2
$\begingroup$

For my thesis I made a prediction model for food security using logistic regression (n=210). One of the predictors included in my final model is 'food expenditures'. The independent variable is originally positively skewed. It had a few outliers, which I removed. The logodds of the independent variable followed a linear relationship with the dependent variable food security. Would I have a reason to do a lognormal transformation on the predictor food expenditures? I read that monetary variables are often transformed, should this also be the case in my research?
I'm looking forward to your answers!

$\endgroup$
0
$\begingroup$

Yes, I have seen that sometimes it helps.

Also sometimes one needs to bin continuous predictor variable if there is a case for quasi - separation, meaning observing very few 0 or 1 values after certain threshold of this continuous variable.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Binning continuous predictors is seldom a good idea! It looses information. I cannot see how it could possibly help with quasi-separation, can you explain? I would model the predictor with a regression spline. $\endgroup$ – kjetil b halvorsen Jun 23 '18 at 20:01
  • $\begingroup$ When you combine observations into binned classes then one can get rid of cases where there is only few 0 or 1 class observations after certain range of x variable. WoE type of transform is often used in credit scoring to have similar effect. There classes with similar WoE score are combined and each WoE vector has increasing trend in its classes with respect to target variable. $\endgroup$ – Analyst Jun 23 '18 at 21:49
  • $\begingroup$ Of course it is true that one loses information if continuous variable is transformed into cat one, but sometimes one must do this thing. $\endgroup$ – Analyst Jun 23 '18 at 21:51

Your Answer

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

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