I am estimating a two-way fixed effects model, using panel data for 40 developing countries over a time period of 16 years.

I'm using an ordinary least squares regression including time-dummies and country-dummies (the command xi: reg in Stata)

Should I use robust standard errors? Why or why not?


Robust standard errors are valid only asymptotically. Since you appear to be doing cross-country analyses (with usually very few observations), you should use it only if you have heteroscedasticity in your data. Stata has a test for heteroscedasticity; the old command was hettest - run it right after your regression. If it comes back significant, then there is heteroscedasticity and your normal standard errors are biased. This is when you would usually start to use robust standard errors, but again, you should have a sufficient amount of observations to do this.

What can you do if you have heteroscedasticity but very few observations? Well, either get more observations, or use additional controls which might take care of the heteroscedasticity. If neither is feasible, maybe there are small sample corrections around.

| cite | improve this answer | |
  • $\begingroup$ When using 40 countries over a 16 year-period; is that too few observations? I have 8 controls plus one dummy variable. $\endgroup$ – Anna May 12 '13 at 9:32
  • $\begingroup$ No that should be enough; among economists that would qualify as "asymptotic" ;). But again, check first if you need robust SEs at all. $\endgroup$ – Nameless May 12 '13 at 10:33
  • $\begingroup$ Helpful read ch. 8.2.3 "Fewer than 42 Cluster" in Angrist and Pischke, Mostly Harmless Econometrics textbook $\endgroup$ – Steve Apr 8 '17 at 20:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.