Do propensity score matching methods need to factor in the index date in a matched cohort context?

Question

I am working on a comparative effectiveness study where we estimated the propensity of treatment between two groups and are exploring matching on the propensity score. The study period is long, encompassing about 7 years of data, and exposed and unexposed patients are selected across the entire period at different points. The propensity score itself is estimated using a high dimensional propensity score estimation method that includes the year of the index date and several other potential confounders. In general, the propensity score balances confounding between the two treatment groups.

In my experience, matching exposed and unexposed patients (using the date of exposure as the index date) in this context is a matched cohort study and would be analyzed as usual like a cohort study. In the past, I would also ensure that exposed and unexposed patients came from a similar point in time in case there were changes in the standard of care or treatment practices over the study period that would confound the treatment effect.

My questions are: can the data be analyzed after matching on the propensity score without needing to further ensure the patients come from a similar point in time? Would this have an impact on the treatment effect estimate or introduce a bias in any way and if so, how should I address it?

Thank you for your time!

Answer 1

If when in calendar time the treatment was administered might change the outcome and the probability of receiving treatment, you absolutely should match on it. You could create time strata and perform matching analyses within each one, you could require that any pair comes from the same time window and perform a paired analysis, or you could just include time as a regular covariate used in creating the propensity score. I think it would make the most sense to measure it as a categorical variable, splitting it using a sensible window length.

Whether this is required depends on the substantive qualities of the data. Are there any causal pathways from date of administration of treatment to the outcome? Maybe those who received the treatment later had less time between treatment and the measurement of the outcome, which could make it less effective than had more time elapsed. If someone else published a study and didn't control for time of administration, would you see that as a weakness of the study, making its results less trustworthy? If if you wouldn't, can you imagine a reasonable but persnickety reader feeling this way? Your goal is to defend your causal effect estimate against any arguments that would weaken the claim you are trying to make, and controlling for time would strengthen your claim (unless you had a reason to think otherwise, which you must articulate in your report).