Estimating
Multilevel Models for Categorical Data Via Generalized Least Squares
Minerva Montero & Valia Guerra
Abstract
Montero et al. (2002) proposed a strategy to formulate
multilevel models related to a contingency table sample. This methodology is
based on the application of the general linear model to hierarchical
categorical data. In this paper we applied the method to a multilevel logistic
regression model using simulated data. We find that the estimates of the random
parameters are inadmissible in some circumstances; large bias and negative
estimates of the variance are expected for unbalanced data sets. In order to
correct the estimates we propose to use a numerical technique based on the
Truncated Singular Value Decomposition (TSVD) in the solution of the problem of
generalized least squares associated to the estimation of the random
parameters. Finally a simulation study is presented to shows the effectiveness
of this technique for reducing the bias of the estimates.
Key words: Multilevel models, Generalized least squares, Truncated
singular value.
PDF
(English)