Issue with time-varying covariates / stsplit in survival analysis of prospective cohort data

[Kai Schulze]

Hello members of Statalist,

First quickly “thanks” for this forum and all the stuff that is posted here - I have found many solutions over the years through the useful answers provided in this forum. I am currently looking into a new topic and would be grateful about some advice.

Dataset:
I am working with an epidemiological prospective cohort dataset that started in the 90’s in the UK with 25,000 individuals. The objective is to estimate the association between energy-intake from processed foods and incidence of CVD (failure event).

For each individual, I have two observations of the covariates. One, when the subject entered the study and the individual ‘baseline’ measurements were taken ("HealthCheck 1", between the years 93-98), and at a second time ("Health Check 2"between the years 98-2000). The outcome is ascertained through hospital admission / health service registry records.

Analysis:
I want to perform Cox-regression with CVD as the event and want to take into account the changes / two time-points of the covariates.

Issue:
I cannot figure out how to enter the time-varying covariates. The tvc - option of stcox doesn't really apply I think it is difficult with non-continous covariates.
I looked at stsplit to split up the data, but figured that I cannot specify the date variable that would indicate the date of the second health check for each individual.

Does anyone have an idea how I can set up this analysis? Any hint / reference would be appreciated.

Thanks a lot in advance!
Kai.

[Steve Samuels]

Welcome to Statalist, Kai!

I don't think that two measurements over eights years are enough to be time-varying covariates. To treat them as such requires the assumption there was little change between the 1st and 2nd exams and after the 2nd. You cannot even estimate the intermediate values as the predictions on a line connecting the 1st and 2nd measurements: that would violate the fundamental principle that a cause must precede effect. Here the predicted values would be functions of something known only at the second Check.

The most defensible approach, in my opinion, is to treat the first measurements as baseline covariates for the entire period of followup. You could follow by a secondary analysis which starts on the date of the second measurement and treats it as a new baseline measurement. However you still risk serious inferential problems if you have dropouts related to health status. See this thread and the reference therein titled "The hazards of time-varying covariates".