanja
Phalombe
Phalombe Boma, Mkhumba
Nazombe
Chikwawa
Ngabu Urban, Ngabu
Chapananga, Lengwe National Park
Mankhwira
Lundu
Chikwawa Boma, Katunga
Kasisi, Majete Game Reserve
Maseya
Nsanje
Mlolo
162
Stellenbosch University https://scholar.sun.ac.za
Malemia, Tengani, Mwabvi Game Reserve
SC Mbenje
SC Makoko, Nyachikadza, Ndamera
Chimombo, Ngabu
Nsanje Boma
Balaka
Msamala and Balaka town
Kalembo
Source: Own compilation from census data
Table A 5: Results showing spatial dependencies in variables for 1987
Variables
Moran's I
SD
p-value Geary's c
Migration
0.010
0.005
0.004***
0.998
Employed
0.015
0.005
0.000***
0.996
Years of schooling
0.100
0.005
0.000***
0.866
Sex
0.000
0.005
0.540
0.997
Age
0.060
0.005
0.000***
0.935
Married
0.081
0.005
0.000***
0.899
Dependency ratio
0.039
0.005
0.000***
0.966
Note: *, **, *** denote significance at 10%, 5% and 1% levels
SD
0.017
0.015
0.007
0.005
0.009
0.007
0.008
p-value
0.909
0.792
0.000***
0.540
0.000***
0.000***
0.000***
SD
.
0.013
0.007
0.005
0.01
0.007
0.008
p-value
.
0.000***
0.000***
0.540
0.003***
0.000***
0.000***
SD
0.006
0.007
0.008
0.005
0.011
0.008
0.009
p-value
0.000***
0.017***
0.000***
0.540
0.890
0.000***
0.000***
Table A 6: Results showing spatial dependencies in variables for 1998
Variables
Moran's I
SD
p-value Geary's c
Migration
.
.
.
.
Employed
0.059
0.005
0.000***
0.924
Years of schooling
0.126
0.005
0.000***
0.841
Sex
0.000
0.005
0.540
0.997
Age
0.043
0.005
0.000***
0.971
Married
0.069
0.005
0.000***
0.913
Dependency ratio
0.033
0.005
0.000***
0.960
Note: *, **, *** denote significance at 10%, 5% and 1% levels
Table A 7: Results showing spatial dependencies in variables for 2008
Variables
Moran's I
SD
p-value Geary's c
Migration
0.420
0.005
0.000***
0.591
Employed
0.025
0.005
0.000***
0.983
Years of schooling
0.107
0.005
0.000***
0.870
Sex
0.000
0.005
0.540
0.997
Age
0.010
0.005
0.005***
1.002
Married
0.043
0.005
0.000***
0.928
Dependency ratio
0.074
0.005
0.000***
0.912
Note: *, **, *** denote significance at 10%, 5% and 1% levels
163
Stellenbosch University https://scholar.sun.ac.za
Table A 8: Population figures for individuals aged (15 years and older)
1987
District
Male
Female
Chitipa
20,100
24,450
Karonga
32,580
37,540
Nkhata Bay,
Likoma
31,780
35,390
Rumphi
22,100
24,750
Mzimba
108,440
117,170
Kasungu
89,990
78,670
Nkhota Kota
41,230
39,810
Ntchisi
30,030
31,700
Dowa
81,260
86,150
Salima
47,140
52,590
Lilongwe
246,180
254,090
Mchinji
63,850
56,860
Dedza
83,060
108,420
Ntcheu
64,030
87,060
Mangochi
103,950
127,900
Machinga
110,070
140,180
Zomba
105,430
128,100
Chiradzulu
48,330
64,420
Blantyre
168,480
145,910
Thyolo
102,140
118,680
Mulanje
132,890
177,440
Chikwawa
68,390
70,110
Nsanje
33,870
39,900
Neno
23,600
29,200
Total
1,858,920 2,076,490
Source: Own computation from census data
1998
Male
Female
29,740
35,780
49,130
56,430
2008
Male
Female
42,900
47,600
66,590
73,190
45,300
49,980
33,720
35,840
161,690
173,590
139,600
125,510
63,140
61,800
44,960
46,330
111,530
114,430
67,400
71,110
381,780
371,920
87,940
86,530
120,660
142,510
92,040
110,580
158,050
179,370
160,250
186,870
151,760
164,380
62,160
76,890
252,610
229,830
120,750
141,180
170,600
210,110
95,930
98,280
49,520
55,770
35,320
38,970
2,685,580 2,863,990
56,980
61,000
45,120
46,780
228,090
244,440
170,160
164,410
79,230
79,550
