An anova method that returns a LRT of two models with associated p-values would be helpful for users and should be easy to implement.
Not sure what would be best with bayesian bootstrapping though, as likelihood is a distribution and not a single value. Mean/median? Another option would be to use the same weights for both models, so we would get a distribution of likelihhood ratios that we could also report. This seems more robust and relatively straightforward to code, but cannot think of a user-friendly implementation that does not require including even more arguments in the model fitting function.
Alternatively/additionally, we could implement a drop1 method that performs a LRT for each predictor, handling paired weights internally.
An
anovamethod that returns a LRT of two models with associated p-values would be helpful for users and should be easy to implement.Not sure what would be best with bayesian bootstrapping though, as likelihood is a distribution and not a single value. Mean/median? Another option would be to use the same weights for both models, so we would get a distribution of likelihhood ratios that we could also report. This seems more robust and relatively straightforward to code, but cannot think of a user-friendly implementation that does not require including even more arguments in the model fitting function.
Alternatively/additionally, we could implement a
drop1method that performs a LRT for each predictor, handling paired weights internally.