Existing methods for fitting generalized additive mixed models to longitudinal repeated measures data rely on Laplace-approximate marginal likelihood for estimation of variance components and smoothing penalty parameters. This is thought to be appropriate due to the Laplace approximation being established as an appropriate tool for smoothing penalty parameter estimation in spline models and the well-known connection between penalized regression and random effects. This paper argues that the Laplace approximation is sometimes not sufficiently accurate for smoothing parameter estimation in generalized additive mixed models leading to estimates that exhibit increasing bias and decreasing confidence interval coverage as more groups are sampled. A novel estimation strategy based on penalizing an adaptive quadrature approximate marginal likelihood is proposed that solves this problem and leads to estimates exhibiting the correct statistical properties.

翻译：暂无翻译