0
$\begingroup$

To perform logistic regression I wish to winsorize outliers in independent/ explanatory variables by flooring and capping independent variables. Can you suggest how I should choose cut-off for winsorization, i.e 1 & 99th or something else. Also is there any impact on decision if the distribution of variable is left/ right skewed and not normal.

Please assume we don't have common knowledge of the variable to choose cut off intuitively.

$\endgroup$
9
  • 2
    $\begingroup$ This seems to be an unusual (and potentially biased) approach to logistic regression. Would you have any references recommending and justifying its use? $\endgroup$ – whuber Apr 8 '19 at 22:25
  • 2
    $\begingroup$ (1) Automatically throwing out outliers based on a percentile and (2) winsorizing explanatory variables in regressions are both unusual procedures and require careful justification. $\endgroup$ – whuber Apr 8 '19 at 22:57
  • 1
    $\begingroup$ What problem are you trying to solve with the capping? That should help guide us towards a principled solution. $\endgroup$ – Matthew Drury Apr 8 '19 at 23:37
  • 1
    $\begingroup$ @whuber: The help page for glmrob claims it can do this, see the argument weights.on.x (with default value none). It can even take a function as arguments, so can be used as a basis for own implementation. But I didn' test it ... here is a promising paper. $\endgroup$ – kjetil b halvorsen Apr 10 '19 at 21:26
  • 1
    $\begingroup$ @kjetilbhalvorsen Thank you! I hadn't noticed that option, and it's precisely what is needed. $\endgroup$ – whuber Apr 11 '19 at 13:06

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.