Causal inference with treatment effect modifier


Hi everyone

I’m posting to ask for references regarding causal inference with treatment effect modifiers in the context of a comparison of treated group vs untreated group.

The context is an analysis to estimate the effect of lipid lowering treatment on LDL-cholesterol (as % reduction from baseline) using routine health care data from diagnosis of familial hypercholesterolaemia (FH) to 2 years follow-up. We will focus on the patients who were not treated before the FH diagnosis. We will compare the patients who were treated after the FH diagnosis with those who were not treated.

Clearly, patients who were not treated are likely to be different from those who were treated in characteristics that affect the outcome (% reduction in LDL cholesterol from the FH diagnosis baseline) and the decision to treat.

The issue is that some of the confounders are only relevant for the treated group. Specifically, characteristics that affect adherence will affect the decision to treat and the LDL-cholesterol reduction, but will not affect the LDL-cholesterol reduction in the untreated group. These characteristics are effect modifiers (i.e. interaction effect) but not prognostic (i.e. main effect).

Hence my questions are:

  1. Should the characteristics that predict the treatment decision and the LDL-cholesterol reduction but only in the treated group be included?
  2. Is propensity score matching the appropriate technique for this situation?

I understand that this is quite a complex area, so I’m not expecting a fully thought out answer. I would be grateful if anyone could direct me towards good papers about this topic.

Thank you very much in advance!

Best wishes



Hi Rita, just to be a trouble maker I’d like to pose the question of whether causal inference can be the true goal of analyzing casually collected data. I tend to think of data from medical practice as providing information about the prevailing outcomes of prevailing treatment strategies, not about what would have happened had different treatments been used in the same patients (the causal question). My personal opinion is that it’s best to rephrase this in terms of estimating associations with best available confounder adjustment.

Observational data gives us major challenges in estimating treatment effects when adherence is perfect. When there is informative (non-random) non-adherence, things are much more challenging.

Lastly, no matter what the goals are, I suggest not taking for granted that percent change in LDL cholesterol is a good metric. First, you can’t do stats on percent change because + and - percent change don’t properly cancel, e.g., a 100% increase is canceled by a 50% decrease. Send, to my knowledge it has never been validated that % change in LDL cholesterol is independent of baseline. It is generally best to jointly analyze the baseline and the raw follow-up LDL values, to answer more relevant questions such as “if two patients getting different treatments started at the same LDL do they end up with different LDLs?”.