Survival analysis with exposure at a defined period and follow up after a more prolonged period

[rollanders]

I am analysing risk för preterm birth after appendectomy. I have data for two cohorts - one with patients that were operated during pregnancy and a reference population that had no operation. I want to analyse the impact on risk of preterm birth related to the length of gestation when the study cohort was operated compared to the reference cohort at identical length of gestation. I will analyse the impact of the appendectomy at various weeks when the exposure occurred.

I have tried to do this comparison using stcox starting and ending the follow up at a specific length of gestation using these commands :

stset grdbs, failure(partuslt37w) origin(time 154) exit(time 195)
stcox i.appect
stdescribe if appect==1
stdescribe if appect==0
stsum, by(appect)

(grdbs is length of gestation in days starting at last menstrual period, partuslt37w is a binary variable identifying partus before 37 weeks (ie prematurity), appect is a binary variable coded 0 for the reference cohort and 1 for the appendicitis cohort. The reason I start at 154 days is that the register do not include births before 22 weeks which is normally regarded as the limit for viability. .

This works for identifiying the differnce in risk of premature birth between the two cohorts, but only during the same followup period as the inclusion. I want to compare the riskdifference of premature births for the included subset also the weeks that followed after the inclusion period. I guess I need to use stsplit and maybe a second stset??

I need help to formulate this command.

[Clyde Schechter]

I'm a bit reluctant to advise you on this because I'm somewhat confused by your explanation of the data. For example, while you explain why you start at 154 days, you never explain why you have specified -exit(time 195)-. That would normally imply that any observation in your data set that did not experience a failure by 195 days of gestation is no longer subject to follow up, hence is censored at time 195. I suppose that is possible, but it seems a rather strange study design: if you are looking to study premature birth, why would you terminate follow-up at 195 days of gestation when there is still ample time remaining for the event to occur. That said, one thing I can say with confidence is that -origin(time 154)- is incorrect. The -origin()- option is used to specify when the risk for the event begins. In your situation , that would be at day zero of the gestation. The proper representation of time 154 in your -stset- command would be -enter(time 154)-. The -enter()- option tells Stata at what point the observation of the subject begins. In principle, that can precede, follow, or be concurrent with the time that risk for the event begins. But in your situation observation clearly follows the onset of risk, and by a substantial amount.

I would also note from an epidemiologic perspective, there are many important risk factors for premature birth, and I suspect that some of those may also be associated with risk of appendicitis, in which case you could have confounding variables on your hand. You should look into this matter, and if my suspicions are born out you should modify your analysis in some way to account for that, perhaps by covariate adjustment, or perhaps by matching if the latter is feasible.

So far, I do not see any need for -stsplit-. However, if it turns out that you do have confounders to deal with and if those confounders are not time-invariant, then you may need to do that in order to accommodate them.