Semiparametric modeling of repeated measurements under outcome-dependent follow-up

Abstract

In regression analysis of repeated measurements that are taken at subject-specific times, the availability of the outcome data may be related to the past outcome and to other measured variables that are not in the intended regression model. In this paper we propose a natural extension of the semiparametric regression procedure of Lin and Ying (J. Am. Stat. Assoc. 2001; 96:103-126) by building a class of 'inverse-intensity-rate-ratio' weighted estimators that accommodate such outcome-dependent follow-up. The estimators have a closed form, are radicaln-consistent, asymptotically normal, and do not require estimation of any infinite-dimensional parameters. We give several simulations to demonstrate the estimator's performance and show a sensitivity study under follow-up with various degrees of dependence on outcome-related variables. We illustrate our approach using data from a randomized health services research study with noncompliance to scheduled visits.