A colleague has proposed using propensity score matching to do feature selection prior to building a model to predict a binary outcome.

There are 28 variables to select the best predictors from. He proposes iterating 28 times, matching cases on the outcome (success/failure) leaving one variable out each time. At each iteration mean difference tests are conducted on the omitted variable to see if that variable is still meaningful to a difference in the outcome beyond chance. He says that variables that are estimated to have a mean difference are likely valuable to predicting our outcome.

I have never seen this approach to feature selection so am wondering if it is valid.



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.