0

I want to test for several, predefined interactions with my main variable of interest in a multivariable adjusted Cox regression model. Is there a recommended way to proceed?

Specifically, should I test these interactions one after an another or include all of them at once? And when I find one significant interaction (<.05) should I test the other interactions with the significant one in the model?

I am sorry if I am asking a too general question, after reading many posts I was not able to find a clear answer. Thank you very much in advance!

newb
  • 1

1 Answers1

1

Provided that you have enough data to evaluate all of the interaction terms without overfitting (typically about 15 events per coefficient you are estimating), include them all in the model and evaluate them in the context of the entire model.

Do not fall into the trap of conflating "statistical significance (p < 0.05)" with practical importance. Even a predictor or interaction that isn't "statistically significant" can improve the quality of the model overall and provide better estimates for the coefficients of the other predictors.

If you don't have enough data to include all the interactions without overfitting, don't fall into the trap of stepwise or other automated selection of predictors/interactions based on p-values. See Frank Harrell's course notes and book for better ways to proceed; pay particular attention to Chapter 4 of each.

EdM
  • 92,183
  • 10
  • 92
  • 267