What is advice of when to use one-vs-all logit or multinomial logit regressions? Most importantly, which one has a higher prediction power? Can one test hypothesis and estimate confidence intervals in one-vs-all approach?

I came from economics and I have just recently started to dive into Machine Learning. I noticed that they use different approaches to estimate discrete choice models. However, I could not find any paper or website where somebody would compare those two approaches. I found this, but it does not answer the question of prediction power: Multinomial logistic regression vs one-vs-rest binary logistic regression

  • 1
    $\begingroup$ We know they are two different models, why should we compare the prediction power? $\endgroup$
    – SmallChess
    Commented Apr 22, 2017 at 12:53
  • 1
    $\begingroup$ Because we still want to predict the same thing: what option will be chosen. Therefore, you could compare, for example, how many right predictions were made on a new set of data using those two approaches. $\endgroup$
    – dart_kaide
    Commented Apr 22, 2017 at 13:06
  • $\begingroup$ To me, they are not predicting the same thing. It's just Apple and Orange. $\endgroup$
    – SmallChess
    Commented Apr 23, 2017 at 11:19


Your Answer

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