Simultaneous

Equation Models-II

BLOCK 3
PANEL DATA MODELS

147
Advanced Topics
in Regression INTRODUCTION TO BLOCK 3
Analysis

Block 3 is on Panel Data Models. It has two units (Units 8 and 9). Unit 8 is
on Introduction to Panel Data. It first outlines the distinction between pooled
cross section data and panel data. Advantages of panel data over pooled data
is then explained. The unit discusses Chow Test used in situations of Linear
Static Panel Data Models. A discussion on differentiating between ‘fixed
effect models’ and ‘random effect models’, with a reference to Policy
Relevant Inference, is then presented in the unit.

Unit 9 is on Estimation of Panel Data Models. Apart from explaining the

estimation methods, the unit discusses a model selection test, for deciding
between the fixed effects and random effects models, viz. the Hausman Test.

148
UNIT 8 INTRODUCTION TO PANEL DATA* Introduction to
Panel Data

Structure

8.0 Objectives
8.1 Introduction
8.2 Panel Data Models
8.2.1 Pooled Cross Section Data
8.2.2 Panel Data
8.2.3 Advantage of Panel Data Over Pooled Data

8.3 Linear Static Panel Data Model

8.3.1 Chow Test

8.4 Fixed Effect Versus Random Effect Panel Models

8.4.1 Fixed Effect Model
8.4.2 Random Effect Model
8.4.3 Policy Relevant Inference

8.5 Let Us Sum Up

8.6 Key Words
8.7 Suggested Books for Further Reading
8.8 Answers/Hints to Check Your Progress Exercises

8.0 OBJECTIVES
After reading this unit, you will be able to:

• differentiate between panel data and pooled data;

• state the distinctive features of time series data, cross-section data and panel
data;
• explain, with illustrations, the panel data models;
• indicate the problem of ‘endogeneity’ in a ‘static panel data model’;
• distinguish between the concepts of ‘self-selection’ and ‘endogeneity’;
• illustrate the framework of a ‘pooled cross section data’;
• elucidate the features of ‘panel data’;
• contrast the advantages of ‘panel data’ over ‘pooled data’;
• define the terms ‘Static Panel Data Model’ and ‘Dynamic Panel Data Model’;
• specify the steps involved in performing a Chow Test;
• discuss the features of ‘fixed effect (FE) model’ and ‘random effect (RE)
model’ in panel regressions;

*
Dr. Poulomi Roy, Jadavpur University. 149
Panel Data • present a comparative profile of the relative contexts in which the FE or the RE
Models
model is appropriate to adopt; and
• outline, with illustration,’ the ‘policy relevant inference’ that could be drawn
from panel data models.

8.1 INTRODUCTION
In applications, econometricians often use either pure cross sectional data or
time series data. A cross sectional data is one which is collected for different
sample units for a same point of time (e.g. NSSO’s 5-yearly data on
manufacturing firms). A time series data, on the other hand, is collected over
different time points for the same set of sample units (e.g. GDP). Such
sample units, in a pure time series data, could themselves be cross sectional
in nature.

A data set that has both cross sectional and time series dimensions are
nowadays very common in empirical research. Such data sets, often used for
policy analysis, could be pooled to form a panel data set. Note that an
independently pooled cross section can be obtained by a random sampling
of a large population at different points of time (usually, but not necessarily,
for different years). From a statistical standpoint, such data sets have an
important feature i.e. they consist of independent sampled observations. Such
independently sampled observations play a key role in our analysis of cross-
sectional data, where, among other things, it rules out correlation in the error
terms across different observations. An independently pooled cross section
data differs from a single random sample. This is in the sense that sampling
from the population at different points of time likely leads to observations
that are not identically distributed. For instance, distributions of wages and
education have changed over time in most countries.

A panel data, or longitudinal data set, thus consists of time series data for
each cross sectional unit in the data set. Panel data is collected on the same
individuals or firms or geographical units over specified periods of time. The
key difference between the panel data and pooled data is that, in case of panel
data, the same cross sectional units are followed over a given time period. In
case of pooled data, different cross section units are observed for a given time
period. Thus, the main features of the three types of data are:

a) Time Series Data: Many observations (large T) on as few as one unit

(small N ): e.g. stock price trends, aggregate national statistics.
b) Pooled Cross Sections: Two or more independent samples of many units
(large N ) are drawn from the same population at different time periods.
c) Panel Data: Observations of multiple phenomena obtained over multiple
time periods (small T ) on many cross section units (large N ).
Hence, for using methods of panel data analysis, we need information on
same cross section units over a given period of time. This is more
150
 difficult to obtain than a pooled cross section data where the sampled Introduction to
Panel Data
units could be different. But in observing the same cross sectional units
over time repeatedly, it helps us in controlling for certain ‘unobserved
characteristics’ of the cross sectional unit.

8.2 PANEL DATA MODELS

Let us now illustrate panel data models with some instances. Suppose that the
population consists of all manufacturing firms in a country operating during a
given three year period. Production function describing the output in the
population of firm can be specified as:
Log ( output it ) = δ t + β 1 log( lobour it ) + β 2 log( capital it ) + β 3 spillover it + quality i + u it
(8.1)

Here, spillover is a measure of foreign firm concentration in a region

containing the firm. The term quality refers to unobserved factors (e.g.
managerial or work quality) affecting productivity. We must note that quality
is a firm specific term and is constant over time. It also has the same effect in
each time period, while uit changes across time and firm. Each firm is
randomly chosen from the population of all manufacturing firms. Thus, in a
panel regression, for a specification like in (8.1), ‘i’ is an indicator of cross
section unit and ‘t’ is an indicator of time. In analysing a panel data set, our
aim is to capture this time constant for firms as a specific unobserved effect.
The error term ‘u’ represents the unobserved shocks in each time period. The
presence of the parameter δ t represents intercepts in each time period,
allowing for aggregate productivity to change over time. The coefficients of
regressors are assumed to be constant.

Another context we can consider relates to the effect of a training programme

on employees (or any other similar programme like providing mid day meals
to children in school) in their subsequent performance. Its specification can
be considered as:
log( performanceit ) = θt + zit γ + δ1 prog it + ci + uit (8.2)
where ‘i’ indicates individual, ‘t’ indicates time period and θ t indicates the
time varying intercept. z it is the set of observable characteristics that affect
not only wage but may also be correlated with program participation. ci
indicates the ability of the individual. Now, suppose at t =1 no one has
participated in the programme. It implies progi1 = 0 for all i. Then, let us say
some individuals are chosen to participate in the programme and their
subsequent performance are observed for the two groups (i.e. the group
which did not undergo training and the group which underwent the training).
The sub-group that participates in the training programme is defined as the
‘treatment group’ and the other one as the ‘control group’. In period t = 1
none received treatment but in t = 2 treatment group received training but the
control group did not receive the training. The term ci included in (8.2) stands
151
Panel Data for an individual ‘i’ who can choose to participate in the programme with
Models
his/her own choice i.e. it can be correlated with the inherent ability (or
proactive initiative) of the individual. This is identified in the literature as the
problem of ‘self-selection’. The important issue in a panel model like (8.2) is
whether unobserved factors of productivity relevance are correlated with the
observable factors? Another issue is whether we can assume at any time point
t, that the unobserved effect is uncorrelated with the error term of other time
periods or not? For instance, the effect of job training on productivity and on
subsequent wages. This is known as the problem of ‘endogeneity’. In the
above example we can see how the self-selection problem can lead to the
problem of endogeneity.

8.2.1 Pooled Cross Section Data

Pooled cross sectional data are obtained by collecting random samples from a
large population independently of each other at different point of time. Panel
data sets have both cross-sectional and time series features (it consists of time
series data for each statistical unit in the cross section). For instance, consider
two cross-sectional household surveys taken: one in 1985 and one in 1990. In
1985, a random sample of households were surveyed with variables like
income, savings, family size, etc. In 1990, a new random sample of
households was taken using the same survey questions. To increase our
sample size, we can form a pooled cross section by combining the two years.
Pooling cross sections from different years is an effective way for analysing
the effects of a new government policy. The idea is to collect data from the
years before and after a key policy change. As an example, we can consider
the data on housing prices taken in 1993 and 1995 i.e. before and after a
reduction in property taxes was effected in 1994. Suppose we have data on
250 houses for 1993 and on 270 houses for 1995. One method of arranging
such a data set is as given in Table 8.1. Observations 1 through 250
correspond to the houses sold in 1993, and observations 251 through 520
correspond to the 270 houses sold in 1995. A pooled cross section is analysed
much like a standard cross section, except that we often need to account for
secular differences in the variables across time. In fact, in addition to
increasing the sample size, the point of a pooled cross-sectional analysis is
often to see how a key relationship has changed over time. With large N and
small T one may introduce separate intercepts for each time period.

