ES1004 Econometrics by Example

Lecture 12: Panel Data Models

Dr. Hany Abdel-Latif

Swansea University, UK

Gujarati textbook, second edition [chapter 17]

11th February 2017

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 1 / 48

 Data Types of Data

Time Series Data

Table: An example

country year x y
Egypt 2000 0.28 0.46
Egypt 2001 0.24 0.51
Egypt 2002 0.26 0.23
Egypt 2003 0.55 0.65
Egypt 2004 0.63 0.95
Egypt 2005 0.25 0.19
Egypt 2006 0.20 0.28
Egypt 2007 0.56 0.78
Egypt 2008 0.50 0.52
Egypt 2009 0.86 0.86
Egypt 2010 0.11 0.20

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 2 / 48

 Data Types of Data

Cross-Sectional Data

Table: An Example

country year x y
Egypt 2000 0.31 0.09
USA 2000 0.30 0.10
Italy 2000 0.30 0.72
Greece 2000 0.17 0.34
Lebanon 2000 0.87 0.28
Algeria 2000 0.24 0.99
Tunisia 2000 0.27 0.51
Kuwait 2000 0.66 0.15
Germany 2000 0.73 0.60
France 2000 0.94 0.88
UK 2000 0.20 0.02

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 3 / 48

 Data Types of Data

Panel Data

Table: An Example

country year x y
Egypt 2000 0.06 0.15
Egypt 2001 0.36 0.91
Egypt 2002 0.02 0.28
Egypt 2003 0.84 0.44
Egypt 2004 0.48 0.24
USA 2000 0.55 0.01
USA 2001 0.15 0.54
USA 2002 0.05 0.19
USA 2003 0.53 0.80
USA 2004 0.19 0.54
USA 2005 0.45 0.26

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 4 / 48

 Panel Data Advantages

The Importance of Panel Data

can take heterogeneity explicitly into account by allowing for

subject-specific variables
more informative, more variability less collinearity, more degrees of
freedom
allows for the study of the dynamics of change

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 5 / 48

 Panel Data Example

Example: Charitable Giving

table17.1 data on charitable giving

47 individuals over 1979-1988
charity: sum of cash and other property contributions
income: gross income
price: one minus the marginal tax rate
age: a dummy equal to 1 if the taxpayer is over 64
MS: a dummy equal to 1 if married
DEPS: number of dependants
find out the effect of the marginal tax rate on charitable giving
[charity function]

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 6 / 48

 Panel Data Example

Example: Data

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 7 / 48

 Panel Data Example

Example: Balanced & Short Panel

balanced panel: the number of time observation is the same of each

individual [equals 10 in this example]
unbalanced panel:
short panel: the number of cross-sectional or individual units N is
greater than the number of time periods
[in this example N=47 & T=10]
long panel: T is greater than N

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 8 / 48

 Panel Data Estimation

Estimation: Five Options

1 individual times series of charity functions

very few degrees of freedom; meaningless statistical analysis
neglect information about other individuals’ charity contributions; they
all operate in the same regulatory environment
2 cross-sectional charity functions
neglect the dynamic aspect of charitable giving

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 9 / 48

 Panel Data Estimation

Estimation: Five Options

3 pooled OLS charity function

pool all 470 observations [47 × 10]
neglect the dual nature of time series and cross-sectional data
assumes that the coefficients remain constant across time and
cross-section
aka constant coefficient model

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 10 / 48

 Panel Data Estimation

Estimation: Five Options

4 fixed effects least squares dummy variable (LSDV) model

pool all observations but allow each individual unit to have its
individual intercept dummy
this is different from the within estimator which we will discuss shortly

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 11 / 48

 Panel Data Estimation

Estimation: Five Options

5 random effects model

assumes that the intercepts values of the 47 individuals are random
drawings from a much larger population of individuals

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 12 / 48

 Panel Data Pooled OLS

Example: pooled OLS

Cit = β1 + β2 Ageit + β3 Incomeit + β4 Priceit + β5 DEPSit + β6 MSit + uit

i = 1, 2, . . . , 47; t = 1, 2, . . . , 10
notice we have two subscripts (i&t)
nonstochastic regressors, or if stochastic, uncorrelated with the error
term
the error terms satisfies the usual classical assumptions

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 13 / 48

 Panel Data Pooled OLS

Example: a priori

we would expect
age, income, price and marital status to have a positive impact on
charitable giving
the number of dependants to have a negative impact
the price variable represents the opportunity cost of giving charitable
contributions
the higher the marginal tax, the lower the opportunity cost

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 14 / 48

 Panel Data Pooled OLS

EViews: import data

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 15 / 48

 Panel Data Pooled OLS

EViews: import data

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 16 / 48

 Panel Data Pooled OLS

EViews: import data

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 17 / 48

 Panel Data Pooled OLS

EViews: import data

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 18 / 48

 Panel Data Pooled OLS

EViews: import data

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 19 / 48

 Panel Data Pooled OLS

Estimation: pooled OLS

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 20 / 48

 Panel Data Pooled OLS

Estimation: pooled OLS

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 21 / 48

 Panel Data Pooled OLS

Estimation: pooled OLS

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 22 / 48

 Panel Data Fixed Effects

Fixed Effects Least Squares Dummy Variables LSDV

one way to take into the heterogeniety among 47 individuals is to

allow each individual to have his or her own intercept

Cit = β1i + β2 Ageit + β3 Incomeit + β4 Priceit + β5 DEPSit + β6 MSit + uit

i = 1, 2, . . . , 47; t = 1, 2, . . . , 10
notice we have added the subscript i to the intercept
differences may be due to special features of each individual, such as
education or religion

