Econometrics Chapter 14
Absolutely — I went through your uploaded Chapter 14 slides on Advanced Panel Data Methods and turned them into deep study notes. The chapter mainly covers Fixed Effects, First Differencing, Random Effects, Correlated Random Effects, policy analysis in panels, feedback tests, and heterogeneous trends.
Chapter 14: Advanced Panel Data Methods — Deep Notes
1. Big picture of the chapter
This chapter is about how to deal with unobserved differences across individuals, firms, states, countries, etc. when you have panel data.
Panel data means you observe the same unit over multiple time periods.
A standard panel model is:
Where:
= outcome for unit i at time t
= explanatory variables
= unobserved time-constant effect for unit i
= idiosyncratic error
Why is important?
Because it captures all the things about a unit that do not change over time, but may still affect the outcome.
Examples:
ability of a worker
culture of a firm
natural advantages of a state
family background of a person
If is correlated with your regressors, pooled OLS becomes biased.
That is why panel methods are powerful: they try to remove or handle .
2. Fixed Effects (FE)
The chapter starts by saying FE estimates the time-demeaned equation by OLS, and it uses within-unit variation.
Core idea
FE removes by subtracting each unit’s time average from each variable.
Original model:
Take the time average for each individual:
Subtract:
Notice: disappears.
That is why FE is also called the within estimator.
Intuition
FE asks:
For the same person / firm / state, when x changes over time, how does y change?
So FE does not compare different people to each other.
It compares each unit to itself over time.
That is the single most important intuition.
3. Example in the slides: training grants and firm scrap rate
The slides mention FE using 1987, 1988, and 1989 for the effect of training grants on firm scrap rate.
Interpretation
Suppose some firms are just better run than others. That managerial quality might affect scrap rate and also affect whether they seek grants.
If we compare different firms directly, that unobserved quality biases the estimate.
With FE, we instead ask:
when a given firm gets a training grant,
does its scrap rate fall relative to its own usual level?
So FE controls for time-constant firm quality automatically.
4. Important discussion points about FE
The chapter lists several important properties and limitations of FE.
(a) Strict exogeneity is required
The original model must satisfy strict exogeneity.
That means:
for all t.
Meaning
The error in period t should not be related to explanatory variables in any period.
This is stronger than just saying current is exogenous.
Why it matters
If future values of x react to current shocks in u, FE can fail.
Example:
a bad shock this year raises poverty
government increases aid next year in response
Then future policy depends on current error. That breaks strict exogeneity.
This becomes important later in the chapter’s feedback test.
(b) R-squared from the demeaned regression is not appropriate
The slides say the R-squared of the demeaned equation is inappropriate.
Why?
Because FE works on transformed data, after removing time means. So that R-squared is measuring fit in the demeaned world, not the original outcome variation.
So do not interpret it like normal OLS R-squared.
(c) FE cannot estimate time-invariant variables
The slides say the effect of time-invariant variables cannot be estimated.
Examples:
gender
race
place of birth
region that never changes
innate ability if treated as fixed
Why?
Because FE subtracts the time average. If a variable never changes over time, then:
So it disappears completely.
This is very important for exams:
FE is great for removing omitted variable bias from fixed unobservables, but the price is that you lose coefficients on time-constant regressors.
(d) Interactions with time-invariant variables can still be estimated
The slides say interactions with time-invariant variables can be estimated, like education interacted with time dummies.
Why?
Because although education itself may be constant, something like:
changes over time through the time dummy D_t.
So FE can estimate how the effect of a time-invariant characteristic varies over time.
(e) Full time dummies create another limitation
If you include a full set of time dummies, then variables whose change is mechanically constant over time cannot be estimated, such as experience in some wage panels.
Why?
Suppose experience rises by exactly one every year for each person. Then it moves exactly with time itself.
If you already include a full set of time dummies, there is no separate variation left to identify experience.
This is a classic multicollinearity / lack of identification issue.
(f) Degrees of freedom adjustment
The chapter says degrees of freedom become:
NT - N - k
because we estimate N time averages in addition.
Intuition
In FE, you are effectively estimating one intercept per unit, or equivalently removing one mean per unit. That costs degrees of freedom.
