1
$\begingroup$

I would like to know if in multivariable Cox regression analysis there is a way to yield only models that include a variable of interest (and if no model is statistically significant to just answer that).

I am studying a cell phenotype. This phenotype "high" is associated with graft loss by the log rank test and in an univariate Cox analysis (versus phenotype "low"). I have 14 demographic variables also associated with graft loss as previously published. Now I would like to explore multivariate models. But I am interested only on models that would include my phenotype variable. If no such model is statistically significant it's fine but I want to know. I have 15 variables including my variable phenotype so I would like to be able to test all models with 2 to 15 variables that include the variable phenotype.

Thank you for your help!

$\endgroup$
1
  • $\begingroup$ How many events (presumably, graft losses) are in your data set? Do you have complete data on all 15 potential predictor variables? $\endgroup$
    – EdM
    Aug 20, 2022 at 1:57

1 Answer 1

0
$\begingroup$

This strikes me as p-hacking, but I'm a stats novice so I could be wrong.

If you decide to move forward with this strategy, I would use the My.stepwise.coxph function in R where you can specify that you want a variable to be included in each regression using the argument in.variable.

Here's the link to the documentation:https://www.rdocumentation.org/packages/My.stepwise/versions/0.1.0/topics/My.stepwise.coxph

$\endgroup$
3
  • 2
    $\begingroup$ Your sense that this is close to p-hacking is correct. Automated model selection like stepwise is not a good idea in general. It's a particular problem with Cox models, that have omitted-variable bias if any outcome-associated predictor is left out of the model. $\endgroup$
    – EdM
    Aug 20, 2022 at 2:03
  • $\begingroup$ Hi. OK, thanks for your answers and warnings! $\endgroup$
    – Elodie
    Aug 22, 2022 at 13:03
  • $\begingroup$ @Elodie another approach to your problem, if you don't have many cases, is to include your phenotype as a predictor and do a ridge-regression penalization on the clinical covariates. That way you don't have to do variable selection (with its p-hacking) and you still account for all your outcome-associated variables. Tihs paper is an example. Or see Frank Harrell's ideas on data reduction in Section 4.7 of his course notes for how to reduce multiple clinical variables into a smaller number of predictors. $\endgroup$
    – EdM
    Aug 26, 2022 at 18:29

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.