I just realise that we do not have a binomial_logit_glm function in the language. There is only a Bernoulli version of it or a categorical logit… but how come there is no binomial_ logit glm function (which are super fast)?
Should that be worked on … so that an issue is needed?
Is the motivation fractional counts from some kind of pre-processing?
If so, I’d suggest writing a different function rather than overloading binomial, which is, strictly speaking, only defined with support over integers. A generalized real variate version would presumably take p(y) propto theta^y * (1 - theta)^(N - y). What’s the normalizing term?