4 Violations of Assumptions
4.1 Assumption 2 in Panel Data
Panel data is characterized by time dependency for each panel unit.
As discussed in Week 6, this is a violation of the regression Assumption 2 (X and Y are i.i.d).
4.1.1 Serial Correlation
Time dependency is often described as autocorrelation or serial correlation.
The main approach to deal with serial correlation is by adjusting standard errors to take into account autocorrelation.
If there is substantial autocorrelation (serial correlation) in the error term, even heteroskedasticity-robust standard errors will be inconsistent.
In panel data as in any other time series data, autocorrelation can be a very serious concern.
We can test for serial correlation after our fixed effects estimation using the Breusch-Godfrey test.
The null hypothesis in this test is that the autocorrelation of the error term is 0.
4.1.2 Cross-sectional dependence
Cross-sectional dependence in panels may arise when e.g. countries respond to common shocks or if spatial diffusion processes are present (think Arab Spring, or shocks from the financial crisis).
If cross-sectional dependence is present, this results, at least, in the inefficiency of the estimators and invalid inference when using standard estimation techniques.
This is another instance of the violation of regression Assumption 2.
If we assume that our earlier two-way fixed effects model specification is consistent, then we can test for residual cross-sectional dependence after the introduction of two-way fixed effects to account for common shocks.
4.2 Panel-corrected Standard Errors
Panel-corrected standard errors (Beck and Katz 1995)
panel heteroskedasticity: each country may have its own error variance
contemporaneous correlation of the errors: the error for one country may be correlated with the errors for other countries in the same year
serially correlated errors: the errors for a given country are correlated with previous errors for that country
4.3 General Approach to Correlation Between Panels
Driscoll and Kraay (1998) propose an estimator producing heteroskedasticity- and autocorrelation- consistent standard errors that are robust to general forms of spatial and temporal dependence. Often known as the SCC estimator.
Panel Corrected Standard Errors (PCSE), while popular in political science, may not work well with shorter panels with large N (ratio of T/N is small).
SCC estimator performs equally well in large N settings.
4.4 Your Roadmap with Panel Data
Is it a fixed effects or random effects model?
Hausman test. But primarily the choice should be driven by theory!
Use robust standard errors, start with HAC.
Check whether there is any cross-sectional dependence:
If not, you can stick to HAC.
If you have cross-sectional dependence, you need to use PCSE or SCC (use SCC).