Panel data analysis

Fixed effect transformation

Consider a model with a single explanatory variable

yit = β1 xit + ai + uit , t = 1, 2, . . . , T (1)

Now for each i, average this equation over time. We get

ȳi = β1 x̄i + ai + ui (2)

1 P
where ȳi = yit
T
Because ai is fixed over time, it appears both in 1 and 2. If we
substract 2 from 1 for each t, we wind up with
Fixed effect transformation

yit − ȳi = β1 (xit − x̄i ) + uit − ūi , t = 1, . . . , T (3)

or
ÿit = β1 ẍit + üit , t = 1, . . . , T (4)
where ÿit = yit − ȳi is the time-demeaned data on y, and similarly
for ẍit and üit
▶ The fixed effects transformation is also called the within
transformation
▶ The important thing about equation (4) is that the
unobserved effect, ai , has disappeared.
▶ This suggests we should estimate (4) by pooled OLS.
▶ A pooled OLS estimator that is based on the time-demeaned
variables is called the fixed effects estimator or the within
estimator .
▶ The latter name comes from the fact that OLS on (4) uses
the time variation in y and within each cross-sectional
observation.
Fixed effect transformation

▶ The between estimator is obtained as the OLS estimator on

the cross-sectional equation 2
▶ We include an intercept, β0
▶ We use the time averages for both y and x.
▶ Then, we run a cross-sectional regression.
▶ We will not study the between estimator in detail because it is
biased when ai is correlated with x̄it .
Fixed effect estimator

▶ Under a strict exogeneity assumption on the explanatory

variables, the fixed effects estimator is unbiased
▶ Roughly, the idiosyncratic error uit should be uncorrelated
with each explanatory variable across all time periods.
▶ The fixed effects estimator allows for arbitrary correlation
between ai 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.
▶ Therefore we cannot include variables such as gender or a
city’s distance from a river.
Example 14.2
▶ Each of the 545 men in the sample worked in every year from
1980 through 1987.
▶ Variables in the data set change over time: experience,
marital status, and union status.
▶ Variables in the data set do not change over time: race,
education, experience.
▶ If we use fixed effects (or first differencing), we cannot include
race, education, or experience in the equation.
▶ However, we can include interactions of educ with year
dummies for 1981 through 1987 to test whether the return to
education was constant over this time period.
▶ We use
▶ log(wage) as the dependent variable,
▶ dummy variables for marital and union status,
▶ a full set of year dummies, and the interaction terms
d81 ∗ educ, d82 ∗ educ, . . . , d87 ∗ educ.
Example 14.2

▶ The estimates on these interaction terms are all positive, and

they generally get larger for more recent years.
▶ The largest coefficient of .030 is on d87educ, with t = 2.48.
▶ In other words, the return to education is estimated to be
about 3 percentage points larger in 1987 than in the base
year, 1980.
▶ Give significant α = 0.05, the other significant interaction
term is d86educ (coefficient=.027, t = 2.23).
▶ The estimates on the earlier years are smaller and insignificant
at the 5% level against a two-sided alternative.
The dummy variable regression
▶ A traditional view of the fixed effects approach is to assume
that the unobserved effect, ai , is a parameter to be estimated
for each i.
▶ Thus, in equation 1, ai is the intercept for person i (or firm i,
city i, and so on) that is to be estimated along with the βj .
▶ Clearly, we cannot do this with a single cross section: there
would be N + k parameters to estimate with only N
observations. We need at least two time periods.
▶ The way we estimate an intercept for each i is to put in a
dummy variable for each cross-sectional observation, along
with the explanatory variables (and probably dummy variables
for each time period)
▶ This method is usually called the dummy variable regression.
▶ After fixed effects estimation with N of any size, the âi are
pretty easy to compute:

âi = ȳi − β̂1 x̄i1 − . . . − β̂k x̄ik , i = 1, . . . , k (5)

Fixed effects of first differencing ?

