3
$\begingroup$

I am working on Sales data. i have binary variable win/loss the opportunities and rest are the activities done by sales force (sales guys) with 40+ variables (different types of activities done for the Opportunity)

I build the logistic model on the available data-set, and i found huge VIF value for different Xi's, then i perform stepwise variable reduction procedure for getting less variable in my model. At the end of this process i got 15 indep variable with dependent variable

Again i build same model on new data-set and again i m getting high VIF around(5610,3374.020669,3270.561737,2.512324,9.922235,...... etc.) for each variable

If u will look at the pairs graph and coefficient result please refer attached picPaire Graph

Result

Please suggest me what should I do further and how to come with my actual model with less error?

I am really stuck for further conclusion.

Improve this question
$\endgroup$
4
  • 2
    $\begingroup$ Often these VIFs, large as they are, would not be a huge concern (such as in a cross-validated predictive setting), but since your model was derived via stepwise regression they call into question the choice of variables. Rather than going further, as requested, would you consider starting over with a better technique of variable selection? $\endgroup$
    – whuber ♦
    Oct 30, 2014 at 14:13
  • 1
    $\begingroup$ appropriate steps depend on your ultimate goals: are you trying to generate the best possible predictions of success or are you trying to understand which factors influence success probabilites and in what way? (please don't say both..) $\endgroup$
    – fabians
    Oct 30, 2014 at 14:50
  • 2
    $\begingroup$ Regularization $\endgroup$
    – Steve S
    Oct 30, 2014 at 15:32
  • $\begingroup$ I am trying to generate best possible predictions of success or in other way can which factors influence success probabilities. My goal of this analysis is to build relationship of all sales force activities with the Opportunity status (win/loss) $\endgroup$
    – user43247
    Oct 31, 2014 at 4:58

1 Answer 1

Reset to default
3
$\begingroup$

A good approach to reduce the dimension of the feature space in regression is partial least-square regression, which finds factors which are both good at explaining the variance in the feature space, but also at predicting the variable of interest.

With a few tweaks, this approach can be used for logistic regression too. For a discussion, see this paper, or this one.

Improve this answer
$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.