5. FE as dummy variable regression
The slides explain that FE is equivalent to running pooled OLS with a dummy for each individual.
This means:
where each is a dummy for a specific unit.
Why this matters
This gives a second interpretation of FE:
Demeaning view: remove each unit’s average
Dummy-variable view: allow each unit to have its own intercept
These are mathematically equivalent.
Intuition
Every unit gets its own baseline level. So firms can have different baseline scrap rates, workers can have different baseline wages, states can have different baseline poverty rates.
That is exactly what a_i represents.
6. FE vs First Differencing (FD)
The slides next compare FE and FD.
First differencing idea
Instead of subtracting the unit mean, FD subtracts the previous period:
Again, a_i disappears because it is constant over time.
Key results from the slides
(a) If T = 2, FE and FD are identical
This is a famous result.
With only two time periods:
demeaning from the average
differencing between the two periods
lead to the same slope estimates up to a scaling factor.
So for two-period panels, FE and FD are basically the same estimator.
(b) If T > 2, FE is more efficient under classical assumptions
The slides say FE is more efficient when classical assumptions hold.
Meaning
If the idiosyncratic errors are well behaved, FE tends to use the information more efficiently than FD.
So in ordinary settings, FE is often preferred.
(c) FD may be better with severe serial correlation
The slides say FD can be better when errors have severe serial correlation, for example a random walk.
Why?
If the error process is highly persistent, differencing may better remove that persistence.
So if errors evolve like:
then differencing can stabilize things more effectively.
(d) If T is very large and N is not large, use caution
The slides say when T is very large and N is not, the panel starts looking like a time series, and dependence problems become important. In such cases FD may be better.
Intuition
A long panel is not just “many repeated cross sections.” It has a real time-series character:
serial dependence
persistence
trends
nonstationarity-type issues
So method choice becomes more subtle.
(e) Good practical advice: compute both
The chapter says it is often a good idea to estimate both FE and FD and check robustness.
This is good exam and real-world advice.
If both tell a similar story, that builds confidence.
7. Unbalanced panels
The slides define an unbalanced panel as one where not all units have the same number of observations.
Example
Some people are observed for 5 years, some for 3, some for 2.
This happens a lot in real data:
missing surveys
firms entering and exiting
states with incomplete reporting
people dropping out
Important result
Dropping units with only one time period does not create bias or inconsistency.
Why?
Because FE and FD need time variation. With only one observation, you cannot identify within-unit change anyway.
So those units contribute nothing to FE/FD estimation.
FE vs FD in unbalanced panels
The slides say FE preserves more data than FD.
Why?
FD needs both t and t-1 to form a difference.
If some years are missing, differencing may throw away many observations.
FE only needs enough observations to compute means and within variation, so it can often keep more usable data.
The slide example
If you observe years 1, 3, 5, 7 only:
FE can still use those years
FD loses everything because there are no adjacent t and t-1 observations
This is a very testable conceptual point.
8. Random Effects (RE)
The chapter then moves to RE. It says the composite error is uncorrelated with the explanatory variables, but is serially correlated within unit i.
Core RE assumption
RE assumes:
for all t.
This is much stronger than FE.
Interpretation
The unobserved individual effect is treated like a random component that is unrelated to the regressors.
Example where RE might be plausible:
randomly sampled classrooms where permanent classroom effect is unrelated to a randomized treatment
But in economics, this is often hard to believe.
Why pooled OLS is consistent under RE
The chapter says if RE assumptions hold, pooled OLS is consistent.
Why?
Because the composite error is uncorrelated with regressors.
But there is still a problem: the error is serially correlated within unit, because appears in every period.
That means ordinary OLS standard errors are wrong unless corrected.
Clustered standard errors
The slides say if OLS is used, standard errors must be adjusted for within-unit correlation using clustered standard errors.
This is extremely important.
Intuition
Observations from the same person or firm across time are not independent.
So standard errors must allow arbitrary correlation within each cluster i.
OLS is consistent but not efficient
The chapter says pooled OLS is not efficient because of serial correlation.
So we want a transformation that accounts for the error structure.
That leads to random effects estimation, which is a form of FGLS.
9. Quasi-demeaning and RE estimation
The slides say the quasi-demeaning parameter is estimated, and FGLS using this estimated is called RE estimation.
