ECONOMETRIC ANALYSIS OF FIRM
PRODUCTIVITY:
Fixed Effects, Random Effects, and Hausman Test
ACADEMIC YEAR 2023–2024
Nicola Ciotti (r0962417)
This paper was written in the course
Advanced Applied Econometrics,
Prof. Dr. Dirk Czarnitzki.
1 Introduction
This paper details the econometric analysis performed to understand the determinants of
productivity in Chinese firms using panel data. The objective of this study is to analyze the
factors influencing productivity in the nonmetallic mineral product manufacturing industry,
estimating the production function. The data used in this analysis is a balanced panel with
N = 2,804 and T = 6 (during 2000-2006) from Malikov, E., Zhao, S., & Kumbhakar, S. C.
(2020).
The dataset is composed of 12 variables:
• Firm: ID of the Firm.
• Year: time variable, from 2000 to 2006.
• Ind: the number of the Industry, in this case it’s always 31, as all fata is taken from the
same industry.
• y: log deflated gross industrial output.
• k: log deflated net fixed assets.
• k1: 1-period lag of k.
• l: log deflated total wage bill plus benefits.
• l1: 1-period lag of l.
• m: log deflated total intermediate inputs, including raw materials and other production-
related inputs.
• m1: 1-period lag of m.
• x: share of the export in output.
• x1: 1-period lag of x.
2 Methodology
Initially, we performed a pooled OLS regression to estimate the relationship between the
industrial output and the explanatory variables. The pooled panel model is as follow:
yit = β0 + β1 kit + β2 lit + β3 mit + β4 xit + γ1 + γ2 + γ3 + γ4 + γ5 + γ6 + ϵit
With the γi as time dummies for the variable Year, from 2001 to 2006 respectively.
An important assumption for this is the strict exogeneity, which implies that the explana-
tory variables must be completely uncorrelated with the error term at all time periods.
E(ϵit |kit , lit , mit , xit ) = 0 ∀i, t
To address potential heteroscedasticity, the robust standard errors were calculated for the
pooled OLS estimates.
1
Now we’ll implement the two basic static linear models in a panel data setting, the fixed
effects and the random effects model, and then we’ll discuss the choice between the two using
a specification test.
The fixed effects model is estimated to account for unobserved heterogeneity across firms,
by allowing each firm to have its own intercept. this allows the firm-specific intercepts αi to
capture the unobserved heterogeneity. In this case the model is written as:
N
X
yit = αj dij + β0 + β1 kit + β2 lit + β3 mit + β4 xit + γ1 + γ2 + γ3 + γ4 + γ5 + γ6 + ϵit
j=1
with dij as dummy variables, where dij = 1 if j = i and 0 elsewhere. The unobserved effect
may be correlated with the explanatory variables, but the assumption of strict exogeneity still
holds.
The random effects model is estimated to handle potential correlation between the regres-
sors and the individual effects. It assumes that the individual-specific effects are random and
uncorrelated with the explanatory variables. The model is specified similarly to the fixed
effects model but the individual effects are now considered as random error term.
yit = β0 + β1 kit + β2 lit + β3 mit + β4 xit + γ1 + γ2 + γ3 + γ4 + γ5 + γ6 + αi + ϵit
In this case the error term consists of two components: an individual specific part, that doesn’t
change with time, and a remainder part, assumed to be uncorrelated over time. It’s also
assumed that these two parts are mutually independent and uncorrelated with the explanatory
variables. If these assumptions holds then this is an efficient parameter estimation.
The specification test is needed to understand how to treat the individual effect αi as fixed
or random, i.e. to decide between the fixed effects and random effects models.
The Hausman test is conducted to compare those two, and it determines whether the
unique errors are correlated with the regressors.
With as the test statistics:
H = (β̂F E − β̂RE )′ [V ar(β̂F E ) − V ar(β̂RE )]−1 (β̂F E − β̂RE )
Where β̂i are the estimated coefficients. In this test the null hypothesis is for the explanatory
variables and individual effect are correlated. In this way we can compare the two estimators,
as the fixed effects one is consistent whether those two are are correlated or not, meanwhile
the random effects estimator is consistent and efficient only if the explanatory variables is
uncorrelated with the individual effect.
3 Analysis and Results
In this section we’ll explore the results got from an analysis using the software Stata.
