How to scale your priors by number of predictors in logistic regression

I wrote up a poor man’s version of this type of prior for a beta and binomial combination on this forum.

The key part is the prior on the centered and scaled predictors on the logit scale.

\gamma \sim normal(0, \sigma_{\gamma}\sqrt{N_{\gamma}})

It’s nothing fancy but by doing that the total sd from the predictors (sd(X\gamma)) gets a prior of \sigma_{\gamma} which you can set according to what is appropriate (or maybe even give it its own prior). I am not sure how to directly related it to ROC AUC.

I thought of the \sigma_y as an indicator of how sure you expect the model to be if the predictors strongly point to one of two outcomes for an observation. Maybe that observation is 3 \sigma_{\gamma} away from the mean on the logit scale. If \sigma_{\gamma} = 1, this implies a probability of inv\_logit(3) = 0.95. If \sigma_y = 2, this implies a probability of inv\_logit(6) = 0.997. Or you could think about how well you think the model will discriminate between an observation that is \sigma_{\gamma} away and an observation that is -\sigma_{\gamma} away.

2 Likes