63,000
63,740
152,060
158,700
89,810
93,730
540,870
534,880
121,150
119,380
159,680
182,870
123,020
140,760
194,000
218,630
196,960
223,400
174,900
192,770
74,070
88,790
299,600
288,050
150,130
171,600
206,610
244,320
117,710
117,230
62,890
64,510
50,140
54,810
3,465,670 3,675,140
164
Stellenbosch University https://scholar.sun.ac.za
Table A 9: Number of people who migrated to other districts (15 years and older)
1987
Male Female
Chitipa
206
281
Karonga
375
411
Nkhata Bay, Likoma
434
418
Rumphi
453
460
Mzimba
1,914
1,877
Kasungu
2,040
1,438
Nkhota Kota
955
814
Ntchisi
464
410
Dowa
1,198
1,105
Salima
956
757
Lilongwe
3,751
3,236
Mchinji
1,203
903
Dedza
1,129
897
Ntcheu
637
668
Mangochi
1,817
1,672
Machinga
2,098
1,888
Zomba
1,800
1,503
Chiradzulu
593
595
Blantyre
4,457
3,101
Thyolo
1,626
1,522
Mulanje
1,462
1,310
Chikwawa
862
687
Nsanje
328
325
Neno
252
283
Total
31,010 26,561
Source: Own computation from census data
Total
487
786
852
913
3,791
3,478
1769
874
2,303
1713
6,987
2106
2026
1305
3,489
3,986
3,303
1188
7,558
3,148
2,772
1549
653
535
57,571
165
2008
Male Female Total
61
64
125
182
185
367
296
213
509
223
199
422
1,153
1,036
2,189
792
664
1,456
244
212
456
164
146
310
275
286
561
306
270
576
3,201
2,590
5,791
295
297
592
250
228
478
281
265
546
705
659
1,364
713
662
1,375
931
836
1,767
259
276
535
4,428
3,805
8,233
733
676
1,409
685
646
1,331
563
527
1,090
257
264
521
346
327
673
17,343 15,333 32,676
Stellenbosch University https://scholar.sun.ac.za
Table A 10: Proportions of people who migrated to other districts (15 years and older)
1987
Male
Female
Chitipa
10.2%
11.5%
Karonga
11.5%
10.9%
Nkhata Bay, Likoma
13.7%
11.8%
Rumphi
20.5%
18.6%
Mzimba
17.7%
16.0%
Kasungu
22.7%
18.3%
Nkhota Kota
23.2%
20.4%
Ntchisi
15.5%
12.9%
Dowa
14.7%
12.8%
Salima
20.3%
14.4%
Lilongwe
15.2%
12.7%
Mchinji
18.8%
15.9%
Dedza
13.6%
8.3%
Ntcheu
9.9%
7.7%
Mangochi
17.5%
13.1%
Machinga
19.1%
13.5%
Zomba
17.1%
11.7%
Chiradzulu
12.3%
9.2%
Blantyre
26.5%
21.3%
Thyolo
15.9%
12.8%
Mulanje
11.0%
7.4%
Chikwawa
12.6%
9.8%
Nsanje
9.7%
8.1%
Neno
10.7%
9.7%
Total
15.8%
12.9%
Source: Own computation from census data
Total
10.9%
11.2%
12.7%
19.5%
16.8%
20.6%
21.8%
14.2%
13.8%
17.2%
14.0%
17.5%
10.6%
8.6%
15.1%
15.9%
14.1%
10.5%
24.0%
14.3%
8.9%
11.2%
8.9%
10.1%
14.3%
Male
1.4%
2.7%
5.2%
4.9%
5.1%
4.7%
3.1%
2.6%
1.8%
3.4%
5.9%
2.4%
1.6%
2.3%
3.6%
3.6%
5.3%
3.5%
14.8%
4.9%
3.3%
4.8%
4.1%
6.9%
4.2%
2008
Female
1.3%
2.5%
3.5%
4.3%
4.2%
4.0%
2.7%
2.3%
1.8%
2.9%
4.8%
2.5%
1.3%
1.9%
3.0%
3.0%
4.3%
3.1%
13.2%
3.9%
2.6%
4.5%
4.1%
6.0%
3.7%
Total
1.4%
2.6%
4.3%
4.6%
4.6%
4.4%
2.9%
2.5%
1.8%
3.1%
5.4%
2.5%
1.4%
2.1%
3.3%
3.3%
4.8%
3.3%
14.0%
4.4%
3.0%
4.6%
4.1%
6.4%
3.9%
Note: As earlier stated, four of the districts in Malawi consist of cities, namely Mzuzu city (in Mzimba
district), Lilongwe city (in Lilongwe district), Zomba municipality or city (in Zomba district) and
Blantyre city (in Blantyre district). A movement from the rural areas of the district into the city (e.g.
from Blantyre rural to Blantyre city) and vice versa is recorded as migration. As a result, the high
migration patterns observed in Blantyre and Lilongwe (which are the two major cities in Malawi) could
largely be rural-urban migration involving people moving rural areas to urban areas within the district
boundaries.