2
3.1 Pooled OLS regression
Next there are the stata results of the pooled OLS regression:
Figure 1: Pooled OLS result
As we can see almost all the coefficients are statistically significant, apart from the ones
related to the dummy variables for the year 2001 (p = 0.066 > 0.05) and the year 2003
(p = 0.187 > 0.05). So, as this result suggests, those two years don’t seem to be statistically
different from the year 2000 regarding the influence on the industrial output.
3.2 Fixed Effects Model
Next there is the Fixed effect model output:
The within R-squared score indicates that the model explains 83.38% of the variation in
the dependent variable within each firm. Meanwhile the between R-squared of 0.9796 indicates
the model explains a really high percentage of the variation between the firms and the overall
R-squared score indicates the model explains 95.52% of the total variation in the data. So
we’ve got a model that can explain quite well the variation of the data, as we can see from
the significant value of the F-test, which means that the model is a good fit overall.
The first four coefficients, relates to the explanatory variables, are all positive, apart from
the one related to x which is negative, and have a significant effect on y at the 1% level.
Specifically, k and l have smaller but significant effects on the dependent variable. The inclu-
sion of year dummies shows varying impacts over different years, but the ones related to the
3
Figure 2: Fixed effects model
year 2001 and 2002 are the only ones not statistically significant, showing a similarity with
the year 2000, which is a result similar with the OLS regression, for the year 2001.
This model value of ρ, which can range from 0 to 1, where an high value indicates that
most of the variability is between individuals (firms), is higher than the value for the Random
Effects suggesting that a considerable portion of the variance is due to unobserved firm-specific
factors. This will be tested with the Hausman test later.
4
3.3 Random Effects Model
As we can see in the radom effects model output in the Figure 3, the within R-squared value
is very close to the one of the fixed effects model. Both the between R-squared and the overall
R-squared values are slightly higher than in the fixed effects model, but the difference is not
that high. This means that this model too seems to indicate a good fit of the model, backed
by the significance of the chi-squared test.
Figure 3: Random effects model
Similarly to what we’ve seen before in the fixed effects model, we’ve a significant effect on
y for the first four coefficients, and we can also notice that their values are a bit lower for k,
l and x but not for m. This also happens for the dummy variables, and we can see that for
the random effect only the one related to the year 2001 is not significant, which is something
that has happened for all of the models seen so far.
The lower σu and ρ in the random effects model suggest that it assumes less unobserved
heterogeneity and attributes more of the variance to within-firm idiosyncratic errors.
5
3.4 Specification Test
As we’ve seen before the null hypothesis for the Hausman test is that the explanatory variables
and individual effect are correlated, which means that the preferred model is random effects.
Since the p-value is 0.0000, we reject the null hypothesis, indicating that the fixed effects
model is more appropriate.
Figure 4: Haussman test results
This suggests that the individual-specific effects are correlated with the independent vari-
ables, making the random effects model inconsistent. This also explains the high value of ρ
in the Fixed effects model, which indicates indicating that these unobserved factors are corre-
lated with the explanatory variables. Meanwhile the value in the random effects model is lower
as it fails to account for the correlation between the unobserved effects and the explanatory
variables.
4 Conclusion
The strong positive impact of m on y suggests that this variable (representing the total in-
termediate inputs) is crucial for the output or productivity measure. The negative coefficient
for x tells us that this variable negatively affects the outcome, which could indicate an inverse
relationship that needs to be managed or further investigated.
Given the results of the specification test the fixed effects model is more suitable for the
data as it better accounts for unobserved heterogeneity across firms. The significant σu and
6
higher rho indicate that firm-specific factors play an important role in explaining the variability
in the dependent variable.
To conclude we can write the final model as:
N
X
yit = αj dij +1+0.03kit +0.082lit +0.796mit −0.018xit +0.034d3 +0.103d4 +0.159d5 +0.204d6 +ϵit
j=1
dropping the dummy variable for the year 2001 and 2002 as not statistically significant.
This study utilizes panel data econometrics to analyze the determinants of productivity
in Chinese firms. By following a systematic approach of estimating pooled OLS, fixed and
random effects models, and conducting specification tests, we derive meaningful insights into
the factors affecting productivity. This methodology can be applied to similar datasets to
understand firm-level economic phenomena.
5 References
Wang, T., Yao, F., Kumbhakar, S. C. (2024). A flexible stochastic production frontier model
with panel data. Journal of Applied Econometrics.
