A New RCM Approach to Event History Analysis

• In Event: 038 - Section on Methodology Paper Session. Advancement in Observing and Modeling Social Processes

Abstract

This paper first describes some major problems in the use of the marginal structural model (MSM) in survival analysis of a nonrepeatable event. The most serious one is a contradiction in the set of assumptions of sequential ignorability in the presence of endogenous sample selection for the hazard rate model, and another problem is the use of numerous selection equations. because, in the hazard rate model, selection bias occurs not only by selection into the treatment state but also by endogenous sample selection by event occurrences. While retaining the framework of Rubin’s causal model, as in the MSM, this paper introduces a new method of assessing the treatment effect in survival analysis by assuming a semiparametric conditional incidence rate model. Unlike the hazard rate model, the conditional incidence rate model not only eliminates the problem of selection bias in the treatment variable at the time of event occurrences, but also solves the problem of endogenous sample selection. This new method yields the estimation of the average treatment effect for the treated for each time of entry into treatment simply as the difference in the log rate between the treatment group and the control group weighted by inverse-probability-of-treatment weights. The overall average treatment effect then becomes the weighted average of the treatment effects of different times of entry into treatment. An illustrative application analyzes the effect of leaving home on the occurrence of premarital sexual initiation.

Author

• Kazuo Yamaguchi, University of Chicago