Is it appropriate to adjust a Cox regression model for the baseline year when using data from open cohorts?

Question

I was wondering whether it is correct to include baseline year as additional covariate in a Cox regression model when using data from an open cohort. I will try to give an example to make this clearer:

Data: open cohort of individuals followed up from 1990 to 2016

Baseline: year of diagnosis of hypertension

Exposure: lifestyle intervention in the year following the diagnosis

Outcome: fatal cardiovascular event

In this case each individual would enter the cohort the year of diagnosis of hypertension and that year would be the baseline year. as individuals are censored at different time the model would account for the differences in the follow-up. As that is the baseline year, all covariates (i.e. age, sex, ethnicity, BMI..) registered in that year would be included. As this is meant to be a non-time varying model, covariates would not change over time and no interaction term would be fitted with time. However, the time period is quite extended and there are changes in guidelines in how to treat the condition over time, therefore, I would be keen to include in the model year of diagnosis (baseline year) as additional covariate (categorical variable with first year as ref). To me this would better account for the possible changes in management over time. However, I could not find a good reference that would answer my question about whether this is a correct approach. Any advice?

Answer 1

Including the year of diagnosis in a Cox model is a perfectly good way to test your hypothesis that two patients with the same covariate values might have different survival depending on date of diagnosis. With a non-time-varying model as you propose, the covariate values of interest are those at the time 0 of entry into the study. The date of diagnosis is no less valid in principle than any other covariate value at entry time. I have done this successfully for Cox models of a disease whose therapeutic approaches have improved over time.

Do think, however, about just what hypothesis you will be testing in this way. Simply including date of diagnosis as a separate covariate will control for overall changes over the decades in cardiovascular mortality. It would not, however, control for things like improvements in the quality of a particular intervention that might have happened during the study. For that you would have to examine an interaction of date of entry with the intervention in question. Note that an interaction of a covariate with the date of entry does not pose the problems that might arise with interactions between a baseline covariate and time elapsed from study entry and the resulting immortality bias. The date of entry is just as fixed at time 0 as are the other covariate values.