# Panel Data: In a fixed effects model, does auto-correlation introduce bias?

Given a panel of countries over time, a fixed effects estimator makes sense to control for country-specific effects. My intuition tells me that if the dependent variable is correlated with lags of the independent variables, then bias will be introduced into the estimator. However, I'm having difficulty rigorously understanding why this would be the case. Additionally, is there any easy way to tell if the bias would be towards or away from zero?

-
Wooldridge in his book gives a very detailed explanation of this. Look in to chapters 10 and 11. In short the key is correlation of lags of independent variables with the error term, not with dependend variable. If the independent variables are strictly exogenous then there is no bias. If they are not the bias is of order 1/T, assuming of course that contemporaneous model holds. –  mpiktas May 5 '11 at 12:05