0
$\begingroup$

I am running a logistic regression analysis to model if a patient has a specific disease or not.

I want to remove outliers because i want my model to be as accurate as possible. For the same I learnt that in SAS proc logistic one can obtain Standardized Deviance Residuals and Likelihood Residuals. The observations with values of aforementioned residuals greater than 2 or less than -2 needs to be investigated as outliers.

My question is do I run Proc logistic on full data set or just the training data set(70% of full data set) for outliers removal? Do I have to remove outliers on full data set (meaning both training & test data sets will have outliers removed) or just remove outliers from training data set and let test data set have outliers?

$\endgroup$
  • 2
    $\begingroup$ The implicit assumption--that removing outliers and (presumably) refitting the model will make it more accurate--is doubtful and probably not correct in general. Perhaps your question would be better formulated as "does how one treat observations with outlying deviance residuals affect the accuracy of a logistic regression model and, if so, how should that be done?" $\endgroup$ – whuber Mar 14 at 18:31
  • $\begingroup$ There is broad disagreement among the authors I have read on the best way to test for outliers and whether you should remove them or not [or transform them which is often preferred to removal]. If you are going to conduct analysis with the full data set I think that is the data set you should search for outliers with, although I have not seen this point addressed. When you find one you should try to figure out why it is occurring which will improve your model. $\endgroup$ – user54285 Mar 14 at 19:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.