Fixed-Effects Conditional Logit Model

[Stephanie Atala]

I am fitting a fixed-effects conditional logistic model using panel data for 453 groups with a 0/1 outcome for each of 6 years. I ended up with a model with 10 predictors and a McFadden Pseudo R-square of 0.244. I was initially planning on using the Brier score but I am aware that with the FE logit, we cannot get predicted probabilities so how can I test if my model has a good predictability power? In other terms, what can I do with my model besides getting the Odds Ratios?

Any insight is appreciated.
Thank you!

[Phil Bromiley]

You'll increase your chances of a useful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

" what can I do with my model besides getting the Odds Ratios?" - plotting predicted values is usually a good idea with logit.

xtlogit with fe will give you predicted probabilities - look at xtlogit postestimation in the documentation.

[Stephanie Atala]

Thank you Phil Bromiley. I think the predicted probabilities given by Stata assume the fixed effects is 0. On another note, for a logit model to be valid, should the prediction from the model be better than a prediction done without the model (i.e., by looking at previous year outcomes without considering the predictors)?

[Richard Williams]

With xtlogit, fe, and also clogit, the margins default is "pu0 - probability of a positive outcome assuming that the fixed effect is zero"

As to why that matters -- some day I am going to try to really really understand Steve Samuels' posts in the following thread. My key takeaway is that margins or predict after clogit or xtlogit,fe is NOT straightforward.

https://www.statalist.org/forums/for...-after-xtlogit

[Joao Santos Silva]

Dear Stephanie,

Indeed there isn't much you can do with the output of the FE logit and certainly you cannot obtain meaningful predictions. See, however, here.

Best wishes,

Joao