The Illusion of the “Self-correcting” Nature of Science

In 2014, Epidemiology published a paper that introduced a method intending to estimate nonlinear causal effects using Mendelian randomization.1 The method (hereafter referred to as residual nonlinear Mendelian randomization) was further developed2 and has been widely used in subsequent work. However, it has also been demonstrated that the method can produce nonsensical findings3 and generate results that are self-refuting4,5: the authors of one of these papers referred to their erstwhile findings as a “logical impossibility.”6 The reasons nonsensical findings are produced are becoming clearer,7 and were hinted at in a commentary published alongside the original Epidemiology paper.8 While two of the papers with clear self-refuting findings have been retracted,9,10 other papers that are clearly equally erroneous have only been corrected. Most concerningly, in the vast majority of cases, papers using residual nonlinear Mendelian randomization have been left to stand with no indication that they are anything but trustworthy. The two retracted papers9,10 presented (fallacious) findings of a striking nonlinear causal effect of vitamin D on cardiometabolic outcomes. According to their results, lower vitamin D was substantially protective against cardiometabolic outcomes for about half the study population. In fact, the “causal effect estimates” obtained using residual nonlinear Mendelian randomization simply reproduced the observational associations. Ironically, the same exposure (vitamin D) in the same study (the UK Biobank) as in the two retracted papers has recently been analyzed using the now discredited approach.11 As expected, this analysis generated nearly identical observational and supposedly causal relationships, strikingly similar to the spurious results advanced in the retracted papers. Some believe that the “self-correcting” nature of science cleanses it from errors that have been introduced, and such a Panglossian view has been put forward in the case of residual nonlinear Mendelian randomization.12 This goes against a wealth of evidence showing this process is, at very best, partial and slow.13 Journals could play an important role in correcting this, for example, by clearly displaying online warnings to readers attempting to access papers presenting methods that have been shown to be, at best, highly problematic. Perhaps if such were done, it would reduce the number of fallacious papers appearing that cite the Epidemiology paper as the source of their methods – as the recent vitamin D paper,11 along with many other recent papers, do.