Table 8.1: Pooled Data on Houses Sold

obsno year hprice proptax sqrft bdrms bthrms

1 1993 85500 42 1600 3 2.0
2 1993 67300 36 1440 3 2.5
3 1993 134000 38 2000 4 2.5
- - - - - - -
- - - - - - -
152
 - - - - - - - Introduction to
Panel Data
250 1993 243600 41 2600 4 3.0
251 1995 65000 16 1250 2 1.0
252 1995 182400 20 2200 4 2.0
253 1995 97500 15 1540 3 2.0
- - - - - - -
- - - - - - -
- - - - - - -
520 1995 57200 16 1100 2 1.5

Source: Woolridge (5th Edition)

In India, many surveys of individuals, households and firms are repeated in

the NSSO’s [National Sample Survey organisation (NSSO)] periodic surveys
conducted on individuals and households at regular intervals. For these
surveys, NSSO randomly samples households at every five year interval. If a
random sample is drawn at each time period, pooling the resulting random
samples gives us an independently pooled cross section. One reason for using
independently pooled cross sections is to increase the sample size.

8.2.2 Panel Data

The unique characteristic of panel data structure is that each cross section
unit is followed over a certain period of time. Panel data sets are fairly easy
to collect for districts, cities, states, and countries. Hence, policy analysis is
greatly enhanced by using panel data sets. For the econometric analysis of
panel data, we cannot assume that the observations are independently
distributed across time. For instance, unobserved factors (such as ability) that
affect someone’s wage in 2010 will also affect that person’s wage in 2011.
Likewise, unobserved factors that affect a city’s crime rate in 2015 will also
affect that city’s crime rate in 2020. For this reason, special models and
methods have been developed to analyse panel data.

In using panel data in an econometric study, it is important to know how the

data should be stored. We must be careful to arrange the data so that the
different time periods for the same cross-sectional unit (person, firm, city,
and so on) are easily linked. For instance, let us suppose that the data set is on
cities for two different years. For most purposes, the best way to enter the
data is to have two records for each city, one for each year. The first record
for each city corresponds to the early year, and the second record is for the
later year. These two records should be adjacent. Therefore, a data set for 100
cities and two years will contain 200 records for each of the variables. The
first two records are for the first city in the sample, the next two records are
for the second city, and so on.

The above method of data arrangement makes it easy to obtain the

differences in the two records for each city and store them in a pooled cross-
153
Panel Data sectional manner for an analysis of the differencing estimation. Most of the
Models
two-period panel data sets are stored in this way. We use a direct extension of
this scheme for panel data sets with more than two time periods. A second
way of organising the two periods of a panel data set is to have only one
record per cross-sectional unit. This requires two entries for each variable,
one for each time period. Creating the differences from T1 to T2 is then easy.
Placing the data in one record, however, does not allow for a pooled analysis
by using the two time periods on the original data. Also, this method of
organisation does not work for panel data sets with more than two time
periods. Table 8.2 presents a two-year panel data set on crime and related
statistics for 150 cities. Cities are numbered as 1,2,…,150. Just as in a pure
cross section, the ordering in the cross section of a panel data set does not
matter. We could use the city name in place of a number. But it is often
useful to have both.

Table 8.2: Panel Data on Crime and Unemployment by City

obsno City Year Murders Population Unem Police

1 1 1986 5 350000 8.7 440
2 1 1990 8 359200 7.2 471
3 2 1986 2 64300 5.4 75
4 2 1990 1 65100 5.5 75
- - - - - - -
- - - - - - -
- - - - - - -
297 149 1986 10 260700 9.6 286
298 149 1990 6 245000 9.8 334
299 150 1986 25 543000 4.3 520
300 150 1990 32 546200 5.2 493

Source: Woolridge (5th Edition)

8.2.3 Advantages of Panel Data Over Pooled Data

Because panel data require replication of the same units over time, panel data
sets, especially those on individuals, households, and firms, are more difficult
to obtain than pooled cross sections. Not surprisingly, observing the same
units over time leads to several advantages over cross-sectional data or even
pooled cross-sectional data. The benefit that we will focus on is of having
multiple observations on the same units which allows us to control for certain
unobserved characteristics of individuals, firms, etc. As we will see, the use
of more than one observation can facilitate causal inference in situations
where inferring causality would be difficult if only a single cross section
were available. A second advantage of panel data is that it allows us to study
the importance of lags in the behaviour or the result of decision making. This
154 information can be significant because many economic policies can be
expected to have an impact only after some time has passed. It therefore Introduction to
Panel Data
follows from here that the advantage of panel data is that we can observe the
‘before and after effects’ of receiving a treatment by the same individual. It
also provides the possibility of isolating the effects of treatment from other
factors affecting the outcome.

Panel data obtained by combining both the cross sectional and time series
data capture both the inter-cross-sectional differences as well as the intra
cross sectional dynamics. It has several other advantages over cross sectional
and time series data. For instance, cross sectional data may be viewed as a
panel with T =1 and time series data may be viewed as a cross section with
N = 1. Hence, panel data combining both cross section and time series data
provides more degrees of freedom and more sample variability than either
only the cross sectional or only the time series data. It hence improves the
efficiency of econometric estimates.

Evaluating the effectiveness of certain programmes by using a cross-sectional

sample typically suffers from the fact that those receiving treatment are
different from those without. In other words, one does not simultaneously
observe what happens to an individual when she receives the treatment or
when she does not. An individual is observed as either receiving treatment or
not receiving treatment. Using the difference between the treatment group
and control group could suffer from two sources of biases: (i) selection bias
due to differences in observable factors between the treatment and control
groups and (ii) selection bias due to endogeneity of participation in treatment.

It is also frequently argued that the real reason one finds (or does not find)
certain effects is ‘due to ignoring the effects of certain variables in a model
specification which are correlated with the included explanatory variables’.
Panel data contain information on both the inter-temporal dynamics and the
individuality of the entities. This therefore allows for one to control for the
effects of missing or unobserved variables.

By pooling random samples drawn from the same population, but at different
points in time, we can get more precise estimators and test statistics with
higher power. Pooling is helpful in this regard only in-so-far as the
relationship between the dependent variable and at least some of the
independent variables remain constant over time. Using pooled cross sections
raises a statistical complication viz. the two populations could have different
distributions. To reflect for the fact that the populations may have different
distributions in different time periods, we allow the intercept to differ across
periods. This is also accomplished by including dummy variables for all but
one year i.e. for the earliest year in the sample which is usually chosen as the
base year. Sometimes, the pattern of coefficients on the year dummy
variables could itself be of interest.

155
Panel Data Check Your Progress 1 [answer within the space given in about 50-100
Models
words]

1) Distinguish between time-series data and cross-section data.

……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
2) Differentiate between pooled data and panel data.
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
3) State the two main advantages of ‘panel data’ over ‘pooled data’.
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
4) What is a ‘statistical complication’ in pooling two cross sections? How is
this complication dealt with in practice?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………

8.3 LINEAR STATIC PANEL DATA MODEL

Suppose for each cross section unit we collect data on same set of variables
for T time periods. Let X be a vector of k exogenous variables which affect Y.
At any time point ‘t’, the population model is like:
Yit = X it β + c i + u it , t = 1,2,….,T, I = 1,2,…,N (8.3)
156
where ci is the unobserved effect and uit is the random error term. (8.3) is a Introduction to
Panel Data
panel regression with ‘i’ as an indicator of cross section unit and ‘t’ as an
indicator of time. The most commonly used method for estimating the
parameters β is the ‘ordinary least square’ (OLS). The OLS assumes that the
explanatory variables are exogenous in nature and they are uncorrelated with
the random error term. Primary motivation behind the panel data is to solve
for the omitted variables problem. In panel data models, we consider time to
account for the unobserved effect (like quality in the example considered
above). We assume that the ‘unobserved effects’ are random variables. This
is an instance of a linear static panel data model. It is a static model because
all explanatory variables are contemporaneous dates corresponding to the
value of Y in period t. In contrast, in a dynamic panel data model, one or
more lagged dependent variables are allowed in the models as a ‘partial
adjustment mechanism’. In this unit, we discuss only the static panel data
models. You may however note that a dynamic panel model, with one lagged
dependent variable and a single regressor X, is defined as:

y it 1 y it 1 X it 2 ci it (8.4)

where ci is for a specific unobserved effect and ε it is the overall random

error term.

