Chapter 14 Advanced Panel Data Methods

T t derrorterm complicate x y
it it
,... 2 , 1 ,
1
= + = |
is interpreted to mean that an increase in x of one unit leads to a prediction, in all cases,
that y will increase by units.
The emphasis is on in all cases:
In a panel, where does variation in X come from?
In other words, What variation identifies ?
I expect the same difference in y if
1. I observe two different subjects with a one-unit difference in x between them, and
2. I observe one subject whose x value increases by one unit.
For example, suppose y is income and x is lives in the South of the United States:
1. If I compare two different people, one who lives in the East (x1=0) and another
who lives in the South (x1=1), I expect the earnings of the person living in the
South to be lower because, on average, all prices and wages are lower in the
South. That is, I expect the coefficient on x1 will be less than 0.
2. On the other hand, if I observe a person living in the East (x1=0) who moves to
the South (x1=1), I expect that the earnings increased, or why else would that
person move? That is, I expect b1 will be greater than 0.
There are really two kinds of information in cross-sectional time-series data:
1. The cross-sectional information reflected in the changes between subjects
2. The time-series or within-subject information reflected in the changes within
subjects

Deciding which specific panel data model adopt requires thinking about kind of variation
in x to be used to IDENTIFY Ask what is the source of variation in my model that
drives this effect?

Are these two sources of variation in X, within and between, likely to give the same
effect on y?


14.1 Fixed Effects estimation

Suppose model is following:

(1) T t u a x y
it i it it
,... 2 , 1 ,
1
= + + = |

As in chapter 13, the concern is that the estimate of
1
| will be biased if x is correlated
with a
i
the fixed, unobserved characteristics.

First differencing was one way to eliminate these fixed unobserved components

An alternative (related) method is a FIXED EFFECT TRANSFORMATION

For each i, average over time for each individual
(2)
i i i i
u a x y + + =
1
|

Subtract (2) from (1)

T t u u x x y y
i it i it i it
,... 2 , 1 , ) (
1
= + = |

That is, we regress the individual-demeaned y on individual-demeaned x


What kind of variation then is this using to identify
1
| ? The WITHIN variationthe
variation in x over time for an individual. All average differences in ys or xs between
individuals have been wiped out.
Fixed effect estimator also called the WITHIN estimator.


Between estimator
This estimator is analogous, but here subtract the mean over time. Now
1
| is identified
by variation in ys, xs, between individuals, not over time for same individual.

When would that be useful? When have something about time period that is specific that
dont observe. Usually, the problem of individual specific error component is more
commonthat is,
i
x is often correlated with a
i


Will come back to the between estimator though when talk about random effectsthere
the idea is that if DONT have problem that
i
x is correlated with a
i
then can use both
sources of variation and get a more efficient estimator.




Dummy Variable Regression

One way to view this model is to think of a
i
as a parameter to be estimated for each
individual. Way we do this is to put in a dummy variation for each i (person, firm, state,
etcwhatever the unit is that we observe over time).

What will be the value of the fixed effect? Mean for that group.

This give us EXACTLY the same estimates of the s, their standard errors, etc. as using a
demeaned transformation.


Fixed Effects or First Differencing?

In last chapter we also talked about differencing the data. That also dealt with
unobserved effects. (Instead of subtracting the mean, we subtract one period from the
other.)

What is the difference?

T=2no difference in the estimated coefficients.










T=3+ The two methods will not give identical coefficients. However, both estimators are
unbiased, consistent

Large N, small T
Which model choose depends on structure of errors over timeis there serial correlation
in the u
it
idiosyncratic errors?
No: fixed effects more efficient than first difference estimator
Yes: first differencing may be betterthe Au
it
may have less autocorrelation

T large, N small

Fixed effect estimator--inference sensitive to violations of assumptions with small n
Use first differencescan appeal to CLT because of large T

`
) (
) )( (

2
1
2
1
1
X
XY
n
i
i
n
i
i i
s
s
X X
Y Y X X
=

=
=
|
X Y
1 0

| | =
Bottom line: Often want to check both and see if results are differentspend time
looking at structure of errors over time if arewill discuss more in time-series chapters


What about missing data?

Often in panels, have an UNBALANCED panelmissing data on some individuals in
some years. Dummy variable/fixed effect regression still works fine, although note that
any individuals with only 1 observation get dropped.

If attrition or reason are missing is randomor at least uncorrelated with u
it
, then not a
problem. However, if IS related to u
it
, then can lead to biased estimates. Will discuss
models to deal with selection later.


Comparison with DD Model

Like with DD models, FE model control for time constant differences in means. FE
models control for any permanent, unobserved variables

Like with DD models, are often concerned about differences in trends in unobserved
variables.

Several ways to deal with that.
- Difference data over time a second time. This will subtract any
unobserved/omitted variables that have a constant trend. See Hoxby paper for an
example.