▶ When T = 2, the FE and FD estimates, as well as all test

statistics, are identical, and so it does not matter which we
use.
▶ When T > 2, the FE and FD estimators are not the same.
▶ When the uit are serially uncorrelated, fixed effects is more
efficient than first differencing (and the standard errors
reported from fixed effects are valid).
▶ When uit are serially correlated. One example is that uit
follows a random walk (which means that there is very
substantial, positive serial correlation). In this case, the
difference is serially uncorrelated, and first differencing is
better.
Practice: Exercises at the end of the chapter 14.
Random effect models

We begin with the same unobserved effects model as before,

yit = β0 + β1 xit1 + . . . + βk xitk + ai + uit (6)

where we explicitly include an intercept so that we can make the

assumption that the unobserved effect, ai , has zero mean

Equation (14.7) becomes a random effects model when we assume

that the unobserved effect ai is uncorrelated with each explanatory
variable.

(xitj , ai ) = 0, t = 1, 2, . . . , T ; j = 1, 2, . . . , k (7)
Random effect models

▶ If we believe that ai is uncorrelated with the explanatory

variables, the βj can be consistently estimated by using a
single cross section: there is no need for panel data at all.
▶ But using a single cross-section disregards much useful
information in the other time periods.
▶ Define the Composite error term as

vit = ai + uit (8)

▶ Equation 6 becomes

yit = β0 + β1 xit1 + . . . + βk xitk + vit (9)

Random effect models

▶ Because ai is in the composite error in each time period, the

vit are serially correlated across time.
▶ Let

σa2 = Var(ai )
σu2 = Var(uit )

▶ under the random effects assumptions

σa2
Corr(vit , vis ) = , t ̸= s (10)
σa2 + σu2
Random effect models
Define
1
σu2

2
λ=1− 2 2
σu + T σa
then
0≤λ≤1 (11)
Then the transformed equation turns out to be

yit −λȳi = β0 (1−λ)+β1 (xit1 − λx̄i1 )+. . .+βk (xitk − λx̄ik )+(vit −λv̄i )
(12)
where the overbar again denotes the time averages.
▶ The random effects transformation subtracts a fraction of
time average ȳi , where the fraction depends on σu2 , σa2 , and
the number of time periods, T .
▶ Then, we can use the generalized least square (GLS) to
estimate the parameters.
▶ It can be proven that the equation 12 are serially uncorrelated.
Random effect models

▶ Since REM only substracts a fraction of the time averages, it

allow the presence of the variables which are constant over
time in the equations.
▶ A more thoughtful explanation is the FEM assume the
unobserved effect ai are uncorrelated with any explanatory
variable whether or not it’s constant over time.
Random effect models

▶ λ is a constant to be estimated:

 1

 
 2

 

 1 
λ̂ = 1 − ! (13)
 σˆa2 
1 + T ·

 

 
σˆ2

u

▶ Pooled OLS is obtained when λ = 0

▶ When λ̂ is close to 0 !
σˆa2
≈0
σˆ2 u

▶ This means σa is super small compared to σu .

▶ So, the unobserved effect ai is relatively unimportant
Random effect models

▶ FEM is obtained when λ = 1

▶ When λ̂ is close to 1
!
σˆa2
→∞
σˆ2
u

▶ This means σa is super large compared to σu .

▶ So, the unobserved effect ai is relatively important
▶ The composite error can be decomposed as

vit − λv̂i = (1 − λ)ai + uit − λûi (14)

▶ the errors in the transformed equation used in random effects

estimation weight the unobserved effect by (1 − λ)
Example 14.4

▶ data: WAGEPAN
▶ Three methods
▶ pooled OLS
▶ random effects
▶ fixed effects
▶ In the first two methods, we can include educ and race
dummies (black and hispan), but these variables are drop out
of the fixed effects analysis.
Fixed effect or random effect ?

Hausman test
▶ Null hypothesis: Random effect model
▶ Alternative hypothesis: Fixed effect model.
▶ If the p-value of the Hausman test is less than the significance
level, then reject the null hypothesis and use the Fixed effect
model
▶ Essentially, the tests looks to see if there is a correlation
between the unique errors and the regressors in the model.
The null hypothesis is that there is no correlation between the
two.
Practice: Exercises at the end of the chapter 14.