8.3.1 Chow Test

Chow Test examines whether parameters of one group of data are equal to
those in the other groups. Simply put, the test checks whether the data can be
pooled. If only intercepts are found different across groups, it becomes a
‘fixed effect model’. Let us consider two groups from a model like y = α + βx
+ ε as follows:

y = α1 + β1 x + ε1 for n1 observations (group 1) (8.5)

y = α2 + β2 x + ε2 for n2 observations (group 2) (8.6)

The null hypothesis is α1 = α2 and β1 = β2. If the null hypothesis is rejected, it

means the two groups have different slopes and intercepts and hence the data
is not poolable. The Chow test is simply an F test used to determine whether
a multiple regression function differs across two groups. We can also apply
this test to two different time periods. The ‘sum of squared residuals’ (SSR)
obtained from the pooled estimation for both the groups combined is
designated as the ‘restricted SSR’ (SSRr). The unrestricted SSR is the sum of
the two SSRs obtained for the two groups separately. A Chow test can also be
computed for more than two time periods. In such cases, we first estimate the
restricted model by doing a pooled regression and obtain the SSRr (i.e. the
restricted SSR). We then run separate regressions (for each of the time
periods, say, T) to obtain the sum of squared residuals for each time period.
The unrestricted sum of squared residuals is then obtained as
SSRUR = SSR1 + SSR2 + ........... + SSRT . If there are k explanatory variables
157
Panel Data (excluding the intercept or the time dummies) with T time periods, then we
Models
are imposing (T –1)k restrictions for the (T +Tk) parameters estimated in the
unrestricted models. Hence, if n = n1 + n2 + ..... + nT is the total number of
observations, then the ‘degrees of freedom’ (df) for the F test are ‘(T –1)k and
(n –T –Tk)’. We compute the F statistic as usual i.e.:

(SSRr –SSRur) (T –1)k / SSRur (n –T –Tk)] (8.7)

You may simply note at this stage that, as with any F test based on sums of
squared residuals, this test is not robust to Heteroscedasticity.

8.4 FIXED EFFECT VERSUS RANDOM EFFECT

PANEL MODELS
With panel data, the most commonly estimated models are the fixed effects
and the random effects models. Let us therefore focus first on the major
differences between these two types of models. Several considerations affect
the choice between the two types of models. For this, first of all, one has to
identify the nature of the variables that have been omitted from the model. If
we have reason to believe that there are no omitted variables, or we believe
that the omitted variables are uncorrelated with the explanatory variables in
the model, then a ‘random effects’ (RE) model is probably the best. It will
produce unbiased estimates of the coefficients, use all the data available and
produce the smallest standard errors. On the other hand, if there are omitted
variables, and these variables are correlated with the explanatory variables in
the model, then ‘fixed effect’ (FE) models provide a means for controlling
the ‘omitted variable bias’. In a fixed-effects model, ‘subjects’ serve as their
own controls. The idea is that whatever effects the omitted variables have on
the subject at one time, they will also have the same effect at a later time. In
this sense, their effect will be ‘constant’ or ‘fixed’. However, for this to be
true, the omitted variables must have time-invariant values with time-
invariant effects. By time-invariant values, we mean that the value of the
variable does not change across time. Gender and race are obvious instances,
but this can also include the ‘educational level’ of the respondent.

Second, one needs to consider the variability within subjects or cross section
of units. If subjects change little across time, a fixed effects model may not
work very well. This is because, there needs to be within-subject variability if
we are to use subjects as their own controls. If there is little variability within
subjects, then the standard errors from fixed effects model could be too large.
Conversely, random effects models will often have smaller standard errors.
But, the trade-off is that their coefficients are more likely to be biased.

Third, one needs to decide whether one wants to estimate the effect of
variables whose values do not change across time. With fixed effects models,
we do not estimate the effect of variables whose values do not change across
time. Rather, we control for them or ‘partial them out’. This is similar to an
158 experiment with random assignment. Though the RE models estimate the
effect of time-invariant variables, the estimates could be biased because we Introduction to
Panel Data
are not controlling for omitted variables. For a more clearer description, let us
consider a situation where y and x x1 , x2 ,.............., xk are observable
random variables with a linear relationship like:
y = α + xβ + c (8.8)

where ‘c’ the unobservable random variable. We are interested in the partial
effect of the observable explanatory variables xj while holding ‘c’ constant.
Our interest is to estimate the vector β . If ‘c’ is uncorrelated with x, then ‘c’
is just another unobserved factor uncorrelated with the explanatory variables.
If cov x j , c 0 for some j, then we cannot consistently estimate β .

8.4.1 Fixed Effect Model

Fixed effect (FE) are thus variables that are ‘constant across individuals’.
These variables are like age, sex, ethnicity which do not change (or change at
a constant rate) over time. FE explores the relationship between the predictor
variables (i.e. explanatory or independent variables) and outcome variables
(i.e. the dependent variable). The relationship between them is explored
within an entity (country, person, company, etc.). Each entity has its own
individual characteristics that may or may not influence the predictor
variables. For instance, being a male or female could influence the opinion
toward certain issue, the political system of a particular country could have
some effect on trade or GDP, the business practices of a company may
influence its stock price, etc. When using FE, we assume that something
within the individual may impact, or bias, the predictor and therefore we
might wish to control for this. This is the rationale behind the assumption of
the correlation between entity’s error term and predictor variables. FE
removes the effect of those time-invariant characteristics so that we can
assess the net effect of the predictors on the outcome variable. Another
important assumption of the FE model is that the time-invariant
characteristics are unique to the individual and are not correlated with other
individuals’ characteristics. In other words, each entity is different and
therefore the entity’s error term and the constant (which captures individual
characteristics) are not correlated with the others. If the error terms are
correlated, then the FE model is not suitable. In that case, we need to model
that relationship using the RE model.

The FE model allows the unobserved individual effects to be correlated with

the included variables. We can therefore model the differences between units
as parametric shifts of the regression function. This could be viewed as
applying only to the cross-sectional units in the study and not for the
additional units outside the sample. For instance, an inter-country comparison
may include the full set of countries for which it is reasonable to assume that
the model is constant. If the individual effects are strictly uncorrelated with
the regressors, then it might be appropriate to model the individual specific
constant terms as randomly distributed across the cross-sectional units.

159
Panel Data
Models
8.4.2 Random Effect Model
The random effects (RE) model is useful when we have reason to believe that
the unobserved effect is uncorrelated with all the explanatory variables. In
such a situation, the time constant’s unobserved effect is uncorrelated with
the explanatory variables and the parameters could be consistently estimated
by using a single cross section. There is therefore no need for panel data. But
using a single cross section disregards much useful information in the other
time periods. We can therefore use the data in a pooled OLS procedure i.e.
just run the OLS of dependent variable on the explanatory variables with the
time dummies. This, too, produces consistent estimators of the parameters
under the RE assumption. But it ignores the fact that the existence of
unobserved effect in the error term in each time period is serially correlated
across time. We can use the GLS method to solve for the serial correlation
problem.

RE assumes that the unobserved effect is uncorrelated with all explanatory

variables irrespective of whether the explanatory variables are fixed over
time or not. Hence, we can include a variable like education even if it does
not change over time. But we are assuming that education is uncorrelated
with the unobserved effect. Hence, in applications of FE and RE, it is usually
informative to compute the pooled OLS estimates. Comparing the three sets
of estimates can help us determine the nature of the biases caused by leaving
the unobserved effect. We must, however, remember that, even if the
unobserved effect is uncorrelated with all explanatory variables in all time
periods, the pooled OLS standard errors and test statistics are generally
invalid. This is because they generally ignore the often substantial serial
correlation in the composite errors. But it is possible to compute the standard
errors and test statistic which are robust to arbitrary serial correlation (and
Heteroscedasticity) in composite error. Note that the FE approach allows for
the arbitrary correlation while the RE approach does not. Hence, the FE
approach is widely thought to be a more convincing tool for estimating the
‘ceteris paribus’ effects.

