I am log-transforming my predictor variables to meet the linearity on the logit assumption for logistic regression. In so doing, I am adding a constant to ensure the minimum value is 1 .

My question is, if I then standardize my columns (e.g., min-max scaling), I will again end up with zeros that I will have to add a constant to, to ensure I can do more logistic regresison diagnostics like testing linearity on the logit, etc.

Is there any other way around having to add a constant twice, as described above?


1 Answer 1


Yes. Rather than torture your data until it fits a model, use a model that does not make assumptions which your data violates.

One idea is to use a spline of the IVs. Another is a classification tree (or one of the offshoots of these). Yet another is a neural net.

  • $\begingroup$ Thanks Peter - my limitation, however, is that my study is based on applying a logistic regression model. So unfortunately, data torture is the only way I see to get to the end goal. When violating model assumptions, I gather the main impact is on statistical inference capabilities, but machine learning prediction per se is not necessarily affected? $\endgroup$ Oct 9, 2023 at 0:52
  • 1
    $\begingroup$ @user19226726 You can spline your features and use a logistic regression. // If data torture is the only way to reach your ultimate goal, it is reasonable to wonder if you’re doing something that should be done. Always remember the “could/should” line in Jurassic Park. $\endgroup$
    – Dave
    Oct 9, 2023 at 1:06
  • $\begingroup$ That strikes me as a really silly notion -- to base a study on a particular method. But splines would be within logistic regression. $\endgroup$
    – Peter Flom
    Oct 9, 2023 at 11:00

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.