In a recent article on the efficacy of antihypertensive therapy, Berlowitz et al (1998) introduced an ad hoc method of adjusting for serial confounding assessed via an intensity score, which records cumulative differences over time between therapy actually received and therapy predicted by prior medical history. Outcomes are subsequently regressed on the intensity score and baseline covariates to determine whether intense treatment or exposure predicts a favorable response. We supply sufficient conditions for interpreting results of the Berlowitz approach based on a causal model for the effect of a time-varying exposure. We also consider a modified version in which the intensity score records cumulative scaled differences over time between therapy actually received and therapy predicted by prior medical history. This leads to a simple, two-step implementation of G-estimation if we assume a nonstandard but useful structural nested mean model which implies that subjects less likely to receive treatment are more likely to be helped by it. These modeling assumptions might apply, for instance, to certain health services research contexts in which differential access to care is a primary concern. We also extend the methods to accomodate repeated outcomes and time-varying effects of time-varying exposures.
Joint work with Sander Greenland, Mary Redman, Nancy Kiviat and Paula Diehr
