科研之友
微信
新浪微博
Facebook
分享链接
摘要
Generalized linear mixed models (GLMMs) comprise a class of widely used statistical tools for data analysis with fixed and random effects when the response variable has a conditional distribution in the exponential family. GLMM analysis also has a close relationship with actuarial credibility theory. While readily available programs such as the GLIMMIX procedure in SAS and the Ime4 package in R are powerful tools for using this class of models, these programs are not able to handle models with thousands of levels of fixed and random effects. By using sparse-matrix and other high performance techniques, procedures such as HYMIXed in SAS can easily fit models with thousands of factor levels, but only for normally distributed response variables. In this paper, we present the %HPGLIMMIX SAS macro that fits GLMMS with large number of sparsely populated design matrices using the doubly-iterative linearization (pseudo-likelihood) method, in which the sparse-matrix-based HPMIXED is used for the linear iterations with the pesudo-variable constructed for the inverse-link function and the chosen model. Although the macro does not have the full functionality of the GLIMMIX procedure, time and memory savings can be large with the new macro. In applications in which design matrices contain many zeros and there are hundreds or thousands of factors levels, models cab be fitted without exhausting computer memory, and 90% or better reduction in running time can be observed. Examples with a Poisson, binomial, and gamma conditional distribution are presented to demonstrate the usage and efficiency of this macro.
- 出版日期2014-6
- 单位Microsoft
