Addressing complications of intention-to-treat analysis in the combined presence of all-or-none treatment-noncompliance and subsequent missing outcomes

Abstract

We study the combined impact that all-or-none compliance and subsequent missing outcomes can have on the estimation of the intention-to-treat effect of assignment in randomised studies. In this setting, a standard analysis, which drops subjects with missing outcomes and ignores compliance information, can be biased for the intention-to-treat effect. To address all-or-none compliance that is followed by missing outcomes, we construct a new estimation procedure for the intention-to-treat effect that maintains good randomisation-based properties under more plausible, nonignorable noncompliance and nonignorable missing-outcome conditions: the 'compound exclusion restriction' on the effect of assignment and the 'latent ignorability' of the missing data mechanism. We present both theoretical results and a simulation study. Moreover, we show how the two key concepts of compound exclusion and latent ignorability are relevant in more complicated settings, such as right censoring of a time-to-event outcome.

Principal Stratification: A Broader Vision

“…Any change in context will lead to a different CACE, therefore limiting the clinical usefulness of the PS approach. As a concrete example, when controls do not have access to active treatment in the study, the current PS method (and any instrumental variable estimator) states that there are no ATs [8,21]. However, once the active treatment is approved, ATs will very likely exist, and the CACE will be different due to the change in the membership of the BC stratum.…”

A Randomized Evaluation of Ohio's Personalized Assessment Reporting System (PARS)

“…Other methods exist for correcting this kind of sample selection issue including principal stratification (Frangakis & Rubin, 1999), instrumental variables adjustment (Angrist, Imbens, & Rubin, 1996), propensity score matching or adjustment (Rosenbaum & Rubin, 1983), Heckman's method for selection correction (Heckman, 1976), and Bayesian methods (Imbens & Rubin, 1997). Unfortunately, each of these methods impose assumptions upon the data, they can produce conflicting results, and their use may fail to produce unbiased estimates of treatment effects (Holland, 1989;LaLonde & Maynard, 1987;Little, 1985;Murnane, Newstead, & Olsen, 1985;Wainer, 1986).…”

Statistical power in randomized intervention studies with noncompliance.

“…Yau and Little (2001) demonstrated the simultaneous modeling of noncompliance and nonresponse using the ML-EM and Bayesian estimation methods, assuming no correlation between noncompliance and nonresponse (i.e., missing at random; Little & Rubin, 1987). Frangakis and Rubin (1999) demonstrated the simultaneous modeling of noncompliance and nonresponse using Bayesian estimation methods, allowing for possible correlation between noncompliance and nonresponse (i.e., missing not at random; Little & Rubin, 1987). Statistical power will be affected not only depending on whether this missingness is taken into account, but also depending on what is assumed for the relationship between noncompliance and nonresponse.…”