Structural Nested Failure Time Models

Abstract

Abstract Structural nested failure time models are causal models for the effect of a time‐dependent treatment or exposure on a survival time outcome in the presence of time‐dependent confounders. A time‐dependent confounder is a repeatedly measured covariate, which acts as a confounder for future exposure measurements but is on the causal path between earlier exposure measurements and ultimate response (i.e. here it acts as intermediate variable). Standard epidemiologic methodology fails if time‐dependent confounders are present. Under the essential condition of no unmeasured confounders (the mathematical definition of which is a central issue) these models allow for unbiased causal inference, generalizing Robins's g ‐computation algorithm. The models use as building blocks time‐dependent accelerated failure time models. Important applications are intricate endogenous (feedback) selection effects in occupational epidemiology (healthy worker effect) and clinical epidemiology (AIDS treatment).