Primary Submission Category: Dynamic Treatment Regimes
Structural Nested Models in Target Trial Emulation
Authors: Fuyu Guo, Oliver Dukes, Mats Julius Stensrud, James Robins,
Presenting Author: Fuyu Guo*
Target trial emulation is a popular method for estimating effects of treatment regimes from observational data. In the emulation, new trials, indexed by time, are initiated at fixed intervals. A subject participates in every trial for which eligibility criteria are met. Current methods treat each time-specific trial separately. For instance, for a trial comparing the regimes “always” versus “never” treat from initiation at t onwards, it is common to fit a hazard or risk ratio (RR) model that includes a treatment indicator and its potential confounders. Subjects are censored if they later change treatment, with inverse probability weighting to adjust for the censoring. If most subjects change treatment, the estimates will be inefficient. In this paper we propose more efficient estimators by introducing regime-specific structural nested target trial emulation models (SNTTEM). Given a regime, a SNTTEM imposes parametric models for all time-specific blip functions of the eligible subjects and leaves those for the ineligible unrestricted. A time-specific blip function quantifies on a mean scale the effect of initiating the regime at a time t versus one period later, as a function of past history. The intersection of all the earlier time-specific RR models constitutes a single SNTTEM with regime “always take the treatment that one took last time”. We show that SNTTEM can be fitted using g-estimation, a method that censors less and is more efficient than current methods.