Temporality and reverse causation
Why it is easy to miss
It produces exactly the association the hypothesis predicts, in the expected direction, often with a dose-response gradient. Nothing in the data announces it, because the data cannot see which came first unless the design recorded it.
The common shapes
- Straightforward reversal: physical activity and depression, where lower mood reduces activity as well as the reverse.
- Preclinical disease altering the exposure: weight loss preceding a cancer diagnosis, making low body mass index look causal.
- Protopathic bias: a drug prescribed for the earliest symptoms of the disease it is later blamed for.
- Health-conscious behaviour changing after early symptoms, which affects several exposures at once.
- Confounding by indication, which is not reverse causation but is frequently confused with it and needs a different remedy.
What reverse causation looks like
Design defences
Measure the exposure before the outcome, in a genuinely prospective design with a clear time zero. Exclude an initial period of follow-up and show the results across several exclusion windows rather than one. Use an exposure measured long before the outcome. Where a genetic instrument exists, Mendelian randomization is attractive here specifically because genotype is fixed at conception and cannot have been altered by later disease, though its own assumptions then need arguing.
Say what would distinguish the two
The useful discipline is to state, in advance, what pattern would be expected under reverse causation and what pattern under the hypothesis, and then report whether the data can tell them apart. Often they cannot, and saying so is more informative than a limitations sentence.
What to watch out for
The most common form is not a simple reversal but a subtler one: undiagnosed early disease changes the exposure before it announces itself, so low cholesterol looks like a risk factor for cancer when it is an early consequence of it. Excluding the first few years of follow-up is the usual defence and it is a judgement, not a fix, because the right lag depends on the disease's preclinical duration. A cross-sectional study cannot establish temporality at all, whatever the analysis.
Clinical example
A cohort finds lower serum cholesterol associated with higher cancer incidence. Excluding the first five years of follow-up removes most of the association, which is what would be expected if undiagnosed cancer lowers cholesterol rather than the reverse.
- Mendelian randomization: genetic anchors for causal inference in epidemiological studies (opens in a new tab) · Hum Mol Genet · doi 10.1093/hmg/ddu328
- Tutorial on directed acyclic graphs (opens in a new tab) · J Clin Epidemiol · doi 10.1016/j.jclinepi.2021.08.001