Causal Multistate Models to Evaluate Treatment Delay.

Abstract

Multistate models allow for the study of scenarios where individuals experience different events over time. While effective for descriptive and predictive purposes, multistate models are not typically used for causal inference. We propose an estimator that combines a multistate model with g-computation to estimate the causal effect of treatment delay strategies. In particular, we estimate the impact of strategies such as awaiting natural recovery for 3 months, on the marginal probability of recovery. We use an illness-death model, where illness and death represent, respectively, treatment and recovery. We formulate the causal assumptions needed for identification and the modeling assumptions needed to estimate the quantities of interest. In a simulation study, we present scenarios where the proposed method can make more efficient use of data compared to an alternative approach using cloning-censoring-reweighting. We then showcase the proposed methodology on real data by estimating the effect of treatment delay on a cohort of 1896 couples with unexplained subfertility who seek intrauterine insemination.