I have a retrospective dataset that has records of the medications a patient was administered to prevent recurrence of heart attack. The goal is to determine if a particular drug or combination of drugs is better at preventing (or prolonging) the recurrence of heart attack.

My plan is to create an indicator variable for each of the drugs taken and include all possible two way interaction between the drugs in a cox regression model treating the problem as a survival analysis problem. Would this be an appropriate approach to answering the question? I'm thinking of using the interaction terms to model the effect of drug pairs. Should I create indicator variables for all pair of drug combination in the data instead and why?

  • $\begingroup$ Do you have data from every possible combination of drugs? If not, using an interaction term may not make much sense for reasons alluded to in this answer to a previous question. $\endgroup$
    – Ian_Fin
    Aug 11, 2016 at 11:18
  • $\begingroup$ @Ian_Fin No, I was planning to include interaction terms for only the pairs that exist. Anyway, do you think it would be better to create an indicator variable for each unique pair rather than using interaction terms? $\endgroup$
    – godspeed
    Aug 11, 2016 at 14:56
  • $\begingroup$ if you have a sufficient number of observations for each unique combination then you could do that. I can't tell you what is sufficient, but if you have combinations where there's only one or two observations then that's certainly not and you'd have to remove those from your analysis. $\endgroup$
    – Ian_Fin
    Aug 11, 2016 at 15:02
  • $\begingroup$ How many drugs combinations do you have? You mention pairwise combinations, but some patients could be taking more than two drugs. I also think you will run into issues with interpretation with many pairwise interactions and their main effects. As you suggest in the last sentence, do a simple analysis of indicators for each drug and see if that answers your question. Simple analysis may show that in the presence of one drug, another loses its effect, etc. You could then consider more complex models, but should ideally have some prospective plan for analysis to avoid testing too many times. $\endgroup$
    – Todd D
    Aug 26, 2016 at 14:51


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.