To sum up, therefore, if the key explanatory variable is constant over time,
we cannot use FE to estimate its effect on dependent variable. In such
situations, we must rely on the RE (or pooled OLS) estimate. We can
however use the RE approach if we are able to assume that the unobserved
effect is uncorrelated with the explanatory variables. Typically, when one
uses random effects, many time-constant controls are included among the
explanatory variables. However, with the FE approach, it is not necessary to
include such controls. RE is preferred to pooled OLS due to its generally
higher efficiency.

8.4.3 Policy Relevant Inference

The choice of fixed or random effects should be based on the basis of the
160 background knowledge and the availability of data. Let us have clarity on
what we mean here by the term ‘policy-relevant inference’. Ideally, policy- Introduction to
Panel Data
relevant inferences are causal inferences about average treatment effects.
Causal inferences tell us what happens if we intervene and change the way
the things are being done. Within the regression modelling framework, and in
the absence of experimental or quasi-experimental data, many issues can be
overcome by making assumptions. But, estimating the treatment effect in an
unbiased manner becomes difficult. A realistic goal is therefore to produce
policy-relevant estimates that may be biased, but are not too much so, so as to
lead to misleading policy recommendations. Recall that the RE approach
requires the strong assumption that the unobserved effect is uncorrelated with
any of the covariates. An important reason why the random effect assumption
fails is that there is usually non-random selection of cross section units. For
instance, if each school had drawn its pupils at random from the pupil
population, then the random effect assumption would hold. But, in reality, a
non-random selection mechanism operates through which parents choose
schools and some schools select which children to accept. Thus, the
probability of selecting a particular school varies systematically according to
a series of factors characterising the child, his/her family, the school itself or
the higher local education authority. Some of these factors will be associated
with pupil attainment, either directly or indirectly, through a mediating
mechanism.

In light of the above, the debate remains inconclusive on whether we should

conclude that the FE approach is always preferable? The answer depends on
circumstances. The FE estimator for β is robust when a school F is not empty.
Therefore, if we have some knowledge about the school selection
mechanism, and we can include measures of these factors in the model as
‘controls’, then we can estimate the average treatment effect using the RE
approach.

Check Your Progress 2 [answer within space given in about 50-100 words]

1) Distinguish between Linear ‘Static Panel Data Model’ and ‘Dynamic

Panel Data Model’.
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
2) For what purpose is the ‘Chow Test’ used? What does it basically seek to
examine?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
161
Panel Data 3) In what contexts, the ‘fixed effects’ or the ‘random effects’ panel data
Models
model used?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
4) Specify the considerations that determine the choice between the FE and
the RE models.
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………

8.5 LET US SUM UP

The unit introduces the panel data models. Panel data refers to observations
on multiple variables obtained over different time periods for the same firms
or individuals. It can be understood by the common expression ‘the two data
sets are drawn from the same panel’. In contrast, pooled cross section data
refers to a time series of cross-sections where the observation on each cross
section do not necessarily relate to the same units. In India, the surveys of
NSSO, conducted on many subjects periodically, usually at an interval of 5
years, are based on independent random samples. They are therefore useful
for methods of ‘pooled data’ analysis and techniques. The unit has introduced
you to the concepts and application of two main leading approaches viz. the
FE approach and the RE approach. The contexts in which the choice between
the two could be made is outlined. Generally, the choice need to be based on
the background knowledge on variables and the nature of availability of data.

8.6 KEY WORDS

Pooled : Refers to data collected on same cross sections at two

Cross different points of time but pooled for the purpose of
Section Data analysis. By combining the two samples, we get increased
degrees of freedom or higher ‘n’. Data is pooled to assess
the impact of a new government policy. In other words,
pooled cross section data helps us in assessing the
before/after effects.
Panel Data : Refers to data collected for two time points on same sample
162 units. In other words, unlike in ‘pooled cross section’, no
 new random samples are used in the two surveys. Data is Introduction to
Panel Data
collected on same variables. This is particularly useful for
assessing the effect of ‘lags’ which is usually there in govt.
policies introduced.

8.7 SUGGESTED BOOKS FOR FURTHER

READING
1) Wooldridge J M (2006). Introduction to Econometrics: A Modern
Approach. Michigan State University. USA.
2) Greene W H (2016). Econometrics. Prentice Hall.

8.8 ANSWERS/HINTS TO CHECK YOUR

PROGRESS EXERCISES
Check Your Progress 1

1) Time series is data is collected over different time points for the same set
of sample units (e.g. GDP for states). Cross section data is collected over
different sample units for a same point of time (e.g. NSSO’s surveys in
India on 5-yearly basis).
2) In case of panel data, the same cross sectional units are followed up over
different time periods. In case of pooled data, different cross section
units (i.e. two independently selected random samples) are observed for
a given time period.
3) One, it allows for causal inference. Second, it allows us to study the
effect of lags in the behaviour or the result of decision making.
4) The complication is that the two samples might have come from
populations with different distributions. The way it is dealt with is by
allowing for different intercept terms or by using a ‘dummy variable’.

Check Your Progress 2

1) It is like Yit = X it β + c i + u it , t = 1,2,….,T, I = 1,2,…,N where ci is for
a specific unobserved effect (like quality), assumed to be a random
variable. uit is the overall random error term, accounting for all other
unobserved factors or effects. When one or more lagged dependent
variables are allowed in the models, as a ‘partial adjustment mechanism’,
it becomes a dynamic panel data model.
2) It is used to test for the feature of ‘poolability’ across data collected in
groups. In other words, it seeks to examine whether the parameters in the
models for the two or more groups are equal.
3) In general, a panel data model is used for determining the effect of
‘omitted variables’. If we have reason to believe, no variable is omitted,
then ‘random effect model’ can be used. If it is not so, applying the
163
Panel Data ‘fixed effects panel data model’ helps in controlling for the ‘omitted
Models
variable bias’.
4) If the key explanatory variable is constant over time, then RE model is to
be applied. Alternatively, we can use the RE model when ‘we are able to
assume that the unobserved effect is uncorrelated with the explanatory
variables’. The FE approach allows for the arbitrary correlation while the
RE approach does not. Hence, the FE approach is a convincing tool for
estimating the ‘ceteris paribus’ effects. Therefore, The choice of fixed or
random effects should be based on the basis of the background
knowledge and the availability of data.

164
UNIT 9 ESTIMATION OF PANEL DATA Estimation of Panel
Data Models
MODELS*

Structure

9.0 Objectives
9.1 Introduction
9.2 Random Effect Estimation Method
9.3 Fixed Effect Estimation Method
9.4 Model Selection: Hausman Test
9.5 Simultaneous Equation Models in Panel Data
9.6 Let Us Sum Up
9.7 Key Words
9.8 Suggested Books for Further Reading
9.9 Answers/Hints to Check Your Progress Exercises

9.0 OBJECTIVES
After reading this unit, you will be able to:

• specify the form of a ‘panel data regression model’;

• state the assumptions behind the ‘random effect estimation models’;
• discuss the various features of the ‘method of random effect estimation’;
• list the assumptions behind the ‘fixed effect estimation models’;
• explain the features of ‘fixed effect estimation model’;
• describe the Hausman’s ‘model selection procedure’; and
• elucidate, with illustrations, the application of ‘simultaneous equation
models’ in panel data contexts.

