ECTA [U06982]
Guy Judge November 2008
Econometric Analysis
Week 10
Pooled and panel data models
ECTA [U06982]
Guy Judge November 2008
Lecture outline
Basics
–
data sets with both cross

section and time
dimensions
Pooled regressions and the use of time period dummies
A note on Chow tests versus interactive dummies in the
presence of heteroskedasticity
Panel data and longitudinal data sets
Attractive features of panel data sets
Fixed effects models
–
differencing and demeaning
Random effects models
An illustrative example
ECTA [U06982]
Guy Judge November 2008
References and recommended reading
Wooldridge, J M (2006) Introductory
Econometrics. A Modern Approach. (Third
Edition) Chapters 13 and 14
Kennedy, P (2003) A Guide to Econometrics.
(Fifth Edition) Chapter 17
Dougherty, C (2007) Introduction to
Econometrics (Third Edition) Chapter 14
Gujarati, D N (2003) Basic econometrics
(Fourth Edition) Chapter 16
ECTA [U06982]
Guy Judge November 2008
Basics
many data sets have both a cross

section and a time
dimension
two subscripts are required for the variables
x
it
i = 1,…..,n
t = 1,…..,T
If n is large but T is quite small (say only 2 or 3) then we
may decide just to apply cross

section methods, but with
intercept (and possibly even slope) dummies to distinguish
observations from different time periods
–
or we might be
able just to
pool
the data.
ECTA [U06982]
Guy Judge November 2008
Pooled regressions with time period dummies
If the time period dummies are insignificant
then the data from the different periods can be
brought together to form one
pooled data
set
We might also think of using
Chow tests
to check
the validity of pooling the data (effectively a test
of structural change)
For example
i = 1,…n; t=1,2
If the data are pooled then we are effectively
incorporating the restriction that
0
and
1
remain
unchanged between periods 1 and 2
it
it
it
u
x
y
1
0
ECTA [U06982]
Guy Judge November 2008
More on pooled models with time dummies
Including a standard dummy variable just allows
the intercept to change
The intercept in period 2 =
it
it
it
it
u
x
D
y
1
0
0
2
1
;
1
0
t
for
D
t
for
D
it
it
0
0
ECTA [U06982]
Guy Judge November 2008
Chow tests
A Chow test looks for evidence of changes in
both intercept and slope parameters
Effectively
Chow tests the restrictions
The pooled data set imposes these restrictions
n
i
t
for
u
x
y
it
it
it
,...,
1
,
1
]
1
[
1
]
1
[
0
n
i
t
for
u
x
y
it
it
it
,...,
1
,
2
]
2
[
1
]
2
[
0
]
2
[
1
]
1
[
1
]
2
[
0
]
1
[
0
;
ECTA [U06982]
Guy Judge November 2008
Interactive slope dummies
The same test could be undertaken using
interactive
slope dummies in a pooled regression
Intercept in period 2 =
0
+
0
Slope in period 2 =
1
+
1
Note: The Chow test requires that the model is free of
heteroskedasticity

which may not always be the case. The
dummy variable approach could in this case be combined with
the use of
heteroskedastic consistent standard errors
for
computing t values to assess the significance of
0
and
1
it
it
it
it
it
it
u
x
D
x
D
y
1
1
0
0
ECTA [U06982]
Guy Judge November 2008
Longitudinal data sets
It T is large, as with so called
longitudinal
data
sets, then we may need to consider the nature of any
trends in the variables.
Unit root tests for mixed cross

section time series
data have been developed, but we shall not consider
them here
–
but we will mention that simpler models
may just incorporate lagged dependent variables in
the regression equations.
So, without concerning ourselves about unit roots,
we will just examine two popular types of panel data
models :
fixed effects
and
random effects
models.
ECTA [U06982]
Guy Judge November 2008
More on longitudinal data sets
If the observations in the different time periods relate to
exactly
the same subjects then we may refer to the data
set as
panel data
. For example, the famous US
National
Longitudinal Survey of Youth
tracks the same individuals
over several years. Another well known US panel data set is
the
Panel Study of Income Dynamics (PSID)
.
Note: sometimes this might mean that we have to deal with
unbalanced panels
when some individuals disappear from
the panel sample.
The
Current Population Survey (CPS)
on the other hand
extracts a
different random sample
each year so the cross

