I am trying to understand how probability works with a threshold for logistic regression.

  1. I understand the basics of how to calculate probability.

    log odds = intercept+value1*coef1
    odds = exp(log odds)
    prob = odds / (1+odds)
  2. I understand that a threshold is used to find the optimal mix of correct predictions (precision, f1, etc.).

However, how do we interpret a probability in light of a threshold? For example, if a threshold is 0.195, and a user has a probability of 0.0975 are they:

  • 50% likely to respond (1) since they are 50% towards the threshold?
  • Or are they still 0.0975% likely to respond (1), irrespective of how we consider the fact that anyone who is more then 0.195% likely is going to respond (1)?

They are predicted to have a $0.0975$ probability of responding $1$. The threshold you have chosen has no effect on the probability, only what you do with the predicted probability later. I should note that these are just the model's estimates, they need not be the true probabilities of responding $1$—models can certainly be wrong!

As a final note, I need to point out that positing a threshold and calling all observations above it, $1$, is not generally a good thing to do. There is more information in the predicted probability than in the attempted classification.

  • $\begingroup$ Thank you for the clarification. I am aware that there is more information to be captured, and am also interpreting different co-efficient. I just wanted to ensure my understanding of the resulting probability was correct as I am creating a nomogram to display the effect of different variables. $\endgroup$
    – sdhaus
    Jun 22 '17 at 0:56

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.