9.1 INTRODUCTION
In a panel data set up, data is collected on each cross section unit ‘i’. It is
done over a period of time ‘T’. Therefore, if X is a vector of ‘k’ exogenous
variables affecting Y (the dependent variable), at any time point ‘t’, the
population model is of the form:
Yit = X it β + c i + u it , i = 1,2,…,N; t = 1,2,….,T (9.1)

where ci is the ‘unobserved effect’ and uit is the ‘random error term’. They
are also called as ‘idiosyncratic error terms’ because they change over ‘time
and cross sections units’ i.e. over i and t. Equation (9.1) is a panel regression
*
Dr. Poulomi Roy, Jadavpur University 165
Panel Data where ‘i’ indicates a cross section unit and ‘t’ is an indicator of time. The
Models
primary motivation behind a panel data analysis is solving for the ‘omitted
variables problem’. In panel data models, we consider ci as ‘time invariant’.
We further assume that the ‘unobserved effects’ are random variables. This is
in the case of a ‘linear static panel data model’. It is a ‘static model’ in the
sense that the explanatory variables are contemporaneous (fixed or pre-
determined) values corresponding to the value of Y in period t. In Equation
(9.1), to enable us to assume that the effect of ci is with a zero mean, we
explicitly add an intercept term. With this, Equation (9.1) takes the form:

Yit = β0 + Xit βi + ci + uit (9.2)

9.2 RANDOM EFFECT ESTIMATION METHOD

In the absence of any ‘omitted variable’, where the omitted variables are
assumed to be uncorrelated with the regressors, the random effect estimation
technique is preferred. The assumptions required for the estimation of the
parameters of the random effect model are the following:

i) For the ith cross section unit, at any time point ‘t’, the model is defined
as:

yit 0 1X it1 2 X it 2 ............ k X itk ci uit (9.3)

where ci is the ‘unobserved effect’ and uit is the ‘random error term’ (i =
1,2,……,N and t = 1,2,…..,T). βj’s are the parameters to be estimated.

ii) Each explanatory variable changes over time (at least for some i). No
perfect linear relationship exist among the explanatory variables.

iii) For each t, the ‘expected value of the idiosyncratic error terms’ given the
Xi’s [for all (∀) ‘t’ and ‘ci’] is ‘zero’ i.e. E ( u it | X i , c i ) = 0 .

iv) var (u it | X i , ci ) = var (u it ) = σ u2 , t = 1,2,…,T.

v) The covariance of the uncorrelated error terms, conditional to Xi’s and

ci’s ∀ t ≠ s is ‘zero’ i.e. Cov (u it , u is | X i , ci ) = 0

vi) ���� |�� ) = ���� ) = 0 i.e. the expected value of ci is constant.

2
vii) var ci | X i i i.e. the assumption of homoscedasticity hold for ci’s.

Under the above assumptions, the random effect model (9.3) can be written
as:

yit = xit β + vit (9.4)

where vit = ci + uit is the ‘composite error term’ such that E (vit | xti ) = 0 , t =
1,2,3,….,T. Note that the defining of the composite error term (as vit = ci +
166
uit) implies that the vit’s are ‘serially correlated across time’. Hence, Estimation of Panel
Data Models
�� �
����������� ���� , ��� ) = �� �� �,
t ≠ s.
� �

Therefore, the usual pooled OLS procedure cannot be applied. In such cases,
we have to apply the GLS method of estimation with ‘auto regressive serial
correlation’. Therefore, the model in (9.4) can be written as:

yi = xi β + vi (9.5)

where vi = ci jT + ui (jT being the TX1 vector of ones) and we define the
unconditional variance matrix vi as: Ω = E (vi , vi' ) . We assume Ω to be a
‘positive definite matrix’ which is constant for all cross section units due to
the assumption of random sampling. This means, for the consistency of GLS,
we need the ‘rank condition’ that Rank (E (X i'Ω −1 X i )) = k . Under this
assumption, E (ci , uit ) = 0 ∀ t=1,2,3,…,T and

( ) 2
E v 2 it = E (ci ) + 2 E (ci uit ) + E (u 2it ) = σ c2 + σ u2

where σ c2 = E (ci2 ) ∀ t ≠ sA.

We have to now derive the variance-covariance of the elements of v.

E (v it , v is )= E (c i + u it , c i + u is )= ( )
E c i2 = σ 2
c

Therefore:

σ c2 + σ u2 σ c2  σ c2 
 
 σ c2 σ c2 + σ u2  σ c2 
( '
Ω = E vi , vi =  ) 
     
 σ 2
σ c2  σ c2 + σ u2 
 c

Thus, Ω = (σ c2 iT jT + σ u2 I T ) iT jT where Ω depends on two parameters

σ c2 and σ u2 regardless of the size of T. For efficiency of feasible GLS, we
'
( '
)
assume that E v i , v i | x i = E ( v i , v i ) . Now the RE assumption
( )
E u i , u ' i | x i , c i = σ u2 I T implies
E (u it , u is | x i , c i ) = 0, t ≠ s ; t , s = 1, 2,3,..., T
2
σ 2 ˆ = (σˆ 2 I + σˆ 2 )i j '
Ω
With the estimated u and c used in Ω we get u T c T T.

Applying the GLS technique, the RE estimator is obtained as:

−1
N  N 
β̂ RE =  ∑ xi' Ω −1 xi   ∑ xi' Ω −1 yi  .
 i =1   i =1 

Now, given that: (i) E (uit | xi , ci ) = 0, (ii) E (ci | xi ) = E (ci ) = 0 and (iii)
ˆ
Rank (E (X i
'
Ω −1
X i )) = k , β RE → β as N → ∞. Deriving the GLS
transformation that eliminates the serial correlation in the errors, requires
167
Panel Data [
sophisticated matrix algebra. For this, let us define: θ = 1 − σ u2 /(σ u2 + Tσ c2 ) ]
Models
which lies between zero and one. Then, the transformed equation becomes:

y it − θ y i = ( x it − θ x i ) β + ( v it − θ v i ) (9.5a)

where the ‘over bar’ denotes the ‘time averages’. This is a very interesting
equation, as it involves ‘quasi-demeaned data’ on each variable. If the model
considered had been of ‘fixed effects’, then the estimator would have
subtracted the time averages from the corresponding variable. Since we are
here considering the random effects transformation, it subtracts a fraction of
that time average, the fraction depending on σ u2 , σ c2 and T. The GLS
estimator is thus simply the pooled OLS estimator of Equation (9.5a). The
transformation in (9.5a) allows for the explanatory variables that are constant
overtime. This is in fact the one advantage of random effects (RE) over ‘fixed
effects’ (or first differencing). This is possible because RE assumes that the
‘unobserved effect’ is uncorrelated with all explanatory variables irrespective
of whether the explanatory variables are fixed over time or not. In many
applications, the whole rationale for using panel data is to allow for the
unobserved effect to be correlated with the explanatory variables.

Check Your Progress 1 [answer within the space given in about 50-100
words]

1) Specify a ‘linear static panel data model’. Why is the model called
‘static’?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
2) Why is it that in the case of ‘linear static panel data models’, the usual
OLS method cannot be applied? Which method is to be applied in such
cases?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
168
3) State the assumptions required to be made for the ‘random effect panel Estimation of Panel
Data Models
models’.
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
4) In Equation (9.5a), what is meant by ‘quasi demeaned data’? What is its
significance?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………

9.3 FIXED EFFECT ESTIMATION METHOD

Let us again consider the ‘linear unobserved effects model’ [(9.2)] for T time
periods:
yit xit ci uit , t=1,2,….,T

The random effect approach to estimate β effectively puts �� into the error
term. This is done under the assumption that �� is orthogonal to xit. It then
accounts for the implied serial correlation in the composite error term:
v it ci u it using a GLS analysis. In other words, the FE analysis
is correlated with xit. The T equations in the model are like:
assumes that c i

y it = x it β + c it jT + u it
(9.6)

where jT is the TX1 vector of one. The assumptions required for the FE model
are the following:

i) For each i, the model is:

y it = β 1 x it 1 + β 2 x it 2 + ......... + β k x itk + c i + u it , t = 1,2,…,T

where the β j are the parameters to be estimated and ci is the

‘unobserved effect’.

ii) The cross section is a random sample.

iii) Each explanatory variable changes over time (for at least some i) with
‘no perfect linear relationships’ existing among the explanatory
variables. 169
Panel Data iv) For each t, the expected value of the idiosyncratic error term, given the
Models
explanatory variables in all time periods, and the unobserved effect, is
‘zero’. That is: E (u it | xi , ci ) = 0 .

