Some Latent-Class Models for Controlling for Selection Bias in Covariate Effects in Event History Analysis: A Project Description

点击翻译按钮获取中文摘要

This project aims to develop new groups of event-history models which control for selection bias in and out of the states of a covariate of the hazard rate of event occurrence, and assess the net effect of the covariate on event occurrence. Selection bias in the effects of covariates -- such as whether men's entry into marriage causes a decrease in the rate of unemployment or rather uncontrolled characteristics of men who become and remain married at a given age predict a lower rate of unemployment, or whether marijuana use in marriage causes an increase in the rate of divorce or uncontrolled characteristics of people who become marijuana users and retain this behavior in marriage predict a higher rate of divorce -- is a central issue in the analysis of nonexperimental data in order to establish causality. Although others have recently introduced some event-history methods/models to handle this problem, this project will develop distinct groups of models to handle aspects of this issue; in particular, mover-stayer models for bivariate hazards and their extensions, and loglinear latent-class models and their extensions. As a related second aim, the project will clarify the characteristics of the new models regarding the reliability and stability of their parameter estimates and the dependence of the reliability and stability on alternative assumptions made to specify the models.The project has five components: (1) development of software for the new mover-stayer models; (2) an elaboration and extension of basic loglinear latent-class models for controlling for selection bias; (3) applications of those models in social science research; (4) an investigation of the stability and reliability of parameter estimates as a function of specific assumptions made in the models; and (5) some comparisons of the results from the models/methods to be developed in the project with those from related existing alternative models/methods.