Fixed effects bias in panel data estimators