Intuition of quasi-demeaning
RE transforms the data like this:
This is neither:
full demeaning like FE (\lambda = 1)
nor no demeaning like pooled OLS (\lambda = 0)
It is somewhere in between.
Meaning of
If the random effect is small
Then , and RE becomes close to pooled OLS.
If the random effect is large
Then , and RE becomes close to FE.
This is a beautiful intuition:
little unit-specific heterogeneity → pooled OLS
lots of unit-specific heterogeneity → FE-like transformation
Major advantage of RE
The slides say RE can estimate the effects of time-invariant variables.
That is one of the biggest practical reasons people like RE.
Examples:
gender
race
region
education if it does not change much
policy regime fixed over time
FE cannot estimate these directly; RE can.
But only if the RE assumption is believable.
10. FE or RE?
The slides clearly say that in economics, unobserved individual effects are seldom uncorrelated with explanatory variables, so FE is usually more convincing.
This is the main practical message.
Why FE is often preferred
Think of wage equations:
ability affects wages
ability also affects education, training, job choice
So is likely correlated with regressors.
If that happens, RE is inconsistent.
FE remains valid under correlation between and , as long as strict exogeneity holds for the idiosyncratic error.
Rule of thumb
Use FE when you think unobserved fixed traits are correlated with regressors.
Use RE only when you can defend no such correlation.
11. Correlated Random Effects (CRE)
The slides then introduce Correlated Random Effects (CRE) and say one advantage is that it allows estimation of time-constant explanatory variables, which FE cannot do.
This is a very important concept.
Big idea of CRE
CRE is a way to relax the unrealistic RE assumption by allowing the individual effect a_i to be related to the regressors through their unit means.
A common CRE/Mundlak form is:
Substitute into the model:
Then estimate by RE or pooled OLS with the group means included.
Why this is powerful
CRE separates two things:
within effect: how changes inside a unit affect y
between effect / correlation structure: how the unit’s long-run average level of x is related to the unobserved effect
So CRE lets you:
get the FE-type within interpretation,
estimate time-invariant regressors,
test whether RE is too restrictive.
Practical intuition
Suppose education is correlated with unobserved ability.
RE says: no problem, assume no correlation.
FE says: remove ability completely, but then lose time-constant variables.
CRE says: let the fixed effect depend on the average regressors, and model that correlation directly.
That is why CRE is often seen as a bridge between FE and RE.
Important slide point
The slides say when using CRE to choose between FE and RE, we must include any time-constant variables that appear in the RE model.
That matters because if RE includes time-constant regressors, the CRE comparison must include them too for a fair test.
12. General policy analysis with panel data
The chapter then shifts to policy evaluation in a general panel setting and says the two-period before-after model is just a special case when .
Core idea
Suppose:
= outcome
= policy or treatment variable
Then estimate a model like:
using FE or FD.
The slides say if may be related to the fixed effect , estimate using FD or FE with cluster-robust standard errors.
Why this matters
This is the panel version of causal policy analysis.
Instead of just one before and one after period, we may have many periods and changing policy intensity.
Examples:
welfare spending by state over time
minimum wage across states and years
pollution rules across counties
school funding across districts
FE controls for time-invariant differences across units.
Including lags
The slides say we can include lags of the policy intervention: w_{i,t-1}, w_{i,t-2}, \dots.
Why include lags?
Because policy effects may not happen immediately.
Example:
job training today may affect employment next year
infrastructure spending today may affect growth two years later
So lags help capture delayed effects.
13. Testing for feedback from the error term to the policy variable
This is one of the most important conceptual parts of the chapter.
The slides warn that policy variables may react to past shocks, and this creates problems.
The problem
Suppose:
y_{it} = poverty rate
w_{it} = government assistance
If poverty suddenly rises because of some unexpected shock in year t, the government may increase assistance in year t+1.
Then future policy responds to current shocks.
That means:
is related to u_{it}.
This violates strict exogeneity.
The falsification test
The slides say if we have at least three time periods, we can test for feedback by including a forward policy variable. If it is significant, there may be feedback from the error term to policy. This is called a falsification test.
Intuition
Future policy should not affect current outcomes.
So if the coefficient on future policy is statistically significant, that is suspicious.
