This is more of a clarification question than anything else, but I am familiar that when dealing with regression algorithms it is useful to eliminate features that are highly correlated. I have been reading that this stays true for both highly +/- features. In a data science interview question video I saw a few months back asked a question on inversely correlated features and I believe the answer was that if two features are inversely proportional then they can both be fed into a logit model. Is this true?
$\begingroup$
$\endgroup$
5
-
$\begingroup$ Do you mean that $r_{XY} = -0.9?$ $\endgroup$– DaveCommented Sep 15, 2021 at 18:26
-
$\begingroup$ Yes exactly. @Dave $\endgroup$– thomaspaneCommented Sep 15, 2021 at 18:27
-
2$\begingroup$ Did you or the interviewer (or answer key) say that both features can be fed into the model? // Why eliminate correlated features, even when the correlation is positive? $\endgroup$– DaveCommented Sep 15, 2021 at 18:27
-
$\begingroup$ The question was asked by the interviewer and the interviewee answered with both features can be left to which the interviewer said was correct. Feeding features that are highly correlated into my logit model in SAS causes the convergence criteria to fail which states that the model coefficients cannot be trusted. @Dave $\endgroup$– thomaspaneCommented Sep 15, 2021 at 18:31
-
1$\begingroup$ That's a problem with the algorithm, not the statistical procedure. $\endgroup$– whuber ♦Commented Sep 15, 2021 at 19:46
Add a comment
|