Panel Data Models

Ani Katchova

2013 by Ani Katchova. All rights reserved.

Panel Data Models Overview

Panel data characteristics, panel data types

Variation types (overall, within, and between variation)
Panel data models (pooled model, fixed effects model, and random effects model)
Estimator properties (consistency and efficiency)
Estimators (pooled OLS, between, fixed effects, first differences, random effects)
Tests for choosing between models (Breusch-Pagan LM test, Hausman test)

Panel Data Models

Panel data model examples

Labor economics: effect of education on income, with data across time and individuals.
Economics: effects of income on savings, with data across years and countries.
Panel data characteristics
Panel data provide information on individual behavior, both across individuals and over time
they have both cross-sectional and time-series dimensions.
Panel data include N individuals observed at T regular time periods.
for all
Panel data can be balanced when all individuals are observed in all time periods (
).
i) or unbalanced when individuals are not observed in all time periods (
We assume correlation (clustering) over time for a given individual, with independence over
individuals.
o Example: the income for the same individual is correlated over time but it is independent
across individuals.

Panel data types

Short panel: many individuals and few time periods (we use this case in class)
Long panel: many time periods and few individuals
Both: many time periods and many individuals
Regressors
Varying regressors .
o annual income for a person, annual consumption of a product
Time-invariant regressors
for all t.
o gender, race, education
for all i.
Individual-invariant regressors
o time trend, economy trends such as unemployment rate

Variation for the dependent variable and regressors

Overall variation: variation over time and individuals.
Between variation: variation between individuals.
Within variation: variation within individuals (over time).
Id Time Variable Individual Overall
Overall
Between
mean
mean
deviation
deviation
i
1
1
1
2
2
2
3
3
3

t
1
2
3
1
2
3
1
2
3

9
10
11
20
20
20
25
30
35

10
10
10
20
20
20
30
30
30

20
20
20
20
20
20
20
20
20

-11
-10
-9
0
0
0
5
10
15

Within
deviation

-10
-10
-10
0
0
0
10
10
10

-1
0
1
0
0
0
-5
0
5

Within
deviation
(modified)

19
20
21
20
20
20
15
20
25

Individual mean

Overall mean

Overall variance

Between variance

Within variance

The overall variation can be decomposed into between variation and within variation.

 Time-invariant regressors (race, gender, education) have zero within variation.

Individual-invariant regressors (time, economy trends) have zero between variation.
We need to check the data to see if the between or within variation is larger for each variable.

Panel data models

Panel data models describe the individual behavior both across time and across individuals.
There are three types of models: the pooled model, the fixed effects model, and the random
effects model.
Pooled model
The pooled model specifies constant coefficients, the usual assumptions for cross-sectional
analysis.

This is the most restrictive panel data model and is not used much in the literature.

Individual-specific effects model

We assume that there is unobserved heterogeneity across individuals captured by .
o Example: unobserved ability of an individual that affects wages.
The main question is whether the individual-specific effects are correlated with the
regressors. If they are correlated, we have the fixed effects model. If they are not correlated,
we have the random effects model.
Fixed effects model (FE)
The FE model allows the individual-specific effects to be correlated with the regressors x.
We include as intercepts.
Each individual has a different intercept term and the same slope parameters.

We can recover the individual specific effects after estimation as:

In other words, the individual-specific effects are the leftover variation in the dependent
variable that cannot be explained by the regressors.
Time dummies can be included in the regressors x.

Random effects model (RE)

The RE model assumes that the individual-specific effects are distributed independently of
the regressors.
We include in the error term.
.
Each individual has the same slope parameters and a composite error term

and
,
Here
,
/
so
Rho is the interclass correlation of the error. Rho is the fraction of the variance in the error due
to the individual-specific effects. It approaches 1 if the individual effects dominate the
idiosyncratic error.

Panel data estimators

The panel data models can be estimated with several estimators.

The estimators differ based on whether they consider the between or within variation in the
data.
Their properties (consistency) differ based on which model is appropriate.
Estimator properties
We prefer estimators that are consistent and efficient. We check for consistency first and then
for efficiency.
Consistency
The distribution of

collapses on

as n becomes large:

Consistency is established based on the law of large numbers.

If an estimator is consistent, more observations will tend to provide more precise and accurate
estimates.

