I am interested in running a CPH model and looking at treatment and a biomarker score. Treatment fails the proportional hazard assumption, looking at both Schoenfeld residuals and log-log plots. I would stratify my analysis on treatment, however, I am mainly interested in the interaction between treatment and the biomarker score. I am not really sure how to or if I can interpret the interaction term if I stratify on treatment. If there is not a good interpretation for an interaction term involving a stratified variable any suggestions on a more appropriate analysis would also be helpful.
The Cox model has an interesting feature: you can interact a non-modeled factor (stratification factor) and a modeled covariate. This will assess whether the effect of the covariate (biomarker) is modified by the strata.
Hi Dr. Harrell,
Thank you for the quick response! I believe I am quite lucky to have blindly fallen on this feature as it will allow me to proceed with my analysis w/o a large strategy change.
From an implementation standpoint, should I stratify on treatment and only include treatment in the interaction term (i.e. not include treatment main effect)?
Correct. Having treatment as a stratification factor is more general than having it as a main effect so you don’t also include the main effect term.