ECON0019: Quantitative Economics and

Econometrics
Lecture 9: Pooled regression, differences in differences
and panel data

Professor Dennis Kristensen

UCL

December 6, 2018

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 1 / 32

Today’s lecture

Pooled regression: Pool data sets (across time) to evaluate changes

over time (e.g., in the return to education)
Further use: Policy evaluation (“Di¤erence in Di¤erences” aka “Dif
in Dif“)
Introduction to panel data: When we have observed the same
observational unit over multiple time periods.

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 2 / 32

Pooled regression: Data with a time dimension

Until now, we have covered regression analysis using only

cross-sectional data
Now, we will take a look at data with a time dimension:
repeated cross sections (focus of today)
panel data (introduction this week, continue next week)
How to use the time dimension of the data to:
measure changes over time (e.g., changes in return to education)
better estimate the e¤ect of public policies (e.g., estimate the e¤ect of
the minimum wage on employment)

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 3 / 32

Repeated cross-sections

Many data collections are repeated on a regular basis – random

sampling from a large population at di¤erent points in time
Examples: census, labor force surveys, etc.
Pooling/combining these di¤erent samples (and using pooled
regression)
increases our sample size
but more importantly allows us to analyze changes over time
we still (possibly...) have random and independent sampling
) we can use standard regression techniques

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 4 / 32

Repeated cross-sections: What Piketty spends his time
on...
Wage inequality rises greatly in the US from early 1980s and onwards

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 5 / 32

Example 1: Changes in wage equation over time

We can use repeated cross sections to learn more about the changes
in the income distribution.
We have two data sets from the US Current Population Survey : One
from 1978 and one from 1985. Let wageit be the wage of individual i
at time t (= 1978 or 1985):
wagei 1978 is individual i’s wage the 1978 data set, i = 1, ..., n1 (= size
of 1st sample)
wagei 1985 is individual i’s wage the 1985 data set, i = 1, ..., n2 (= size
of 2nd sample)
These are two di¤erent individuals! The two samples consist of
di¤erent (randomly sampled) individuals.
We can combine the two data sets to analyse how returns to
education and the gender wage gap changed between 1978 and 1985.

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 6 / 32

Example 1: Pooled wage equation

log(wageit ) = β0 + β1 educit + β2 y 85it + β3 (educit y 85t ) + β4 expit

+ β5 expit2 + β6 unionit + β7 femalei + β8 (femalei y 85t ) + uit

where y 85t = 1 if t = 1985, 0 otherwise; expit = experience of individual
i in year t, etc
Intercept for 1978 is β0 , while intercept for 1985 is β0 + β2
Returns to education: β1 in 1978, while β1 + β3 in 1985
Gender wage gap (with men as baseline): β7 in 1978, while β7 + β8
in 1985
Other explanatory variables: Assumed constant e¤ect over time (and so
no interactions with y 85)

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 7 / 32

Example 1: Equivalent representation as structural break
regression
On the previous slide, we explicitly keep track of which data set a
given individual belongs to through subscript t.
But we can also pool the two data sets into one, and then include an
additional regressor that for each individual tells us which year he/she
was "observed". Let y 85i = 1 if individual i belongs to 1985 data
set, 0 otherwise.
Then the following regression is equivalent to the previous one
log(wagei ) = β0 + β1 educi + β2 d85i + β3 (educi y 85i ) + β4 expi
+ β5 expi2 + β6 unioni + β7 femalei + β8 (femalei y 85i ) + ui ,
where wagei is individual i in the pooled data set, etc., for
i = 1, ..., n1 + n2 (= size of pooled sample)
Since each sample is random, then so is the pooled.
So a pooled regression is just a special version of structural break
regression.
D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 8 / 32
Example 1: STATA output
Changes in return to education and gender wage gap

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 9 / 32

Example 1: STATA output
Estimate the changing return to education/being female w and w/o controls

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 10 / 32

Example 1, …ndings:
Returns to education increased; wage gap decreased

Educated and women saw signi…cant increases in their wages. (Chow

test rejects that y 85 was insigni…cant)
The change in wages for the educated in the US (and Britain too)
over the 1970s-today was more extreme than in other countries
Evidence that it was related to government policies (deregulation,
reduction in trade barriers in both the US and UK)?
Important assumption of our model:
the e¤ect of the (uninteracted) explanatory variables remained
constant
We can check this assumption by interacting remaining explanatory
variables with y 85.
If all explanatory variables are interacted with time, we are estimating
two separate regression models, one for 1978 and one for 1985.
D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 11 / 32
Example 1: Fully interacted model
Estimate the changing return to education/being female, without, with full set of
interactions

