I'm looking for a "hack" in R that would allow me to calculate the log-likelihood of a GLM fit on a separate test set easily regardless of the distribution. For instance for a Gamma GLM, this is how we could extract the training likelihood:
x_train <- rnorm(1000, mean=100)
y_train <- rnorm(1000, mean=100)
mdl <- glm(y_train ~ x_train, family=Gamma(link="log"))
loglik_train <- logLik(mdl)
If I have test data available, what is a quick way to calculate the log-likelihood of observing these new datapoints given the parameterized model estimated from the training set?
I wish to avoid doing the full calculation (making predictions, calculating for each prediction the likelihood and then summing) since I will have different distribution GLMs & link functions et cetera...
Thanks a lot for any suggestions!
