I am working on a project to determine the variables that better predict the binary outcome. I am using conditional random forest and permimp::permimp to assess the importance of variables for my subgroup analysis. Now, I want to make probability predictions from a properly constructed random forest, based on each new patient's values.

For instance, if I initially used 30 variables to fit the random forest and, through permutation, identified 7 variables as important. Now, when we have a new patient, and we want to predict the probability of their response, we may only have the 7 important variables collected, not all 30. Can we still predict the probability of the patient being a responder based on these 7 variables alone, or do we need all 30 variables? Alternatively, should we refit the random forest based only on these 7 variables?

Not all variables might be collected because the setting involves clinical data. Hence, I could train the model based on one study's data, while a clinician is presumably interested in the probability of developing the disease, but he/she might not have the same data collected.

  • $\begingroup$ With the random forest models I've used, importance is measured continuously and predictors are not dichotomized into Important and Not. Is your type of random forest different in this regard? $\endgroup$
    – rolando2
    Jan 25 at 20:30
  • $\begingroup$ Why did you decide to use a random forest for this problem? $\endgroup$
    – dipetkov
    Jan 26 at 6:03
  • $\begingroup$ @rolando2 I am using conditional random forest and variable importance is calculated using permutations with permimp R package $\endgroup$
    – Kate
    Jan 27 at 14:20


Your Answer

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

Browse other questions tagged or ask your own question.