Question about binary outcome using FE model

[MB Ross]

This is feeling like more of a silly question by the moment but here it goes…

I am using `xtreg` to run a linear probability model of job characteristics on the likelihood of transitioning from employment to unemployment. The outcome variable is obviously a binary variable that equals 1 if the person is unemployed and 0 otherwise. The question I am having is about whether I have set everything up correctly in terms of the outcome variable.

Specifically, in periods where the outcome variable is equal to 1 then there will be no entry for the job characteristic. If that is the case, Stata would then disregard this observation, right? If I’m thinking about this correctly, and I’m not sure I am, then I would need to make the last month of employment before entering unemployment have a 1 for unemployed because then I will actually have an observed job characteristic. Thoughts?

[Clyde Schechter]

Perhaps the problem is that you are using the wrong outcome variable. If what you are trying to model is the transition to unemployment, then it seems you want a variable that is coded 1 if the person is unemployed this month but was employed the previous month, and 0 otherwise.

Which leads to another thought about a different approach to modeling. Presumably the job characteristics don't change much, if at all, in the time leading up to the transition to unemployment. It might make more sense to have a variable counting the number of months of employment in the job before becoming unemployed as the outcome variable, with predictor variables representing the job characteristics, and do a survival analysis. That seems a more natural way to model this, to me.

[William Lisowski]

I agree with your end result, but I think about it slightly differently.

My approach would be to use an outcome in period t that is 1 if the person is unemployed in period t+1, 0 if s/he is employed in period t+1. (Which is what you're suggesting, effectively, but expressed in a way that sounds less ad hoc, because it pertains to every period.)

So in a population limited to employed people (because of the loss of observations when job characteristics are missing) I would be fitting a model of "becomes unemployed between period t and period t+1" to "job characteristics in period t", and any other background characteristics to control for, measured at period t (or earlier).

Again, my approach leads to the same place as yours, just using what I think is a more standard explanation.

My post crossed Clyde's. His coding will rely on using the lagged value of the job characteristics, getting them from the previous period, for the model I described. Doing so will automatically exclude the observations from succeeding periods when unemployment continues, which I would want to do in any event since being unemployed precludes transitioning to unemployment. But if I were including other characteristics measured in the period when unemployment began (e.g. lives somewhere different than the previous period) then Clyde's coding is perhaps more convenient. Depends in a sense on whether the objective is predicting transition to unemployment in period t+1 based on characteristics measured in prior periods, or explaining transition to unemployment in period t based on charcteristics measured in that period and prior periods.

[MB Ross]

I think both these suggestions are valuable. I have already run the survival analysis and will include that in the paper but wanted to use the fixed-effects model as a precursory estimation strategy. The strategy suggested by William is what I had been thinking but I needed the affirmation that this was the correct way of thinking about things. Thanks to you both!