[Under these assumptions {(i) to (iv)}, the FE estimator of βj is

consistent with a fixed T as N → ∞ ].

v) Var(uit | xi , ci ) = Var(uit ) = σ u2 ∀t = 1,2,……,T

vi) For all t ≠ s , the idiosyncratic errors, conditional on all explanatory

variables and unobserved error term ci, are uncorrelated.
[Under the assumptions {(i)-(vi)}, the FE estimator of β j is the ‘best
linear unbiased estimator’ (BLUE)].

vii) Conditional on xi and ci, the uit are ‘iid’ i.e. independent and identically
distributed as N (0, σ u2 ).

If we add the assumption (vii), then the FE estimator is normally distributed

with the t and F statistics having the exact t and F distributions. Without the
assumption (vii), we can rely on asymptotic approximations. In other words,
without making this special assumption (i.e. the viith one), the approximations
require large N and small T.

Now to estimate β of Equation (9.6) under the above assumptions, we

transform the equation to eliminate the ‘unobserved effect’ c i . We obtain
this FE transformation by first averaging Equation (9.6) as:

yi = xi β + ci + ui (9.7)

1 1 1
where yi y u , xi xu , ui uu ,
T T T

Now, subtracting (9.7) from (9.6) we get:

( yit yi ) ( xit xi ) (uit ui )

(9.8)

Or, yit = xit β + uit , t = 1,2,..,T (9.9)

yu
where,  yu yi , 
xit xit xi , uit uit ui

Note that in (9.8), the ‘time demeaning’ has removed the individual specific
effect ci. The ‘fixed effects transformation’ is also called as the ‘within
transformation’. The important thing about Equation (9.9) is that the
unobserved effect ci, is eliminated. This suggests that we should estimate
(9.9) by pooled OLS. Thus, our ‘fixed effects estimator’ (also called the
‘within estimator’), is a pooled OLS estimator, based on the time-demeaned
variables. Now we proceed to estimate (9.9) by the pooled OLS method as
follows. Since, E ( x'it , uit ) = 0 t = 1,2,….,T, under a strict exogeneity
assumption on the explanatory variables, the fixed effects estimator is
170
unbiased. Alternatively, at least roughly, the idiosyncratic error uit are Estimation of Panel
Data Models
uncorrelated with each explanatory variable across all time periods. Note that
the ‘fixed effects estimator’ allows for the arbitrary correlation between c i
and the explanatory variables in any time period. Because of this, any
explanatory variable that is constant over time for all i gets swept away by
the fixed effects transformation: xit = 0 for all i and t. In other words, we
cannot include variables like gender or a city’s distance from a river. Note
also that the FE estimator is the pooled OLS estimator from the regression of
yit on xit , t = 1,2,….,T; i = 1,2,…,N. In order to ensure that the FE
estimator is well behaved asymptotically, we need a standard rank condition
on the matrix of time demeaned explanatory variables viz.

 T
( 
)
Rank  ∑ E x' it , xit  = Rank E x' it , xit [ ( )] = k .
 t =1 

This implies that the ‘time constant variables’ are not allowed in FE analysis
unless they are interacted with ‘time varying variables’ such as time
dummies. Hence, when analysing individuals, factors such as gender or race
cannot be included among xit. We can now define the FE estimator as:
N 1 N N T 1 N T
ˆ xi' 
 xi xi' 
 yi x it' 
 xit xit 
 yit (9.10a )
FE
i 1 i 1 i 1i 1 i 1t 1

and cˆ i = y i − x i βˆ FE (9.10b)

When we estimate the time-demeaned Equation (9.9) by the pooled OLS

method, we have NT total observations and k independent variables. There is
no intercept term in (9.9) since it is eliminated by the fixed effects
transformation. Therefore, we should apparently have ‘NT – k’ degrees of
freedom. However, this calculation is incorrect. For each cross-sectional
observation i, we lose one df because of time-demeaning. In other words, for
each i, the demeaned errors uit add up to zero when summed across t. Hence,
we lose one degree of freedom. There is no such constraint on the original
idiosyncratic errors u it . Therefore, the appropriate degrees of freedom is
‘NT–N – k = N(T –1) – k’. Note that the significance of FE estimators can be
tested by using t-statistic. Note also that, unlike here, in case of RE
estimation we do not lose any df i.e. in the RE’s case the df is full (NT – k).
Note further that although the time-constant variables cannot be included by
themselves in a fixed effects model, they can interact with variables that
change over time i.e. with year dummy variables. Let us now consider the
following wage equation:

log( wageit ) educi exp erit exp erit2 uit

0 1 2 3
(9.11)

In the above wage equation, education is constant over time for each
individual in the sample. However, we can interact education with each year
171
Panel Data dummy to see how the return to education has changed over time. But we
Models
cannot use fixed effects to estimate the return to education in the base period.
This means we cannot estimate the return to education in any period. We can
only see how the return to education in each year differs from that in the base
period.

In applications of FE and RE, it is usually informative to compute the pooled

OLS estimates. Comparing the three sets of estimates can help us determine
the nature of the biases caused by leaving the unobserved effect, ci, ‘entirely’
in the error term (as done in the pooled OLS for FE) or ‘partially’ in the error
term (as done in the RE transformation). But we must remember that, even if
ci is uncorrelated with all explanatory variables in all time periods, the pooled
OLS’s standard errors and test statistics are generally invalid. This is because
they ignore the often substantial serial correlation in the composite errors, vit
= ci + uit.

Check Your Progress 2 [answer within the space given in about 50-100
words]
1) In Equation (9.2), what does ‘ ci is orthogonal to uit ’ mean?

……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
2) In what respects, the assumptions made for the FE differ from those
made for the RE effect?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
3) How is the FE estimate obtained in principle? To what effect?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
172
4) What is the d.f. in estimating the FE Equation (9.9)? Estimation of Panel
Data Models
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
5) What is a limitation of the FE estimate?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………

9.4 MODEL SELECTION: HAUSMAN TEST

The key consideration in choosing between a random effects and fixed effects
approach is whether ci and xit are correlated. It is therefore important to have
a method for testing this assumption. Hausman (1978) proposed a test based
on the difference between the random effects and fixed effects estimates.
Since FE is consistent when ci and xit are correlated (whereas RE is
inconsistent), a statistically significant difference is interpreted as ‘evidence
against the random effects’.

If our interest is in a time-varying explanatory variable, then one should use

RE rather than FE. But situations in which Cov (xit , ci ) = 0 should be
considered the exception rather than the rule. In most cases, the explanatory
variables are themselves a result of selection processes and would likely be
correlated with the unobserved factor, ci. In practice, one applies both the
random effects and the fixed effects, and then formally test for the statistical
significance of the differences of the coefficients on the time-varying
explanatory variables.

Hausman (1978) first proposed such a test and some econometrics packages
routinely compute the Hausman test under the full set of random effects
assumptions. The idea is that one uses the random effects estimates unless the
Hausman test rejects Cov (xit , ci ) = 0 . In practice, a failure to reject means
‘either that the RE and FE estimates are sufficiently close and it does not
matter which of the two methods is used’. A rejection, using the Hausman
test, is taken to mean that the key RE assumption, Cov (xit , ci ) = 0 , is false. In
such cases, the FE estimates are used.

Hausman’s specification test compares an estimator βˆ1 to be consistent with

an estimator βˆ 2 that is efficient. The null hypothesis it tests is that ‘the
173
Panel Data estimator βˆ 2 is efficient (and consistent) of the true parameters’. If this is the
Models
case, the conclusion made is that, there should be no systematic difference
between the two estimators. If there exists a systematic difference in the
estimates, we then have reason to doubt the assumptions on which the
efficient estimator is based. The Hausman statistic is distributed as χ 2 and is
computed as:
' 1
H Vc Ve
c e c e (9.12)
where β c is the coefficient vector from the consistent estimator, β e is the
coefficient vector from the efficient estimator, Vc is the covariance matrix of
the consistent estimator and Ve is the covariance matrix of the efficient
estimator. The ‘degrees of freedom’ for the statistic is the ‘rank of the
difference in the variance matrices’. When the difference is positive definite,
it is the number of common coefficients in the models being compared.

