Logit Model Problem

[Christian Roberto]

Hi. I have been searching the web for the past few weeks. I really appreciate if someone can help me.

I am doing logit regression right now and using panel data. I have a few question:
1. How to choose between fixed and random effect? Which one is more suitable to my model? Is there any test like Hausmann test in OLS to determine which effects are better?
2. I have tried to do the regression with xtlogit command. But i cannot run goodness of fit test using the 'estat gof'. But i can use the command after I regress with logit command. Any help? How to measure the goodness of fit in xtlogit? (Other command such as 'estat class' is also unable)
3. Can anyone help to intepret the correctly classified table? I have data 1 for company who do mergers and 0 who are not. But on the table, on the + row, the result is 0 for D column and 0 for D~ column, leaving the sensitivity 0%. Any advice?

Thank you for your great help. Your advice is really useful to me.

Cheers

[Carlo Lazzaro]

Christian:
welcome to the list.
I would start off with reading -xtlogit- entry in Stata .pdf manual, which reminds us that only conditional -fe- are avilable for -xtlogit.

[Clyde Schechter]

Adding to Carlo's good advice, with regard to your second question, you are asking for the impossible here.

-estat gof- following a logistic model does a Hosmer-Lemeshow model calibration analysis. When you run -xtlogit, fe-, you get a conditional logistic regression which provides relative odds of a positive outcome within a panel. But that analysis does not and cannot create predicted probabilities. Therefore it is impossible to calculate the predicted risks, nor the deciles of predicted risk that underlie the Hosmer-Lemeshow procedure. Similarly, you would not be able to generate a classification table following -xtlogit, fe- for the same reason: there are no predicted probabilities.

[DILAWAR AHMAD BHAT]

Clyde Schechter data-cke-saved-href="#" href="#" class="b-bbcode-user js-bbcode-user">Clyde Schechter E=Clyde Schechter;n1341281]Adding to Carlo's good advice, with regard to your second question, you are asking for the impossible here.

-estat gof- following a logistic model does a Hosmer-Lemeshow model calibration analysis. When you run -xtlogit, fe-, you get a conditional logistic regression which provides relative odds of a positive outcome within a panel. But that analysis does not and cannot create predicted probabilities. Therefore it is impossible to calculate the predicted risks, nor the deciles of predicted risk that underlie the Hosmer-Lemeshow procedure. Similarly, you would not be able to generate a classification table following -xtlogit, fe- for the same reason: there are no predicted probabilities.[/QUOTE]

@ Clyde , How can we get a classification table in panel logit model? is there a way to generate probablities using xtlogit?

[DILAWAR AHMAD BHAT]

Christian:
welcome to the list.
I would start off with reading -xtlogit- entry in Stata .pdf manual, which reminds us that only conditional -fe- are avilable for -xtlogit.

@ Carlo, How can i get a classification table in panel logit model? is there a way to generate probablities after using xtlogit?

[Rich Goldstein]

see

help xtlogit postestimation##predict

[Clyde Schechter]

Re #4: No, that is impossible. The -xtlogit- results are conditional on base probabilities and actual probabilities therefore cannot be estimated. If you follow Rich Goldstein's advice, you will see that you can get predicted probabilities conditional on the assumption that the fixed effect is zero, or conditional on the assumption that there is only 1 positive outcome per panel. Those are not what one usually bases a classification table on, but depending on what you intend to do with the results, they might be suitable for particular purposes.

Another possible approach is to do -xtlogit, re-. From there you can get predicted probabilities and create a classification table. So if a random effects model is allowable in your circumstances, try this.