It suggests that future policy is picking up the effect of shocks already happening today.
That means policy is reacting to the same shocks, so your policy variable may be endogenous.
Very simple intuition
If tomorrow’s policy predicts today’s outcome, something is wrong with exogeneity.
This is a super useful exam explanation.
14. Heterogeneous trend model
The slides then ask: what if time trends are unique across individuals?
This is a deeper issue.
Problem
Even after controlling for fixed effects, different units may have different trends over time.
Examples:
one state has steadily rising wages
another has steadily falling wages
one firm is on a long-term improvement path
another is stagnating
If these trends are correlated with policy, basic FE can still be misleading.
Model idea
Instead of just:
we allow:
where g_i is a unit-specific trend.
So each unit gets:
its own intercept
its own slope over time
This is much more flexible.
Slide message
The chapter says this lets policy be correlated not only with level differences across units, but also with trend differences. It can be estimated by first differencing, and we need T \ge 3.
Why T \ge 3?
Because with only two periods, you cannot identify a unit-specific trend in a meaningful way.
You need enough time points to separate:
baseline level
trend
policy variation
Intuition
Regular FE handles different starting points.
Heterogeneous trend models handle different starting points and different trajectories.
That is a deeper form of control for unobserved heterogeneity.
15. Applying panel methods to other data structures
The last slide says panel methods can be used in other contexts where constant unobserved effects must be removed, and gives the example of wage equations for twins.
Why twins?
Twins share many fixed family and genetic characteristics.
If you compare twins, you can difference out those common fixed factors.
So even though it is not a standard person-over-time panel, the same logic applies:
remove constant unobservables
use within-pair differences
This is a very important general lesson:
Panel methods are not only about time. They are about removing fixed unobserved effects.
16. The whole chapter in one connected story
Here is the chapter as one logical flow:
Step 1: Problem
Pooled OLS fails when unobserved fixed factors a_i are correlated with regressors.
Step 2: FE solution
Use within-unit variation by demeaning. This removes a_i.
Step 3: FD alternative
Difference across time. Also removes a_i. With two periods, FE and FD are the same.
Step 4: Unbalanced panels
FE is often better than FD because it keeps more observations.
Step 5: RE
If a_i is truly uncorrelated with regressors, RE is more efficient and can estimate time-invariant variables.
Step 6: FE vs RE
In economics, correlation is usually likely, so FE is often safer.
Step 7: CRE
CRE allows correlation between a_i and regressors in a structured way, while still allowing time-invariant regressors.
Step 8: Policy analysis
Use FE/FD to evaluate policies over time with panel data.
Step 9: Feedback problem
Policy may respond to past shocks, violating strict exogeneity. Test this using future policy variables.
Step 10: Heterogeneous trends
Sometimes units differ not only in levels but also in trends, so the model must allow unit-specific trends.
17. Exam-style understanding: when to use what
Use FE when:
unobserved individual effect is likely correlated with regressors
you care about within-unit changes
time-invariant variables are not your main focus
Use FD when:
T = 2
you suspect strong serial correlation
you want a robustness check against FE
Use RE when:
you can justify that unobserved effects are uncorrelated with regressors
you need coefficients on time-invariant variables
you want more efficiency under valid assumptions
Use CRE when:
you want something more realistic than RE
you want to allow correlation between regressors and unobserved effects
you still want to estimate time-invariant variables
Use heterogeneous trends when:
units likely have different underlying time paths
policy adoption may be correlated with those trends
18. Most important points to memorize
Here are the highest-yield takeaways from the chapter:
FE removes time-constant unobserved heterogeneity by using within-unit variation.
FE cannot estimate time-invariant regressors.
FE and FD are identical when T=2.
For T>2, FE is usually more efficient, but FD can be better with strong serial correlation.
FE keeps more data than FD in unbalanced panels.
RE requires a_i to be uncorrelated with regressors.
RE can estimate time-invariant variables.
In economics, FE is often more believable than RE.
CRE is a smart compromise: it models correlation between a_i and regressors while preserving time-invariant variables.
In policy analysis, strict exogeneity can fail if policy responds to past shocks.
A significant future policy variable suggests feedback/endogeneity.
Sometimes FE is not enough because units may have different time trends.