[Rodrigo Primor]

But for survival analysis, i cant have left censored data, right? Im saying this because i dont know when was the foundation of the companies in my dataset and i only have data since 2011).
In the literature that i have read (e.g. ohlson, 1980), they simply use a panel logit model, so survival analysis is the only way?

[Steve Samuels]

I apologize: survival analysis is the wrong approach here. I mistook your setup for a survival problem because it superficially looked like one: there's one failure per firm, for example.

So by all means use panel binary regression. In Stata, the major commands are xtlogit, xtprobit, and xtcloglog. xtlogit comes in three varieties: fixed effect (fe) (clogit), random effects (re), and population-averaged effects (pa). With just one outcome per firm, I would guess (but don't know) that the re models won't work well. There are many panel-data experts on Statalist who can offer better advice.

I don't have specialized knowledge of panel models, but you can use linktest after any model, and Paul Allison describes a kindred test by Stukel in this document. Both require an initial model, then test for the significance of the square of the linear predictor. Predicted probabilities after xtlogit, fe don't actually describe the unconditional probability of an event. With just one event per firm in your data, the pc1 prediction might have value in assessing fit.

One suggestion: have a sufficient number of events to consider models with interaction and polynomial terms. With fixed effects, you can't add time-constant terms to the model, but you can interact those terms with the time-varying ones. You could even try fractional polynomials (fp).

[Rodrigo Primor]

Thanks for the advice but i think that survival analysis can also be applied in this case. For example, Shumway (2001) and Bonfim (2009) used an hazard model to predict bankruptcy

The papers are : Forecasting Bankruptcy: A simple hazard model - Tyler Shumway
CREDIT RISK DRIVERS: EVALUATING THE CONTRIBUTION OF FIRM LEVEL INFORMATION AND OF MACROECONOMIC DYNAMICS - Diana Bonfim

If you can take a look and confirm the use of hazard models i would be gratefuf.

And one more question: If set data as time-series using tsset but with an ID panel variable specified, it's equal to xtset?

[Steve Samuels]

I'm sorry, but I don't have access to those papers nor time to look them up (FAQ 12). I did see the Shumway paper years ago and found nothing original in it, just a rewording of well-known theory. I don't remember how it dealt with the issue of experience prior to time zero. . The answer to your second question is "Yes".

[Rodrigo Primor]

Thanks Steve, i think i'm more aware of what i need to do

Best regards