So I am not sure if this question makes sense at all, but I will try to explain.

I want to build a logistic regression scoring model that learns automatically (updates) when sample is updated with new data. I have done variable selection using correlations among variables, statistical tests and univariate analysis combined with apriori knowledge about variables and their potential influence on the response.

But I have one problematic variable, for which I know that it should be, say, strongly negatively correlated with response, but in my sample it turns out to be positively correlated.

I probably wouldn't care about it that much if I just wanted to do a one-time analysis, and just wouldn't include this variable in the model. But is there any reasonable way how to include it so that it is only "used" when sample data also confirms that it is negatively correlated?

  • 1
    $\begingroup$ +1 It's a good question, but maybe it would help to point out that apart from using knowledge of the variables and the response, the process you describe for variable selection tends not to work well and can identify entirely the wrong variables. You can learn a great deal about effective approaches by searching our site: for a start, try stats.stackexchange.com/search?q=model+variable+selection. $\endgroup$ – whuber Jan 3 '18 at 13:56
  • $\begingroup$ Could you please specify exactly which part of the process (tools?) seems unreliable ? $\endgroup$ – Iden Jan 8 '18 at 15:51
  • 1
    $\begingroup$ The use of "correlations among variables" and relying on (presumably formal) "statistical tests" for variable selection in regression models are known to produce misleading results. $\endgroup$ – whuber Jan 8 '18 at 17:24

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.