Efficiency
Efficiency (minimum variance) is usually established relative to specific classes of estimators.
o Example: OLS is efficient (minimum variance) among the class of linear, unbiased
estimators (Gauss-Markov Theorem).
o Maximum likelihood (given correct distributional assumptions) is asymptotically
efficient among consistent estimators.
Pooled OLS estimator
The pooled OLS estimator uses both the between and within variation to estimate the
parameters.
The pooled OLS estimator is obtained by stacking the data over i and t into one long regression
with NT observations and estimating it by OLS:

If the true model is the pooled model and the regressors are uncorrelated with the error terms,
the pooled OLS regressor is consistent.
If the true model is fixed effects then the pooled OLS regressor is inconsistent.
We need to have panel-corrected standard errors.

Between estimator
The between estimator only uses the between variation (across individuals).
It uses the time averages of all variables.
o If an individual has a work experience of 9, 10, and 11 years measured over 3 periods
then the average experience is 10.
This is an OLS estimation of the time-averaged dependent variable on the time-averaged
regressors for each individual.

The number of observations is N. The time variation is not considered and the data are
collapsed with one observation per individual.
This estimator is seldom used because the pooled and RE estimators are more efficient.
Within estimator or fixed effects estimator
The within estimator uses the within variation (over time).
It uses time-demeaned variables (the individual-specific deviations of variables from their
time-averaged values).

o If an individual has a work experience of 9, 10, an 11 years measured over 3 periods, the
average experience is 10. So the time-demeaned values are -1, 0, and 1.
This is an OLS estimation of the time-demeaned dependent variable on the time-demeaned
regressors.

Some software packages estimate:

The number of observations is NT.

The individual-specific effects cancel out.
Here, is the average of the individual effects.
A limitation of the within estimator is that time-invariant variables are dropped from the model
and their coefficients are not identified.
o A female/male will have values of 1/0 for the female dummy variable, so the values
minus the mean values (calculated over time) for each individual will be zero.
o If we are interested in the effects of time-invariant variables, we need to consider
different models (OLS or between estimators).

First-differences estimator
The first-difference estimator uses the one-period changes for each individual.
It uses first-differenced variables (the individual-specific one-period changes for each
individual).
o If an individual has a work experience of 9, 10, and 11 years measured over 3 periods
then the first difference experience are missing (.), 1, and 1.
This is an OLS estimation of the one-period changes of the dependent variable on the oneperiod changes in the regressors.
,

The number of observations is N(T-1). We lose the first observation for each individual
because of differencing.
The individual-specific effects cancel out.
A limitation of the first-differences model is that time-invariant variables are dropped from the
model and their coefficients are not identified.

Random effects estimator

This is an OLS estimation of the transformed model:
1

The number of observations is NT.

The individual-specific effects are in the error term.
Note that
0 corresponds to pooled OLS and
1 corresponds to the within (fixed effects)
estimator.
The random effects estimates are a weighted average of the between and within estimates.
The random effects estimator is fully efficient under the random effects model.

Models and estimators

Estimator/true model
Pooled OLS estimator
Between estimator
Within or fixed effects estimator
First differences estimator
Random effects estimator

Pooled model
Consistent
Consistent
Consistent
Consistent
Consistent

Random effects model

Consistent
Consistent
Consistent
Consistent
Consistent

Fixed effects model

Inconsistent
Inconsistent
Consistent
Consistent
Inconsistent

The fixed effects estimator will always give consistent estimates, but they may not be the most
efficient.
The random effects estimator is inconsistent if the appropriate model is the fixed effects model.
The random effects estimator is consistent and most efficient if the appropriate model is
random effects model.

Choosing between fixed and random effects

Breusch-Pagan Lagrange Multiplier test

This is a test for the random effects model based on the OLS residual.
or equivalently
,
is significantly different from zero.
Test whether
If the LM test is significant, use the random effects model instead of the OLS model.
We still need to test for fixed versus random effects.

Hausman test
The random effects estimator is more efficient so we need to use it if the Hausman test
supports it. If it does not support it, use the fixed effects model.
Hausman test tests whether there is a significant difference between the fixed and random
effects estimators.
The Hausman test statistic can be calculated only for the time-varying regressors.
The Hausman test statistics is:

o It is chi-square distributed with degrees of freedom equal to the number of parameters for
the time-varying regressors.
o If the Hausman test is insignificant use the random effects.
o If the Hausman test is significant use the fixed effects.