Misspecifying the Shape of a Random Effects Distribution: Why Getting It Wrong May Not Matter

被引:167
作者
McCulloch, Charles E. [1 ]
Neuhaus, John M. [1 ]
机构
[1] Dept Epidemiol & Biostat, San Francisco, CA 94107 USA
关键词
Maximum likelihood; mixed models; parametric modeling; LINEAR MIXED MODELS; RANDOM-EFFECTS MISSPECIFICATION; MIXING DISTRIBUTION; LONGITUDINAL DATA; CLUSTERED DATA; II ERROR; TRIAL; MIXTURES; OUTCOMES; DISEASE;
D O I
10.1214/11-STS361
中图分类号
O21 [概率论与数理统计]; C8 [统计学];
学科分类号
020208 ; 070103 ; 0714 ;
摘要
Statistical models that include random effects are commonly used to analyze longitudinal and correlated data, often with strong and parametric assumptions about the random effects distribution. There is marked disagreement in the literature as to whether such parametric assumptions are important or innocuous. In the context of generalized linear mixed models used to analyze clustered or longitudinal data, we examine the impact of random effects distribution misspecification on a variety of inferences, including prediction, inference about covariate effects, prediction of random effects and estimation of random effects variances. We describe examples, theoretical calculations and simulations to elucidate situations in which the specification is and is not important. A key conclusion is the large degree of robustness of maximum likelihood for a wide variety of commonly encountered situations.
引用
收藏
页码:388 / 402
页数:15
相关论文
共 46 条