- Include interaction between individual fixed effect and a trend variable


This is commonly used in DD style papers that use states or areas as the unit of
observation

y
s,t
= + X
s,t
+
s
Time +
s
+
t
+
s,t


Note that
s
is a vector of s different coefficients on timeone for each state.
This model still also includes state fixed effects and year fixed effects. In
practice, sometimes papers will choose between state specific time trends and
year fixed effects.
Strengths and Weaknesses of First Differenced/Fixed Effect Models

STRENGTHS: Controls for unobserved, time invariant effects that are correlated with
error. A huge advantage when omitted variable bias is an issue.

WEAKNESSES:

1. Amplify Measurement error in x If x is not measured perfectly, have a noisy
measure a noisy measure much of the variation in x
i
may be due to
measurement error, rather than true underlying variation. (Amplifies the ratio of
noise to signal)

What is effect of measurement error? Attenuated coefficient

See proof for thisdo with FD model


2. Cant estimate effect of permanent characteristics.

In this model, can only obtain estimates of things with a x
i

For example, being in the South may lead schools to have different lower test scores.
Cant estimate this effect because there is no change over time: x
i
=0

Similarly, cant include a dummy variable for each state and state level, permanent
characteristicswhy? Perfect collinearity.

Often have to be careful about this in fixed effect models. Dummy variable trap
can be harder to recognize on occasion. Check for dropped variables and identify
reason drop.

3. Less variation in differences. Many times there is less variation in changes over time
that there is across individuals in a cross section. More variation in levels of
unionization across districts than variation in how much unionization changed.
What is the consequence? Little variation in xlarger standard errors for |
1

4. Source of variation for estimation is less clear

Fixed effect estimation removes some of the variation since subtracting the mean
differences across unit of observation



14.2 Random Effects Models

Recall original model:

(1) T t u a x x y
it i itk k it it
,... 2 , 1 , ...
1
1 0
= + + + + + = | | |


Again use fixed effects if a
i
is correlated with x
itj


In other words, going back to our within and between discussion, we thought that people
who worked for non-profits were different from those that didnt. As a result, we wanted
to look at the within- variation onlythat means that our fixed effect estimator was
identified by people who SWITCHED from one sector to the other. NOT by variation
across individuals.

But what if we thought wasnt the case that a
i
was correlated with Xitj

In that case want to use all the variation to get more efficient estimates.

Two ways to think about random effects models:

- Random effects model is a matrix weighted version of the between- and the
within-(fixed effect) estimators.
- Random effects model is a GLS version of Pooled OLS model, accounting for
fact that errors are serially correlated

Random effects model key assumption:

cov(x
itj
, a
i
) = 0, t=1, 2, . . . .T; j=1,2,,k

Note that either using single cross section or pooled data will give us consistent estimates
of betas.

However, doesnt exploit all the variation if use only cross section.

Not going to derive the random effects estimator.

But again, think about combining the variation within an individual over time and the
variation between individuals at a point in time. How do we combine these two sources
of variation? Weighted average. Weight is this:

2 2
2
1
a u
u
To o
o

+
=

Recall FE estimator is this:
T t u u x x x x y y
i it
k
i
k
it k i it i it
,... 2 , 1 , ) ( ... ) (
1 1
1 0
= + + + + = | | |


R.E. estimator is this:
) ( ) ( ... ) ( ) 1 (
1 1
1 0 i it
k
i
k
it k i it i it
v v x x x x y y | | | + + + + =


Comparing RE estimates and FE estimates:
=1 for fixed effect. As quantity under square-root sign approaches zero=1
- T is big lots of variation across time for each individual more like fixed
effects
- o
2
a
is biglots of variation in the fixed effects more like fixed effect
estimate
- o
2
u
is small relative to o
2
a
idiosyncratic variation is smallmore of the
variation is from fixed effect


Summary of RE:
- Random effects estimators are a weighted average of the between estimator
(variation between individuals in a cross section) and the within/fixed effect
estimator (variation within individuals over time)
- Random effects estimators will be consistent and unbiased if fixed effects are
not correlated with xs. Fixed effects estimators will always be consistent and
unbiased (under usual GM assumptions)
- Random effects estimators will be more efficient (have smaller standard errors)
than fixed effects estimators because they use more of the variation in X
(specifically, they use the cross sectional/between variation)






14.3 Applying Panel Data methods to Other Data Structures


One common application is a cluster sample For example, may have a
i
that are
correlated within the same cluster (e.g., students in same school)
Solution is to use panel-robust/cluster/sandwich standard errors (different names)

Also, with fixed effect models, are not accounting to the potential autocorrelation within
the errors for each i. Very common to also use clustered standard errors with fixed effect
models.


Basic overview:

With random effects, use precise distributional assumptions
- within-school correlation takes the same form for all schools,
- each pupil within a school is correlated equally with any other pupil in the
school.

With clustered standard errors don't model within school correlation explicitly
- allow for arbitrary correlation within schools,
- the form of this correlation can vary from school to school.

