Re: Biases in Randomized Trials

    loading  Checking for direct PDF access through Ovid


Mansournia et al.1 linked the terminology regarding biases in clinical trials (described by the Cochrane risk-of-bias tool) to that in epidemiologic studies using causal diagrams assuming two-treatment comparisons. Targeting per-protocol effect that would have been observed if all the patients had adhered to the protocol of a randomized trial, they showed that per-protocol and as-treated estimates have selection and confounding biases in epidemiologists’ terminology unless sufficient confounders are adjusted for. They argued that intention-to-treat estimates in randomized trials are unbiased for the intention-to-treat effects within the study population despite nonrandom initiation of post randomization treatments.1
According to recent literature on causal estimands in confirmatory trials (e.g., ICH-E9(R1) concept paper), recognizing the “bias” in intention-to-treat analysis for per-protocol effects, which was ignored by Mansournia et al.,1 is useful for clinical trialists; the bias is viewed as measurement (misclassification) bias in epidemiologists’ terminology. In fact, Greenland2 formalized nonadherence to treatment allocation as a misclassification problem and introduced instrumental-variable methods to epidemiologists for misclassification correction. As g-estimation with an instrument (randomization) is becoming accepted as a method to regain unbiasedness for per-protocol effects from intention-to-treat estimates, recognizing measurement bias in intention-to-treat estimates, selection bias in per-protocol estimates, and confounding bias in as-treated estimates, all of which are adjusted with g-methods,1 would be informative for trialists and epidemiologists.
Another concern about intention-to-treat analysis is generalizability or transportability. Mansournia et al.1 argued that protocol-unspecified treatment O causes the intention-to-treat effect to deviate from that in other settings, rather than cause bias in intention-to-treat estimates. However, another causal-inference school calls this deviation a “generalizability bias.”3 Besides terminology, one may be interested in joint intervention on randomized A and nonrandomized O, such as time-varying treatments’ effects or direct effects, where the problem is not external (i.e., generalizability) but internal validity (i.e., absence of bias).
Again, g-methods can estimate the more clinically relevant effects in other settings that are possibly different from intention-to-treat and per-protocol effects on the study population. Particularly, borrowing a concept from stochastic dynamic regimes,4 we can specify rules that do not depend on observed unintentional treatment patterns but incorporate intentional patterns: for example, “start chemotherapy A = a, and if patent’s performance status is 3 or higher when considering any subsequent treatment, then initiate O = o1 and o2 with probabilities 0.75 and 0.25, respectively; otherwise initiate O = o2.” Their effects are estimated by inverse-probability weighting,4 and would mitigate “generalizability bias” for effects in wider situations, which is actually realized by real-world trialists and epidemiologists.
    loading  Loading Related Articles