reverse causation

The MR Dictionary, maintained by the MRC Integrative Epidemiology Unit at the University of Bristol, defines reverse causation as the phenomenon where an association in the direction of a hypothesised causal relationship between an exposure and an outcome is observed, but is due at least in part to a prevalent outcome, potentially undiagnosed or present only as a precursor, influencing the exposure. The arrow runs the other way. What the study records as the thing being measured is partly a product of the thing it is trying to predict, so the association is real and the explanation is inverted.

Sattar and Preiss, writing in Circulation, argue it is more common than researchers imagine and less well understood than confounding, which reports of observational studies usually do acknowledge. Their worked case is blood pressure: systolic pressure below 120 mm Hg goes with higher mortality in people over 80, which appears to contradict the trial evidence, but serial measurements show pressure declining faster in the five years before death than in those who survive, with a marked fall in the final two, and the same whether or not a person is on blood pressure medication. The pattern repeats elsewhere. Illness makes people tired, so they sit down and watch more television, which makes sedentary time look more dangerous than it is. Ill health makes people stop drinking, producing the sick quitter whose abstinence looks harmful. Cholesterol falls in advance of a cancer diagnosis rather than because of low cholesterol causing cancer, a reading confirmed when statin trials showed no rise in cancer rates.

There is no single definitive method against it, only a set of defences that depend on what data exist. The authors list the steps taken by the Global BMI Mortality Collaboration: look only at never-smokers, remove people with known chronic disease as far as is practical, and discard every death in the first five years of follow-up, on the reasoning that the earliest deaths are the ones most likely to reflect illness already present at the start. They add a fourth, comparing age groups, since the youngest are least exposed to the problem. Genetic instruments offer another route, and the MR Dictionary names bidirectional analysis and Steiger filtering as the tests used there. None of it closes the question. Because subclinical illness is by definition unrecorded, the influence of reverse causation cannot be entirely removed, which is why an observational study, however large, cannot settle the direction of an arrow on its own.