Trade offs:

- Random effect give more efficient estimates if modelling of the correlation
caused by clustering is correct. If it isn't, coeffs and SEs are wrong. (recall RE
relies on zero covariance between ai and xits)

- With clustered standard errors, get consistent estimates across a broad range of
possible forms of the correlation, but they won't be as efficient when you know
the exact form.



In STATA, reg yvar xvar, cluster(school)
Panel Robust Sandwich Standard Errors/Cluster-Robust standard errors


Version of robust standard errors in a panel context

In a panel context, very likely that Cov[uit, uis]>0 for t s.
If ignore this serial correlation, will greatly underestimate standard errors and
overestimate t-stats

May have already done some panel transformation (f.d., f.e., r.e.) to get the transformed
variables:



Notation:
Stacking observations over time periods for a given individual



Then stack by the N individuals



Three different ways we can write the OLS estimator:




Most convenient to use varies with contest

When with this estimator be consistent?




Assume independence over I. Now the exogeneity condition for consistency is

]=0

This is a stronger assumption than ]=0

(Discussed this with serial correlation?)

The panel-robust estimate of the asymptotic variance matrix (that is, one that controls for
both autocorrelation and heteroskedasticity) is analogous to the usual robust standard
errors:



where the are the usual predicted errors

Again, this assumes independence over i, but permits V[uit] and Cov[uit, uis] to vary
with i,t,s


An equivalent expression is


So if use traditional robust standard errors, will adjust for heteroskedasticty, but not
autocorrelation. In panel setting, problem of autocorrelation is much bigger.
Summary:

Pooled OLS is
true model
it it it
u x y + =
1
|

Random Effects is
true model
it i it it
u a x y + + =
1
|

a
i
not correlated
with x
i

Fixed Effects is
true model
it i it it
u a x y + + =
1
|

a
i
correlated with x
i

Pooled OLS Consistent
Efficient
Consistent Inconsistent
Within (Fixed
Effect)
Consistent Consistent Consistent

Usually most
efficient, especially
with large N, small
Tlikely to require
clustered errors to
get correct se
First Difference Consistent Consistent Consistent

Can be more
efficient than F.E.
with small N, large
T, depending on
structure of
autocorrelation
Random Effects Consistent Consistent
Efficient
Inconsistent
Testing for Which Model to Use

1. Choosing Between Fixed Effects and Pooled OLS

Is it appropriate to include set of dummy variables?
Use F-test to decide, just as would with any set of variables


2. Compare Between Random Effects and Pooled OLS

If use OLS on pooled data, have T*n observationscan overstate precision of
estimatesreally have same n people observed T times. Need to account for serial
correlation in errors

Let Vit = ai+uitperson specific error and time specific error

Look at correlation in this over time for same individual:

corr(vit, vis) = o
2
a
/(o
2
a
+o
2
u
)

This correlation over time can be substantial OLS stats for pooled model ignore this.
Need a GLS estimator (remind me what this is?)

Breusch Pagan LM test for Unobserved heterogeneity

Is there unobserved heterogeneity? Ho: o
2
a
= 0

NOTE: This is NOT asking if the person specific error (the unobserved heterogeneity) is
correlated with the Xs. This is only asking IF IT EXISTS. How do we know it exists?
Because the errors for an individual are related.

One way to approachtest for serial correlation in the vi = ai + uit

or we can do a test directly based on o
2
a


in Stata, do xttest0

3. Choosing Between Random Effects and Fixed Effects

Why might we consider random effects instead of fixed effects?

1. When are looking at a variable that doesnt vary over timecant use a fixed
effects estimate.

For example, if were looking at effect of gender on wages, couldnt use a fixed
effect model. Could include random effects if had a panel. Better than OLS,
although note that if key variable of interest is not determined exogenously, will
still have biased coefficients

2. When are sure that x is exogenously determinedsay by an experiment

3. If Hausman test indicates that rejection of fixed effects is appropriatebut be
careful

Hausman test

This is a test we will see again in IV contextthe structure is the same

Need two estimators to perform a Hausman test

1. An estimator (R.E.) that is consistent AND efficient under HO (no correlation)
and inconsistent under HA
2. An estimator that is consistent under both (F.E.)

Basically look at how different the two estimates areif they are really different,
likely FE is more appropriate.

Test statistic:
2 1
~ )

( )]

( )

( [ )'

(
k RE FE RE FE RE FE
V V _ | | | | | |



If have just one X, this is just a t-test for the difference in the coefficients

( )
2 2
)

( )

( /

RE FE RE FE
se se H | | | | =

Problem is may not be really different because even though point estimates are pretty
different if sample variation in FE is so big cant say are statistically different. But thats
not very satisfyingbetter to have a good argument based on theory for why arent
worried about correlation between xs and ai, and then use Hausman test to back you up.

Note: this is NOT about whether ai is fixed or random. It is about whether ai is
correlated with xitj