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 12 / 32

Policy analysis with pooled data

We are often interested in e¤ects of events/policy reforms that a¤ect

some people but not others at a given point in time
Example: Changes in minimum wages
In the US, some states might raise the minimum wage at a point in
time, but not others
Repeated cross-sections can be used to analyze the impact of such
policy reforms
Key idea: Collect a random sample of individuals before and after
reform. Compare individuals a¤ected by reforms to those not a¤ected
by reforms.
This approach is sometimes referred to as the quasi-experimental
approach
Possible to estimate the e¤ect of the policy under transparent (and
sometimes credible) assumptions

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 13 / 32

The quasi-experimental approach

yit = outcome of individual i in period t (e.g., y = is employment, t is

time)
2 periods: before (t = 1) and after (t = 2) the policy reform
2 groups: called treatment group (T ) and control group (C ):
before: (t = 1)
- neither group is treated by the policy change (e.g., before the
minimum wage hike)
after (t = 2):
- the treatment group is treated by the policy (e.g., after the
minimum wage hike)
- the control group is not treated by the policy (no minimum wage
hike)
We wish to estimate the e¤ect of the treatment on y

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 14 / 32

The quasi-experimental approach

Estimate the following regression (w/o controls for now )

yit = β0 + δ0 d2it + β1 dTit + δ1 d2it dTit + uit

where dT = 1 if treated, = 0 if control; d2 = 1 if after the reform,

= 0 if before the reform.
Parameter of interest is δ̂1 : Measures the added growth after the
reform in the treatment relative to controls.
Predicted values:
8
>
> β̂0 d2it = 0, dTit = 0
<
β̂0 + δ̂0 , d2it = 1, dTit = 0
ŷit =
>
> β̂0 + β̂1 , d2it = 0, dTit = 1
:
β̂0 + β̂1 + δ̂0 + δ̂1 , d2it = 1, dTit = 1

But the predicted values are just sample means for di¤erent groups:
e.g., β̂0 = mean of control group before reform (d2it = 0, dTit = 0).
D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 15 / 32
Quasi-experiments and di¤erence in di¤erences

Predicted values are just sample means for di¤erent groups )

another procedure to estimate δ̂1 is the following:

Control group : ȳ1,C = β̂0 and ȳ2,C = β̂0 + δ̂0

Treatment group : ȳ1,T = β̂0 + β̂1 and ȳ2,T = β̂0 + β̂1 + δ̂0 + δ̂1 .
Which explains why OLS estimator of δ1 is sometimes referred to as
di¤erence in di¤erences estimator,

δ̂1 = (ȳ2,T ȳ2,C ) (ȳ1,T ȳ1,C ).

Before After Time e¤ect

Control β̂0 β̂0 + δ̂0 δ̂0
Treatment β̂0 + β̂1 β̂0 + β̂1 + δ̂0 + δ̂1 δ̂0 + δ̂1
Treatment e¤ect β̂1 β̂1 + δ̂1 δ̂1

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 16 / 32

Di¤erence-in-Di¤erences (Dif-in-Dif) as regression

The Dif-in-Dif estimator can be written as the OLS estimator in a

regression model. Which should we use?
Advantages of regression over simple di¤erences in di¤erences:
convenient way to obtain standard errors
straightforward to control for (time-varying) observable characteristics
to regression
Empirical example: E¤ect of change in worker compensation laws
Research question: How does more generous bene…ts a¤ect the
duration of a worker’s sick leave?
Meyer, Viscusi and Durbin (1995) collected data from Kentucky and
Michigan on how long workers stayed on bene…ts after being injured.
They collected data before and after the two states raised the cap on
how much an invidual could earn and still receive bene…ts.

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 17 / 32

Example 2: Change in worker’s compensation

Treatment group: High Earners; Control group: Low Earners.

