4
$\begingroup$

I am running a fixed effects regression with a very unbalanced panel data. There are a lot of large residuals. For half of my observations, the residuals are large. However, I do not want to simply remove them as the model is not statistically significant when omitting these observations.

If I just rely on added variable plots to look at only extreme outliers, I can't exactly tell which observation in the cluster is extreme, unlike in cross sectional data.

So I was thinking if I also check influence by something like Cook's distance.

But how do we identify influential observations in fixed effects regression. Is there a command like Cook's distance as in ordinary least squares.

Improve this question
$\endgroup$
3
  • $\begingroup$ Run your fe regression as a lsdv regression, which gives you access to what is implemented in ols $\endgroup$
    – Christoph Hanck
    May 1, 2015 at 7:28
  • 1
    $\begingroup$ "Like for half of my observations I get large residuals. So I do not want to simply remove them as the model will lose significance." The first point is not a reason to remove observations from an analysis. The second point is not a reason to keep observations in an analysis. $\endgroup$
    – AdamO
    Feb 9, 2018 at 17:46
  • $\begingroup$ Try DFBetas omitting clusters instead of single observations to assess the influence of a cluster of abnormal values rather than the influence of a single abnormal value. $\endgroup$
    – AdamO
    Feb 9, 2018 at 17:50

1 Answer 1

Reset to default
-1
$\begingroup$

Yes, you should be able to look at measures like Cook's Distance or Leverage. There is usually a command to obtain these in whatever statistical software you are using (you don't mention what you are using).

Improve this answer
$\endgroup$
2
  • 2
    $\begingroup$ Far better is to use statistical principles when specifying and fitting the model and to not "play with the data". One of the principles is choosing a model that is unlikely to be ruined by outliers so you don't have to worry about this so much, i.e., semiparametric regression models such as the proportional odds model for continuous $Y$. $\endgroup$
    – Frank Harrell
    Jun 2, 2018 at 11:57
  • $\begingroup$ @FrankHarrell no argument here. Agreed 100%. $\endgroup$
    – StatsStudent
    Jun 20, 2018 at 7:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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