166
Stellenbosch University https://scholar.sun.ac.za
Table A 11: Effects of land reform policy on migration
Male
Female
SEM
SAC
SEM
SAC
Time
-0.170*** -0.102***
-0.028
-0.016
(0.025)
(0.038)
(0.023)
(0.039)
DID
0.015
0.015
0.006
0.017
(0.016)
(0.016)
(0.017)
(0.015)
Schooling
-0.013
-0.013
-0.037*** -0.029***
(0.010)
(0.010)
(0.010)
(0.010)
Age
0.088
0.088
-0.179*** -0.151***
(0.056)
(0.056)
(0.051)
(0.052)
Age squared
-0.002**
-0.002**
0.002***
0.002**
(0.001)
(0.001)
(0.001)
(0.001)
Married
0.384***
0.382***
0.074
0.076
(0.125)
(0.123)
(0.116)
(0.110)
Employed
-0.084**
-0.085**
-0.103*** -0.109***
(0.039)
(0.038)
(0.023)
(0.023)
Assets
0.033***
0.033***
-0.012
-0.008
(0.010)
(0.010)
(0.010)
(0.010)
W*migration
0.564**
0.447
(0.260)
(0.284)
Lambda
0.755***
0.528*
2.024***
0.789***
(0.155)
(0.302)
(0.145)
(0.152)
Observations
356
356
356
356
Spatial units
178
178
178
178
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; standard errors in parenthesis
167
Stellenbosch University https://scholar.sun.ac.za
Table A 12: Effects of land reform policy on agricultural employment
Male
Female
SEM
SAC
SEM
SAC
Time
-0.387*** -0.352**
-0.436*** -0.324***
(0.033)
(0.137)
(0.043)
(0.123)
DID
0.028
0.027
0.053**
0.050**
(0.021)
(0.021)
(0.021)
(0.021)
Schooling
0.074*** 0.073***
0.054*
0.054*
(0.024)
(0.024)
(0.028)
(0.028)
Age
0.286
0.279
1.018***
1.031***
(0.542)
(0.543)
(0.341)
(0.343)
Age squared
-0.004
-0.004
-0.016*** -0.017***
(0.009)
(0.009)
(0.006)
(0.006)
Married
-0.156
-0.153
-0.166
-0.151
(0.309)
(0.309)
(0.271)
(0.269)
Dependency
-0.268*
-0.265*
0.096
0.101
(0.142)
(0.142)
(0.113)
(0.111)
W*employed
0.113
0.342
(0.438)
(0.352)
Lambda
0.334
0.275
0.611***
0.479
(0.319)
(0.412)
(0.218)
(0.309)
Observations
356
356
356
356
Spatial units
178
178
178
178
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; standard errors in parenthesis
168
Stellenbosch University https://scholar.sun.ac.za
Table A 13: Effects of land reform policy on government employment
Male
Female
SEM
SAC
SEM
SAC
Time
0.079**
-0.002
-0.011
-0.024
(0.038)
(0.021)
(0.012)
(0.017)
DID
0.028*** 0.024***
0.013**
0.012**
(0.007)
(0.007)
(0.006)
(0.005)
Schooling
0.007
0.011
0.033***
0.033***
(0.008)
(0.007)
(0.007)
(0.007)
Age
-0.277*
-0.264*
0.432***
0.436***
(0.162)
(0.159)
(0.087)
(0.088)
Age squared
0.005*
0.004*
-0.007*** -0.007***
(0.003)
(0.003)
(0.001)
(0.001)
Married
-0.102
-0.113
0.111
0.110
(0.092)
(0.089)
(0.070)
(0.070)
Dependency
0.021
0.016
0.042
0.041
(0.044)
(0.043)
(0.029)
(0.029)
W*employed
0.799***
0.324
(0.130)
(0.369)
Lambda
0.922*** 0.830***
0.707***
0.582**
(0.054)
(0.115)
(0.174)
(0.288)
Observations
356
356
356
356
Spatial units
178
178
178
178
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; standard errors in parenthesis
169
Stellenbosch University https://scholar.sun.ac.za
Table A 14: Effects of land reform policy on self, wage and private employment (OLS only)
Male
Female
Self
Private
Self
Private
Time
0.228***
0.185***
0.213***
0.108***
(0.020)
(0.018)
(0.019)
(0.013)
Treated
-0.018
-0.007
0.002
-0.014
(0.015)
(0.014)
(0.015)
(0.010)
DID
-0.043**
-0.040**
-0.021
-0.041***
(0.021)
(0.019)
(0.020)
(0.014)
Schooling
-0.001
0.020***
0.008*
0.011***
(0.006)
(0.005)
(0.004)
(0.003)
Age
-1.436***
-0.298
-0.691***
-0.106
(0.310)
(0.284)
(0.161)
(0.110)
Age squared
0.023***
0.005
0.011***
0.002
(0.005)
(0.005)
(0.003)
(0.002)
Married
-0.274*
-0.047
-0.241**
0.090
(0.159)
(0.146)
(0.099)
(0.068)
Dependency
-0.134
-0.324***
-0.030
-0.129**
(0.106)
(0.097)
(0.080)
(0.055)
Constant
22.764***
4.543
10.835***
1.62
(4.826)
(4.416)
(2.428)
(1.658)
R-squared
0.580
0.572
0.657
0.428
Observations
356
356
356
356
Notes: *, **, *** denote significance at 10%, 5% and 1% levels; standard errors in parenthesis
170
Anderson Gondwe