On the causal structure of information bias and confounding bias in randomized trials

Randomized trials are undoubtedly different from observational studies, but authors sometimes propose differences between these designs that do not exist. In this article we examine two claims about randomized trials: first, a recent claim that the causal structure of exposure measurement (information) bias in a randomized trial differs from the causal structure of that bias in an observational study. Second, a long-standing claim that confounding bias cannot operate in a randomized trial - if randomization was perfectly implemented. Using causal diagrams (causal directed acyclic graphs), we show that both claims are false in the context of an intention-to-treat analysis. We also describe a previously unrecognized mechanism of information bias, and suggest that the term 'information bias' should replace the terms 'measurement bias' and 'observation bias'.