Hausman test can be used to decide whether to use the RE model or the FE
model. It states that: “if δˆ RE denotes the vector of RE estimates (without the
coefficient on the time constant variable or aggregate time variables), and if
δˆFE denotes the corresponding FE estimates, then:
1 1
H ˆ ˆ Avarˆ ˆ FE Avarˆ ˆ RE ˆ ˆ ~ 2
FE RE FE RE M

(9.13)”

The null hypothesis tested is H0: assumptions of RE models hold’. If we

reject H0 then it implies that the FE model must be considered. If you fail to
reject H0, then the implication is that the RE model is better.

Check Your Progress 3 [answer within the space given in about 50-100
words]

1) What is a key consideration in deciding between the choice for the FE or

RE model in panel data regressions? How is this achieved?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
2) How is the test result determined in the Hausman’s test?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
174 ……………………………………………………………………………
3) Outline the procedure adopted in the Hausman’s test. Estimation of Panel
Data Models
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………

9.5 SIMULTANEOUS EQUATION MODELS IN

PANEL DATA
The need for considering the ‘simultaneous equation models (SEM)’ arise
also in panel data contexts. In any regression modelling, generally, an
equation is considered to represent ‘a relationship’ to describe a phenomenon.
Many situations involve considering a ‘set of relationships’ to explain the
behaviour of variables. For instance, consider labour supply for a group of
married women already in the workforce. In place of the demand function,
we can write the ‘wage function’ in terms of ‘hours worked’ and the other
‘productivity variables’. With the equilibrium condition imposed, the two
structural equations become ‘labour supply function’ and the ‘wage function’
equations. In such a case, in addition to allowing for simultaneous
determination of variables within each time period, we can allow for the
‘unobserved effect’ in each equation. For instance, in a labour supply
function, it would be useful to allow for an ‘unobserved taste for leisure’ that
does not change over time.

The approach to estimating the SEMs with panel data involves two steps: (i)
eliminate the ‘unobserved effects’ from the equations by using the fixed
effects transformation (or first differencing) and (ii) find instrumental
variables for the endogenous variables in the transformed equation. This can
be very challenging because, for a convincing analysis, we need to find
instruments that change over time. Let us now consider an SEM for panel
data like:
y it1 = α 1 y it 2 + z it1 β 1 + ci1 + u it 1 (9.14)

y it 2 = α 1 y it 2 + z it 2 β 1 + ci 2 + u it 2 (9.15)

where i denotes cross section, t denotes time period, and, zit1 and zit2
represents linear functions of a set of exogenous explanatory variables. The
most flexible analysis allows the ‘unobserved effects’, ci1 and ci2, to be
correlated with all explanatory variables, including the elements in z.
However, we assume that the idiosyncratic structural error terms, uit1 and uit2,
are uncorrelated with z in both the equations and across all time periods. This
is in the sense of being truly exogenous. Further, except under very special
circumstance, yit2 is correlated with uit1, and yit1 is correlated with uit2.
175
Panel Data Now, let us consider the Equation (9.14). This cannot be estimated by the
Models
OLS since the composite error term cit1 + uit1 is potentially correlated with all
explanatory variables. Suppose we take their difference over time to remove
the unobserved effect cit1. That is:
∆y it1 = α 1 ∆y it 2 + ∆z it1 β 1 + ∆u it1 (9.16)

Now, with the differencing (or time-demeaning), we can only estimate the
effects of variables that change over time for at least some cross-sectional
units. The error term in the above equation is uncorrelated with ∆z it1 by
assumption but ∆yit 2 is not necessarily uncorrelated with ∆u it1 . Therefore, we
need an ‘instrumental variable (IV)’ for ∆yit 2 . It is possible that such IVs
comes from other equations. For instance, elements of ∆z it 2 that are not also
in ∆z it1 are natural IVs for ∆yit 2 .

Let us now consider the example of labour supply of married working

women. With the equilibrium condition imposed, the two structural equations
are:
hours = α 1 log(wage ) + β10 + β11educ + β 12 age + β13 kidslt 6
(9.17)
+ β 14 nwifeinc + u1

and
log(wage) = α 2 hours + β 20 + β 21educ + β 22 exp er
(9.18)
+ β 23 exp er 2 + u 2

The variable age is the woman’s age in years, kidslt6 is the number of
children less than six years old, nwifeinc is the woman’s non-wage income
(like husband’s earnings), and educ and exper are years of education and
prior experience respectively. All variables except hours and log(wage) are
assumed to be exogenous. Note that this is a tenuous assumption since educ
might be correlated with the omitted variable ‘ability’ in either equation. But
for illustrative purposes, we ignore the omitted variable problem here. Note
the functional form in this system (where hours appear in level form but
wage in logarithmic form) which is popularly used in labour economics. This
is for the simple reason that hours of work would change less but changes in
wages could be sharper (or minute) requiring to be captured by taking their
logarithmic values.

The first equation is the ‘supply function’. It satisfies the ‘order condition’
because the two exogenous variables, exper and exper2, are omitted from the
labour supply equation. These exclusions are crucial restrictions in the sense
that it amounts to assuming that ‘once wage, education, age, number of small
children, and other income are controlled for, past experience has no effect
on current labour supply’. The ‘wage equation’ is also identified as at least
one of age, kidslt6, or nwifeinc has a non-zero coefficient in (9.17). After
differencing, the labour supply function becomes:
176
 ∆hours = β 0 + α 1∆ log( wageit ) + ∆(other factorsit )β1 + ∆uit (9.19) Estimation of Panel
Data Models
We can use Δexpeerit as an instrument for ∆ log( wageit ) . However, because
we are looking at people who work in every time period, Δexpeerit = 1 for all
i and t (since each person gets another year of experience after a year passes).
We cannot therefore use Δexpeerit as an IV, as it takes same value for all i
and t. In a panel data set up, often, participation in an experimental
programme is used to obtain IVs. Let us now consider another example of
‘job training’ and ‘worker productivity’ where we want to estimate the effect
of ‘another hour of job training’ on ‘worker productivity’. For any two years
(e.g. 1987 and 1988), consider the simple panel data model:

log(scrapit ) = δ 0 + β1∆hrsempit + ∆uit (9.20)

where scrapit (the ‘scrap rate’ i.e. number of items out of 100 that must be
scrapped) is regressed on hrsempit. Normally, we would estimate this
equation by OLS. But ∆uit could be correlated with ∆hrsempit . For instance,
a firm might hire more skilled workers, while at the same time reducing the
level of job training. In this case, we need an instrumental variable for
∆hrsempit . Generally, such an IV is hard to find. But we can use the fact that
in ∆hrsemp it , some firms received job training grants in one of the two years
(1988). If we assume that grant designation is uncorrelated with ∆uit
(something that is reasonable because the grants are given at the beginning of
the year), then ∆grant it is a valid IV provided ∆hrsempit and ∆grant it are
correlated. Job training and worker productivity are jointly determined, but
receiving a job training grant is exogenous in Equation (9.20).

Let us consider another example of SEM with panel data. In order to estimate
the causal effect of increase in prison population on crime rates at the state
level, let us consider the ‘instances of prison overcrowding litigation as
instruments for the growth in prison population’. For the equation estimated
in first differences, we can write an underlying ‘fixed effects model’ as:
log(crimeit ) = θ t + α 1 log( prisonit ) + z it1 β1 + ci1 + u it1 (9.21)

where θ t denotes different time intercepts, and crime and prison are measured
per 100,000 people. The prison population variable is measured on the last
day of the previous year. The vector zit1 contains log of ‘police per capita, log
of income per capita, the unemployment rate, proportions of black and those
living in metropolitan areas, and age distribution proportions’. Differencing
(9.21) gives the equation to be estimated as:
∆ log(crimeit ) = ξ t + α 1∆ log( prisonit ) + ∆z it1 β 1 + ∆u it1 (9.22)

Simultaneity between ‘crime rates’ and ‘prison population’ (or more

precisely in their growth rates), makes the OLS estimation of (9.22)
inconsistent. Hence, we can estimate (9.22) by the ‘pooled 2SLS method’.
After the pooled 2SLS estimation, we obtain the residuals rˆit 1 . Then, by
177
Panel Data including one lag of these residuals in the original equation, we can estimate
Models
the equation by 2SLS with rˆit 1 acting as its own instrument. The first year is
lost because of the lagging. Hence, the usual 2SLS t statistic on the lagged
residual is a valid test for serial correlation. Let us now consider another
example of a panel data model to examine the effects of cigarette smoking on
earnings as:
log( wageit ) = z it γ + δ 1cigsit + ci + u it (9.23)

We want to know the causal effect of smoking on hourly wage. For

concreteness, we assume cigsit is measured as ‘average packs per day’. This
equation has a causal interpretation i.e. holding the factors in z it and ci fixed,
what is the effect of an exogenous change in cigarette smoking on wages?
Thus, Equation (9.23) is a structural equation.

