|
Abstract:
|
Generalized linear mixed models (GLMMs) provide a convenient framework for modeling of longitudinal and correlated data. However, a common concern is the appropriateness of modeling assumptions, such as normality of the random effects and linearity. To relax these assumptions, I propose an adaptive generalized additive mixed model with a nonparametric prior on the random effects distribution. The proposed prior is based on a novel local generalization of the Dirichlet process, deemed the kernel local partition process (KLPP). This prior automatically allows for random effects and basis function selection, while locally borrowing information across subjects, leading to a much sparser representation of the data than global partitioning-based methods, such as the Dirichlet process.
|