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 18 / 32
Example 2: E¤ect of change in worker compensation laws
The Dif-in-Dif estimate is the coe¢ cient on afhigh (= δ̂1 )

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 19 / 32

Dif-in-Dif assumption: only common trends
Why not just estimate change in duration of treated group (high
earners) before and after reform?
Because any e¤ects that we observe maybe due to other underlying
trends/changes over time.
Dif-in-Dif takes care of these trends – but only if they are common to
both treated and control group.
This is called the common trend assumption – choice of control group
is crucial

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 20 / 32

Dif-in-Dif with group-speci…c trends

What if there are group speci…c trends? One could, for example,
imagine that high earners were a¤ected by economic, demographic or
health-related trends di¤erently compared to the low earners.
Or that the composition of the two groups changed (more female
high-earners, for example)
Then we can’t distinguish between the e¤ect of the treatment and
the e¤ect of these trends.
These can be controlled for by including control variables in our
Dif-in-Dif regression
Meyer, Viscusi and Durbin (1995) did just so (gender, age, industry,
etc), but found controlling for these had little e¤ect on their their
estimate of δ1 .

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 21 / 32

Panel Data

We now consider data where we have repeated observations for the

same units over time.
This is called panel data (aka longitudinal data).
Estimators using panel data have the following advantages
1 More observations (and so more precise estimates)
2 Informative about dynamics
3 Allow us to control for unobserved invidividual speci…c e¤ects (and so
reduce the risk of omitted variable bias)
1.-2. are shared with pooled data sets, 3. is not.

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 22 / 32

Panel data example: crime and unemployment

Suppose we are interested in the relationship between the

unemployment rate and crime
Having collected data on crime and unemployment in 46 US cities for
1982 and 1987 (so we have 46 2 = 92 observations).
We decide to run the following SLR,

crimeit = β0 + β1 unemit + vit

where crimeit and unemit is the level of crime and unemployment,

respectively, in city i in year t; vit is the error term
Note: crimei 1 and crimei 2 are now two observations of the same city
(in contrast to pooled data)
We might guess that crime is high when unemployment is high (so
β1 > 0)

“Idle hands are the devil’s work”

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 23 / 32
Data on unemployment and crime

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 24 / 32

Regression of crime on unemployment

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 25 / 32

Crime rate (per 1000 people) versus unemployment rate,
actual and predicted values

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 26 / 32

Omitted variable bias in cross-sectional regression

So unemployment appears to do a poor job at explaining crime – so

much for our hypothesis...
We could try to pool the two samples to get more precise estimates
But would the resulting pooled regression capture a causal e¤ect?
Unlikely since we expect E [vit junemit ] 6= 0 because of city-speci…c
crime factors that we haven’t controlled for.
Speci…cally, we expect that each city has speci…c characteristics that
are correlated with unemployment and crime rates (think Detroit...).

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 27 / 32

Panel data regression

This can formally be modelled by

crimeit = β1 unemit + β0 y 87it + ai + uit ,

| {z }
vit

where y 87it is a year dummy (= 1 if t = 2) and

ai : unobserved city-speci…c characteristics that are constant over time
(e.g., location). Lots of city speci…c variables don’t change much over
time.
uit : Other unobserved factors that vary over time.
If ai is correlated with unem, our OLS estimator of β1 is biased.
So what should we do? Use time-variation in our sample to control
for ai !

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 28 / 32

First di¤erences speci…cation

The model on the previous slide can be represented in terms of the 2

cross sections:
crimei 1 = β1 unemi 1 + ai + ui 1
| {z }
vi 1

crimei 2 = β0 + β1 unemi 2 + ai + ui 2
| {z }
vi 2

Subtract the …rst from the second regression,

∆crimei = β0 + β1 ∆unempi + ∆ui

where ∆crimei = crimei 87 crimei 82 ∆unempi = unemi 2 unemi 1 ,

and ∆ui = ui 2 ui 1 .
First-di¤erencing eliminates ai
Omitted variables bias problem solved if E [∆uit j∆unemit ] = 0

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 29 / 32

First-di¤erence regression of crime on unemployment

∆crimei = β0 + β1 ∆unempi + ∆ui

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 30 / 32

Crime change (per 1000 people) versus unemployment
change, actual and predicted values

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 31 / 32

What’s next?

The …rst di¤erence estimator is a special (T = 2) case of the

so-called …xed e¤ects estimator
ai are the …xed e¤ects – called so because they are …xed over time
With …rst di¤erence models, we can only estimate the e¤ect of
time-varying variables.

D Kristensen (UCL) ECON0019: Lecture 9 December 6, 2018 32 / 32