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 23 / 48

 Panel Data Fixed Effects

Fixed Effects Regression Model LSDV

the term ‘fixed effects’

each taxpayers’ intercept is different from the intercepts of other
taxpayers
however individual intercepts do not very over time [time invariant]

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 24 / 48

 Panel Data Fixed Effects

Fixed Effects Regression Model LSDV

done easily by introducing differential intercept dummies

Cit = β1 + β2 D2i + β3 D3i + · · · + β4 6D46i + β47 Ageit + β48 Incomeit +

β49 Priceit + β50 DEPSit + β51 MSit + uit

46 dummies represents 47 individuals to avoid dummy variable trap

(perfect multicollinearity)
individual 1 is our benchmark or reference category; any individual can
be chosen for that purpose though

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 25 / 48

 Panel Data Fixed Effects

Estimation: LSDV

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 26 / 48

 Panel Data Fixed Effects

Estimation: LSDV

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 27 / 48

 Panel Data Fixed Effects

Estimation: LSDV

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 28 / 48

 Panel Data Fixed Effects

Pooled OLS vs LSDV

a test to find out if the fixed effects model is better than the OLS
pooled model
the pooled model is restricted
neglects the heterogeneity effects that are explicitly taken into account
in the fixed effects model
we can use the restricted F test

2 − R 2 )/m
(Rur r
F = 2 )/(n − k)
(1 − Rur
m number of parameters excluded (46 here)
k number of parameters estimated in the unrestricted model (52 here)

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 29 / 48

 Panel Data Fixed Effects

Pooled OLS vs LSDV

(0.7632 − 0.2245)/46
F = = 20.672
(1 − 0.7632)/(418)

for 46 df in the numerator and 418 in the dominator, this F is highly

significant
confirms that the fixed effect model is superior to the pooled
regression model

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 30 / 48

 Panel Data Fixed Effects

Limitations to the Fixed Effects LSDV model

every additional dummy variable will cost an additional degree of

freedom (very costly in small samples)
too many additive and multiplicative dummies may lead to
multicollinearity

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 31 / 48

 Panel Data Fixed Effects

Limitations to the Fixed Effects LSDV model

we assume that uit ∼ N(0, σ 2 ) but since we have i and t we may

have to modify this classical assumption
there are several possibilities
1 assume the error variance is the same for all cross-sectional unit or it is
heteroscedastic
2 for each subject, we can assume there is no autocorrelation over time,
or we can assume autocorrelation of the AR(1) type
3 at any given time, we can allow the error term of individual 1 to be
non-correlated with the error for say individual 2, or we can assume
that there is such correlation

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 32 / 48

 Panel Data WG Estimator

Fixed Effects Within Group Estimator

we can eliminate the fixed effect term by expressing both the

regressand and regressors as deviations for their respective (group)
mean values
run the regression on the mean corrected variables

Cit − C i = B2 (Ageit − Age i ) + β3 (Incomeit − income i )

+β4 (Priceit − Price i ) + β5 (Depsit − Deps i )+
β6 (MSit − MS i ) + uit

notice how fixed effects or individual effect intercept term

β1i drops out

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 33 / 48

 Panel Data WG Estimator

Estimation: Demeaned Variables

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 34 / 48

 Panel Data WG Estimator

Estimation: WG Estimator

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 35 / 48

 Panel Data WG Estimator

Estimation: WG Estimator

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 36 / 48

 Panel Data WG Estimator

FEM Robust Standard Errors

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 37 / 48

 Panel Data WG Estimator

FEM Robust Standard Errors

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 38 / 48

 Panel Data WG Estimator

FEM Robust Standard Errors

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 39 / 48

 Panel Data Random Effects

Random Effects Model REM

FEM assumes individual specific coefficient β1i is fixed for each

subject [i.e., time-invariant]
REM β1i is a random variable with a mean value of β1 (no i subscript
here) and the intercept of any cross-section unit is expressed as

β1i = β1 + i
where i is a random error term with mean 0 and variance σ2

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 40 / 48

 Panel Data Random Effects

Random Effects Model REM

Cit = β1 + β2 Ageit + β3 Incomeit + β4 Priceit + β5 DEPSit + β6 MSit + wit

where

wit = i + uit

the 47 individuals are drawn from a much larger universe and have a
common mean value for the intercept (= β1 )
differences in the individual values of the intercept for each donor are
reflected in the error term i

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 41 / 48

 Panel Data Random Effects

Random Effects Model REM

wit = i + uit

the composite error term wit has two components

i the cross-section or individual-specific error component, and
uit the combined time series and cross-section error component
[aka idiosyncratic term because it varies over cross-section i.e.,
individual as well as time]

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 42 / 48

 Panel Data Random Effects

Estimation: REM

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 43 / 48

 Panel Data Random Effects

Estimation: REM

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 44 / 48

 Panel Data Random Effects

Estimation: REM

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 45 / 48

 Panel Data Fixed Effects vs Random Effects

Hausman Test

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 46 / 48

 Panel Data Fixed Effects vs Random Effects

Hausman Test

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 47 / 48

 Questions & Answers

Dr. Hany Abdel-Latif (2017) ES1004ebe Lecture 12 Panel Data Models 48 / 48