Likelihood ratio test to compare Linear[Mixed]Model

[kleinschmidt]

Likelihood ratio tests can be used to assess the suitability of random effects terms, but it's currently not possible to directly compare a LinearModel and a LinearMixedModel to assess the suitability of no random effects vs. some random effects. This would require adding methods to determine whether a LinearModel is nested within a LinearMixedModel and maybe a specialized test method to make sure you're getting the right likelihood value (e.g. can't use the deviance since the baseline is different for the two).

[palday]

@dmbates Do you have any philosophical objections to this? Practically, the biggest thing is that we need to look at the log likelihoods directly and not the deviances (LinearModel takes the additive constant for the saturated model into account unlike LinearMixedModel).

[dmbates]

Nope. No philosophical objections.