Efficient estimation with large number of fixed effects in stcox

[Crystal Yang]

Hi,

I am running a proportional hazard model using stcox in Stata 13. I need to control for a large number of fixed effects (over 3000). I don't actually care about the estimation
of the coefficients on the fixed effects, so I thought something like areg (where fixed effects are absorbed and not estimated) might be
helpful - but it seems like this doesn't exist for stcox. Is there an efficient way to control for a large number of fixed
effects using stcox?

Thanks,
Crystal

[Steve Samuels]

Welcome to Statalist, Crystal!

stcox can't fit a fixed effects model. You can, however, model survival semi-parametrically via a piecewise-constant exponential model. See Germán Rodríguez's Princeton page on piecewise exponential models. I empirically demonstrated the theoretical equivalence of Poisson regression to a piecewise exponential, with an application to gsem here.

Then you can use xtpoisson, with the fe option. However, because this is conditional likelihood, your prediction options are limited. Better, I think is a random effects model:

1. Best: fit a multilevel mixed-effects piecewise-constant exponential model via mestreg in Stata 14. You then get options like stcurve.. For data preparation, very similar to that on Germán's page, see William Mason's class notes .

2. In Stata 13, fit the multilevel mixed effects with the equivalent Poisson model in gsem.

3. Fit a frailty model in stcox, which is a random effects model with a single Gamma distribution for the random effect. This is very limiting, compared to the mixed-effect options of mestreg and gsem.

[Steve Samuels]

I should have added a further advantage of mestreg: you might get lucky and find that your piecewise model is well-approximated by one of the built-in parametric distributions.

[Terry Ishitani]

Hello Dr. Samuel,

I have some questions on mestreg in Stata 14. Using mestreg, is it possible to run multilevel mixed-effects piecewise model with period specific effects? This particular procedure is written on page 123 in Event History Analysis with Stata by Blossfeld. Any foreseeable challenges and problems?

Thank you for your consideration in advance.