The ordinary LRT: the objective difference against a chi-squared reference
with Df equal to the number of parameters the larger model adds.
Usage
# S3 method for class 'admFit'
anova(object, ...)Details
Both fits must come from the same estimator and, for the quadrature
estimators, the same node count (n_nodes) or, for the Monte Carlo ones
(admc, adirmc), the same sample size (n_sim). Each scores its own
approximation to the likelihood, so objectives from different ones are not
comparable and the comparison is refused rather than reported.
The stratum resolution is checked per covariate, and only where the two fits overlap. A covariate a model does not read cannot put the two on different scales: if its prediction does not move across a source's nodes, the mixture those nodes collapse to is a sufficient statistic for it, so its objective is the same at either resolution. That is what makes the nested pair of a covariate test comparable — the null model drops the term, so its sources are not cut along it. Two fits that both read a covariate and cut it differently, including one cutting it and the other integrating over it whole, are refused.
Testing a variance AT ZERO puts the null on the boundary of the parameter space, where the exact reference is a chi-bar-squared mixture rather than a chi-squared. The p-value reported there is CONSERVATIVE – too large – so a significant result stays significant, but treat a borderline one with care.