Repeated observations in survival analysis

[Robert Bohler]

I am doing a survival analysis in Stata looking at one-year mortality outcomes after participation in a jail-based program. I have 350 unique individuals that attempted the program a total of 550 times and I am using a propensity-matched control group. For those who attempted the jail-based program more than once, my initial analysis only included the last program attempt, as I would think that including previous attempts as observations in the model would lead to "survival bias" since, by definition, these individuals would not have experienced the event ("death") to be able to attempt the program again. However, I have received feedback from a reviewer that, by not including all the treatment attempts, this leads to survival bias. This does not make sense to me. Am I making the right decision on the study design to only include the last program attempt? If not, should I account for repeated treatment attempts by just using the vce(cluster) option or should I be doing a multilevel survival model to account for observations nested within individuals? Any help would be greatly appreciated!

[Clyde Schechter]

I agree with the reviewer.

As for how to respond, you don't describe the particular survival analysis you have done. If your original analysis was parametric, e.g. based on -streg-, then I would go to a multi-level version of that using -mestreg- with a random intercept at the person level. If your original analysis was semi-parametric, based on -stcox-, then, as far as I know, there is no Stata program that will do a multi-level version of that. However, you can account for the repeated measures on the same person by including a frailty, using -stcox-'s -shared()- option.