section data are
not matched
to the same individuals. But
because they are
independently
drawn the data can be
pooled with the possible addition of time dummies.
ECTA [U06982]
Guy Judge November 2008
Attractive features of panel data sets
Panel data can enable us to examine issues not amenable to study
using only cross

section or time series data sets.
For example with production functions we can deal
simultaneously with issues of economies of scale and of
technological change.
Cross

section labour market data can tell us who is unemployed
in any particular year and time series data can tell us how overall
unemployment changes from year to year. Panel data enables us to
track individuals
to help us answer questions about
unemployment
duration
, turnover rates etc.
* Panel data can help us deal with issues of
heterogeneity
in the
micro units.
Unobserved factors
affecting different people can
cause
bias
in cross

section studies, but with panel data
differencing
or
demeaning
can control for these factors.
We can introduce some allowance for
dynamic adjustment
in
models by including lagged variables.
ECTA [U06982]
Guy Judge November 2008
Analysis of the CRIME2 data set in Wooldridge
CRIME2.xls contains data on (amongst other
variables) the crime rate (
crmrte
) and the
unemployment rate (
unem
) for 46 US cities in the
years 1982 and 1987
Question
: Would you expect cities with a higher
unemployment rate to have a higher or lower crime
rate?
So if, like Wooldridge p 460, you used just the
1987 data and ran a regression of
crmrte
on
unem
would you expect
1
to be positive or negative?
ECTA [U06982]
Guy Judge November 2008
Replicating Wooldridge’s results on p460
EQ( 1) Modelling crmrte by OLS (using crime2sorted.xls)
The estimation sample is: 47
–
92
Coefficient Std.Error t

value t

prob Part.R^2
Constant 28.378 20.76 6.18 0.000 0.4651
unem

4.16113 3.416

1.22 0.230 0.0326
sigma 34.5999 RSS 52674.6416
R^2 0.0326151 F(1,44) = 1.483 [0.230]
log

likelihood

227.266 DW 1.11
no. of observations 46 no. of parameters 2
mean(crmrte) 103.873 var(crmrte) 1183.71
Comment: Not only does the coefficient of
unem
have the
“wrong” sign, it is also not significantly different from zero
–
see
the very low t, F and R^2 values.
Note: the figures in brackets on p460 of Wooldridge are standard
errors, not t

values.
ECTA [U06982]
Guy Judge November 2008
Further comments on Wooldridge’s results
Wooldridge notes that this simple model is deficient in
many ways. To improve it we might
introduce
additional regressors
including demographic
factors such as the age distribution in each city, gender
balance, educational data, law enforcement efforts
try a different
functional form
include a
lagged dependent variable
(the crime rate in
the earlier year)
But Wooldridge wants to use this model to demonstrate
how, with two periods of data, we can control for
individual unobserved
fixed effects
and thus remove this
potential form of bias.
ECTA [U06982]
Guy Judge November 2008
Dealing with individual heterogeneity
Suppose we have an individual
unobserved factor
, or set of factors, that
affects crime in different cities, but remains unchanged (
fixed
) between
the different time periods. Denoting this by a
i
, following Wooldridge, we
can write
Here a
i
picks up all those factors unique to the individual cities that
don’t change, or don’t change much, between periods
–
including the
demographic factors noted on the previous slide. Wooldridge calls this the
unobserved heterogeneity
error term.
The usual error term u
it
picks up all other factors that disturb the
observed values of the dependent variable both across the different cities
and between the years

Wooldridge calls this the
idiosyncratic
error term.
You can see that this model also includes a dummy variable to allow for
intercept shifts (common to all cities) between periods.
it
i
it
it
it
u
a
unem
d
crmrte
1
0
0
87
ECTA [U06982]
Guy Judge November 2008
Replicating Wooldridge’s results p462
The coefficient on
unem
, although positive, is still not significant.
Here pooled OLS has not solved the omitted variables problem.
EQ( 2) Modelling crmrte by OLS (using crime2.in7)
The estimation sample is: 1

