1
$\begingroup$

I am performing a Logit Model on binary variable SALE, when running on stata some of my explanatory variables were dropped as they were "perfect predictors". after some reading i believe the problem to be that of separation. some authors propose Firths method to deal with this. I can install the stata extension and run the tests again no problem but i want to know:

  1. will this work for a relatively small dataset (n=207)?
  2. are there any downsides to this approach? (bias or overstatement of errors, anything like that)
Improve this question
$\endgroup$
1
  • 1
    $\begingroup$ No time to provide an elaborate response, but Greenland recommends log-F(1,1) prior instead of Firth's bias correction method. And you can implement it by simply modifying your data, adding a few rows of data to the data before analysis. See andrewgelman.com/wp-content/uploads/2014/09/… . Firth's method can actually lead to coefficients larger than the ML estimate in certain situations. $\endgroup$
    – Heteroskedastic Jim
    Commented Jul 29, 2018 at 21:15

1 Answer 1

Reset to default
0
$\begingroup$

Partially answered in comments:

No time to provide an elaborate response, but Greenland recommends log-F(1,1) prior instead of Firth's bias correction method. And you can implement it by simply modifying your data, adding a few rows of data to the data before analysis. See http://andrewgelman.com/wp-content/uploads/2014/09/Greenland-Mansournia.Firth-and-stronger-penalization.25June2014.pdf . Firth's method can actually lead to coefficients larger than the ML estimate in certain situations. – Heteroskedastic Jim

Improve this answer
$\endgroup$

Your Answer

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

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