FRIDAY, February 7, 2003
Time: 1:30 - 2:30 PM
Constant Hall Room 1052

Liam O'Brien
Department of Biostatistics, Harvard School of Public Health, Boston

We present a multivariate logistic regression model for the joint analysis of longitudinal multiple source binary data. Longitudinal multiple source binary data arise when repeated binary measurements are obtained from two or more sources, with each source providing a measure of the same underlying variable. Longitudinal studies with multiple sources generally produce a relatively large number of observations per subject, creating two challenges for their analysis. First, the proliferation of association parameters requires that parsimonious models for the within-subject association must be adopted. We consider regression models for the marginal pairwise odds ratios for obtaining a parsimonious within-subject association structure. Second, an additional complication arises with estimation, since maximum likelihood estimation may not be possible without making unrealistically strong assumptions about third- and higher-order moments. To circumvent this, we propose the use of a generalized estimating equations approach. Finally, we present an analysis of multiple informant data obtained longitudinally from a psychiatric interventional trial that motivated the development of this model.