92
Coefficient Std.Error t

value t

prob Part.R^2
Constant 93.4203 12.74 7.33 0.000 0.3766
d87 7.94041 7.975 0.996 0.322 0.0110
unem 0.426546 1.188 0.359 0.720 0.0014
sigma 29.9917 RSS 80055.7841
R^2 0.0122119 F(2,89) = 0.5501 [0.579]
log

likelihood

441.902 DW 1.16
no. of observations 92 no. of parameters 3
mean(crmrte) 100.791 var(crmrte) 880.929
If we were to ignore the heterogeneous effects and just pool the
data (but including the time period dummy) we would find
ECTA [U06982]
Guy Judge November 2008
A first

differenced model
Because a
i
is constant over time we can remove its
effect by first

differencing the equation. Notice that
the initial intercept term
0
gets removed too,
leaving us with
Here I am using Wooldridge’s labels
ccrmrte
and
cunem
respectively for
crmrte
and
unem.
Wooldridge’s results for this regression are replicated
in PcGive and shown on the next slide.
it
it
it
u
cunem
ccrmrte
1
0
ECTA [U06982]
Guy Judge November 2008
Replicating Wooldridge’s results p464
Comments:
(1)
in this regression the estimate of
1
is positive and statistically
significant
(2)
the positive and statistically significant estimate of
0
shows
evidence of a secular increase in crime across all cities between
1982 and 1987.
EQ( 3) Modelling ccrmrte by OLS (using ccrime2.xls)
The estimation sample is: 1

46
Coefficient Std.Error t

value t

prob Part.R^2
Constant 15.4022 4.702 3.28 0.002 0.1960
cunem 2.21800 0.8779 2.53 0.015 0.1267
sigma 20.0508 RSS 17689.5501
R^2 0.1267 F(1,44) = 6.384 [0.015]*
log

likelihood

202.169 DW 1.15
no. of observations 46 no. of parameters 2
mean(ccrmrte) 6.16375 var(ccrmrte) 440.348
ECTA [U06982]
Guy Judge November 2008
The fixed effects model with time demeaned data
Define the
time demeaned
data series for y as
where
(and similarly for x and u)
then if we have
and
then we can estimate
Again the unobserved effect has disappeared.
t
it
it
y
y
y
T
t
it
t
y
T
y
1
1
it
i
it
it
u
a
x
y
1
0
t
i
t
t
u
a
x
y
1
0
it
it
it
u
x
y
1
ECTA [U06982]
Guy Judge November 2008
Some comments on these two approaches
If T=2 the first

differencing approach and the fixed effects
demeaned data approaches are equivalent
(see W3 p491)
But with T>2 the choice depends on the relative efficiency of
the two methods.
Wooldridge says that the first

differences method is better if
the differenced u term is serially uncorrelated
–
but he
advises you to try both approaches and look to explain why
the results differ (if they do).
ECTA [U06982]
Guy Judge November 2008
Random effects models
it
i
it
it
u
a
x
y
1
0
Again suppose that we have
But here we assume that
(Wooldridge allows for k separate x variables, so each must be
uncorrelated with a).
Writing v
it
as the composite error
We have
The composite error v
it
is serially correlated
0
)
,
cov(
i
it
a
x
it
i
it
u
a
v
it
it
it
v
x
y
1
0
2
2
2
)
,
(
a
u
u
is
it
v
v
corr
ECTA [U06982]
Guy Judge November 2008
More on the random effects model
If we knew these variances we could calculate
and use
Generalised Least Squares
based on the
quasi

demeaned data
and
This purges the disturbance term of the serial correlation.
In practice this means that we use
Feasible GLS
estimation
(Kennedy calls it EGLS
–
estimated GLS) where
is
estimated as part of the process. Wooldridge comments
(p495) that the algebra is fairly unpleasant
–
but most
econometric software packages will do this for you.
2
1
2
2
2
/(
1
a
u
u
T
i
it
y
y
i
it
x
x
ECTA [U06982]
Guy Judge November 2008
Some final comments
Kennedy proposes a strategy in which you use a Hausman
exogeneity test to see if the random effects estimator is
unbiased (if this null is not rejected then use the RE model,
otherwise use the FE approach).
Wooldridge says the key issue is whether one can plausibly
assume that the a
i
are uncorrelated with the x variables (in
which case one can use the RE model via FGLS estimation).
But it is just a question of which estimator is more efficient so
the fixed effects estimator would still be unbiased and
consistent.
Σχόλια 0
Συνδεθείτε για να κοινοποιήσετε σχόλιο