The presence of the individual heterogeneity, ci, in Equation (9.23)

recognises that cigarette smoking might be correlated with individual
characteristics that also affect wage. An additional problem is that cigsit
might also be correlated with u it . In this example, the correlation could be
from a variety of sources. But, simultaneity is one possibility because if
cigarettes are treated as a normal good, then, as income increases, holding
everything else fixed, cigarette consumption increases. Therefore, we might
add another equation to Equation (9.23) that reflects cigsit . It may depend on
income which depends on wage. However, if Equation (9.23) is of interest,
we do not need to add equations explicitly. But we need to find some
instrumental variable. To get an estimable model, we must first deal with ci
since it can be correlated with zit as well as cigsit . We can use the FE method
to eliminate ci before addressing the correlation between cigsit and uit. The
approach that can be used to estimating the SEMs with panel data is to use
the fixed effects transformation and then to apply an IV technique such as
pooled 2SLS. A simple procedure is to estimate the time-demeaned equation
by pooled 2SLS like:
yit1 = α 1 yit 2 + zit1 β 1 + uit1 , t=1,2,….,T (9.24)

where zit 1 and zit 2 are IVs.

Check Your Progress 4 [answer within the space given in about 50-100
words]

1) What are the two factors to be considered in the approach to estimate the
SEMs in panel data regressions?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
178
2) How is the ‘unobserved effect’ removed from a composite error term like Estimation of Panel
Data Models
cit1 + uit1 in Equation (9.14)? What is a consequent disadvantage of this
approach and how is this dealt with?
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………
……………………………………………………………………………

9.6 LET US SUM UP

The unit introduces the methods of estimating the panel data models. Two
cases of ‘random effects’ and ‘fixed effects’ are separately discussed. The set
of assumptions governing the estimation process are listed. The method of
converting or transforming the data into ‘quasi de-meaned data’ is indicated.
For making a choice between the ‘fixed effects’ model and the ‘random
effects’ model, the procedure for applying the Haussman’s test is outlined.
Since situations of simultaneity affects the estimation process in panel data
contexts, methods of dealing with such cases is explained with many
illustrations.

9.7 KEY WORDS

Linear Static : It is a model like in Equation (9.1). In this, a

Panel Data Model separate term for ‘unobserved effects’, ci, assumed
to be random variables, is introduced. Further, in
order to be able to assume the effect of ci to be
‘zero’ on an average, we introduce a separate
intercept term as in Equation (9.2).
Quasi-Demeaned : Refers to the procedure adopted for transforming the
Data variables to remove the effect of their means. It is a
partial (quasi) cleaning process since the ‘random
effects transformation subtracts a fraction of the time
average effect’.
Fixed Effects : Refers to a ‘pooled OLS estimator’ based on the
Estimator time-demeaned variables’ used in the panel
regression models.
Hausman Test : This is a test procedure proposed by Hausman which
helps us to make a choice between a ‘random
effects’ model and a ‘fixed effects’ model. The test
statistic [as in Equation (9.12)] is distributed as
χ2 with the d.f. equal to ‘the rank of the difference in
the two covariance matrices’ i.e. the covariance
matrix of the ‘consistent estimator’ and the ‘efficient 179
Panel Data estimator’. It is based on the difference between the
Models
random effects and fixed effects estimates.

9.8 SUGGESTED BOOKS FOR FURTHER

READING
1) Wooldridge J M (2006). Introduction to Econometrics: A Modern
Approach. Michigan State University. USA.
2) Greene W H (2016). Econometrics. Prentice Hall.

9.9 ANSWERS/HINTS TO CHECK YOUR

PROGRESS EXERCISES
Check Your Progress 1

1) A linear static panel data model is like Yit = β0 + Xit βi + ci + uit. Here
Xi’s are fixed or pre-determined (and hence the model called as the
‘static’ model) corresponding to the value of Y in period t. The ci’s stand
for the ‘unobserved effect’ and uit’s are the ‘idiosyncratic error terms’.
The term ‘idiosyncratic’ refers to the random character of the
observation-specific ‘error term’. The zero mean characteristic of ci’s is
achieved by the inclusion of the explicit intercept term β0.
2) The OLS method cannot be applied because of the composite error term
vit = ci + uit which implies that the error terms are serially correlated
across time. In such cases, we have to apply the GLS method of
estimation with ‘auto regressive serial correlation’.
3) The assumptions required for RE model are: (i) the model is defined as
in equation (9.3) with a term for ‘unobserved effect’ (ci ) and ‘random
error term’ (uit), (ii) explanatory variables change over time but not with
perfect collinearity, (iii) the expected value of error terms (uit) given Xi
and ci is ‘zero’, (iv) var (uit) = σ u2 (v) the covariance of uncorrelated
error terms conditional to xi and ci is ‘zero’, (vi) E (ci ) = 0 and (vii)
var(ci | xi ) = σ c2 i.e. the assumption of homoscedasticity hold for ci’s.

4) The term ‘quasi’ means ‘partial’. The term ‘demeaned’ can be split as
‘de-meaned’ which means removing the effect of mean. Hence, the term
‘quasi demeaned data’ refers to the partial adjustment made in each
variable to remove the effects of their means. In effect, it is a partial
cleaning process attained by the estimation process of GLS.

Check Your Progress 2

1) It means the two are correlated with each other. In mathematical

terminology, it means that they are not parallel (when they will never
meet or cross) but perpendicular and hence intersecting with each other.

180
2) Firstly, in terms of the effect on the estimated parameters. They are in the Estimation of Panel
Data Models
first stage consistent and in the later stage ‘best’ i.e. BLUE. With the
additional assumption on the normality of uit’s (conditional on xi and ci)
i.e. making the nature of error terms i.i.d., the t and F statistics have the
exact distributions without the need for approximation of large N and
small T.
3) By a process of ‘time demeaning’ transformation for removing the effect
of ‘average’ in Equation (9.6). The transformation eliminates the
‘unobserved effect’ ci as in the transformed Equations (9.8) or (9.9). The
estimation of (9.9) becomes the ‘pooled OLS’ method. A major
difference between the FE and the RE methods is that, FE is consistent
when ci and xit are correlated, whereas the RE estimate is not.
4) N(T –1) – k.
5) The pooled OLS’s standard errors and test statistics are generally invalid,
because, they ignore the serial correlation in the composite errors: vit = ci
+ uit.

Check Your Progress 3

1) It is ‘whether ci and xit are correlated’. It is achieved by the Hausman

test. It is based on the difference between the random effects and fixed
effects estimates.
2) A statistically significant difference ‘between the random effects and
fixed effects estimates’ is interpreted as ‘evidence against the random
effects’.
3) Hausman’s specification test compares an estimator βˆ1 to be consistent
with an estimator βˆ that is efficient. The null hypothesis it tests is that
2

‘the estimator βˆ 2 is efficient (and consistent) of the true parameters’.

Hausman statistic is distributed as χ2 and is computed as:
H = (β c − β e ) (Vc − Ve ) (β c − β e ) where β c is the coefficient vector
' −1

from the consistent estimator, β e is the coefficient vector from the

efficient estimator, Vc is the covariance matrix of the consistent estimator
and Ve is the covariance matrix of the efficient estimator.

Check Your Progress 4

1) Eliminating the ‘unobserved effects’ from the equations by using the

‘fixed effects transformation’ and (ii) finding instrumental variables for
the endogenous variables in the transformed equation.
2) By considering the difference equation over time as in Equation (9.16).
The error term in (9.16) is uncorrelated with ∆z it1 (by assumption) but
∆yit 2 is not necessarily uncorrelated with ∆u it1 . As a result, we need an
‘instrumental variable (IV)’ for ∆y it 2.

181
Panel Data
Models

182