Balanced vs Unbalanced Fixed Effect Model

[Colin McAweeney]

Hello,
I have a two datasets that I'll be analyzing with the fixed effect model and at least 4 independent variables (~500 and ~60 observations). However, there are outliers in both sets and when removed they become unbalanced. I am unsure how to proceed, should I keep the outliers in my dataset and the fixed effect model will account for them or should I run an unbalanced xtreg, fe? If it is unbalanced, are there other operations that need to be completed before xtreg, fe?

Thanks!

[Carlo Lazzaro]

Colin:
- as a general rule, the so called outliers should not be removed unless you're 100% sure that they mirror errors in data entry;
- Stata can handle both unbalanced and balanced panels with no problem.

[Phil Bromiley]

With your sample size, you can look carefully at each "outlier" and try to figure out what is going on.

While I agree with Carlo that we should be conservative about outliers, there are cases where outliers appear to reflect cases that differ from the population about which you wish to generalize. For example, data on firms will often have a very few observations with assets near zero generating return on assets of 10 or even 100 or more, when mean return on assets is normally around .10. If ROA is the dv, with squared error criteria, that observation of 100 will outweigh a very large number of observations of .05 or .1. I don't want my results to be completely determined by one or two observations.

I don't know of any great solutions to this problem. Any non-random dropping or recoding (a la winsorizing) of data creates problems in consistency. Reweighting and robust estimators likewise have drawbacks and may not have been programmed for some kinds of dependent variables/models. But, I am uncomfortable reporting large sample results that depend very heavily on one or a very small number of observations.