Discussion of Likelihood Inference for Models with Unobservables: Another View

摘要

Lee and Nelder provide considerable food for thought, considerable light and some heat, heat produced by their over-promotion of h-likelihood. I support Lee and Nelder's goal of attempting a unified analysis based on full probability modeling, but note that the Bayesian formalism is best suited to this task.rnUse of the full probability calculus, empowered by modern computing, brings in (most) relevant uncertainties, produces properly shaped and calibrated confidence regions and enables addressing nonstandard goals such as ranking. However, I caution that full probability modeling isn't always available or valid and in many situations compromises are necessary.rnWhatever the approach to analysis, care, evaluation, and sophistication are needed, especially when structuring inferences for latent effects. Polemic and over-promotion distract from the important issues and goals. These should be replaced by aggressive scientific evaluations and energetic discourse.