Semiparametric Estimation of Relative Causal Effects in Randomized Controlled Trials With Noncompliance.

Abstract

Randomized controlled trials (RCTs) are the gold standard for causal inference and are widely used. However, valid analyses of RCTs are often complicated by non-compliance, which can lead to confounding bias and biased causal effect estimation. The main challenge comes from compliance datasets in both the treatment and control groups both following two-component mixture models. The maximum nonparametric likelihood estimator is inconsistent in a two-component mixture model even if the mixture proportion and one of the components are completely known, but the other component is unknown. In this paper, we instead assume parametric models for the ratios of risks among compliers assigned treatment, never-takers and always-takers, and leave the baseline compliers not assigned treatment unspecified. We develop a novel two-step maximum likelihood estimation procedure by making full use of the observed covariates and latent compliance classes, which theoretically can produce asymptotic root $n$ consistent estimators. In particular, our proposed estimator for the conditional local risk ratio always lies within the range of the parameter. Our numerical results show that the proposed method is generally more reliable than existing alternatives.