Collaborative Research: Hybrid Population-Average and Individual-Specific Models for Clustered Longitudinal Data

University Of Washington, Seattle WA

Investigators

Abstract

This collaborative project with Scott (0088061) will develop a class of longitudinal data models appropriate for modeling clustering and heterogeneous patterns in such data. This work extends the modeling paradigm of Banfield and Raftery (1993), in which clusters are identified via an explicit statistical model. The extension will incorporate a parsimonious class of models developed by Scott and Hancock (1998) to identify population-level patterns in individual-specific differences. The study will develop algorithms to estimate the models, and will incorporate the approach of Bensmail et al. (1997) to establish a Bayesian model formulation. Inference to assess the validity of the models will be developed, as will information criterion, and Bayes factor approaches. A further outcome of this work will be an extension of the lme software in Splus to incorporate these new methods (Pinheiro and Bates 1995). An important part of the project will be the development of a case study of long-term trends in wage inequality. This will include comparisons of two important economic periods and will contrast the findings under the new models to results from more traditional mixed effects models. In addition, the implications of the findings for wage inequality and labor market segmentation will be explored and published in an applied journal to provide a bridge for subject matter researchers to this methodology. In much of social, behavioral and biostatistical research, the goal is to understand the structure of the heterogeneity in a population, and in so doing yield insight into social phenomena. In longitudinal studies subjects can become increasingly differentiated over time, and the identification of natural groupings, or clusters, in the subjects has important consequences. They may yield evidence suggesting the presence of several distinct phenomena--that is, different social processes could exist for different subjects. The accurate identification of these clusters depends on the specification of the structure within a subject's responses. Methods that describe this aspect of an individual's profile have been sparse and often hard to interpret substantively. This research will extend the statistical models for clustered longitudinal data developed in Banfield and Raftery (1993) to include a new, parsimonious, and interpretable representation of individual behavior. This extension represents a hybrid population-average and individual-specific approach to modeling the heterogeneity in individual profiles.