Reverse-causation
Description
Reverse-causation is the causal-inference failure mode in which the causal arrow runs opposite to the assumed direction: the analyst infers that A causes B when in fact B causes A. The observed association is entirely real, and — unlike confounding — no third variable is involved. The error is purely directional, which is what makes it so hard to detect: the same data are equally consistent with A → B and B → A, so direction cannot be recovered from the association alone. Four roles compose the shape. The assumed cause A is treated as the lever but is actually the effect. The assumed effect B is treated as the outcome but is actually the cause. The true arrow B → A is the inverted reality, observationally equivalent to the assumed one. The temporal or instrumental resolver is the only thing that can fix the direction — evidence that B precedes A in time, or an instrument / natural experiment that perturbs A without perturbing B. Without such a resolver, the two directions are indistinguishable from the data. The diagnostic question — could the outcome be producing the supposed cause, rather than the other way around? — separates reverse-causation from the adjacent failure modes. It is distinct from confounding (which needs a third variable and is fixed by causal adjustment), from a feedback-loop (which is genuinely two-way, where reverse-causation claims a one-way arrow points the wrong way), and from mere non-causal correlation (where there is no arrow to reverse at all). Asserting reverse-causation is a positive claim requiring its own evidence, not a generic skepticism. The shape recurs wherever the time-order of cause and effect is ambiguous. In epidemiology, an observed link between low cholesterol and cancer was long read as “low cholesterol raises cancer risk” when the true arrow ran the other way — preclinical, undiagnosed cancer lowers cholesterol (the protopathic pattern, where an early symptom of the disease is mistaken for a cause). In economics, more police in high-crime cities was read as “police cause crime” when high crime causes cities to hire more police. The recurring corrective is the same: establish which variable moved first, or find exogenous variation in the supposed cause.Triggers
User-initiated: User asserts that A causes B from observational data where B could plausibly drive A, or asks “which came first?” Vocabulary cues: “reverse causation,” “the arrow runs the other way,” “cause and effect reversed,” “does X cause Y or Y cause X,” “protopathic.” Agent-initiated: Agent notices a causal claim drawn from cross-sectional or observational data where the outcome could be producing the supposed cause. Candidate inference: “could this be reverse causation? What establishes the temporal order, or what exogenous variation moves the supposed cause independently of the outcome?” Situation-shape signals: Cross-sectional associations interpreted causally. “Treatment” variables that could be responses to the outcome. Health/behavior links where the sick or affected change their behavior. Any policy claim where the policy is plausibly a response to the problem it’s blamed for.Exclusions
- Third-variable spuriousness — when a common cause drives both variables with no direct A–B link, the failure is confounding, not reverse-causation. Reverse-causation needs no third variable; the A–B link is real, only inverted.
- Bidirectional causation / mutual influence — when A and B each genuinely cause the other, the shape is a feedback-loop, two-way by construction. Reverse-causation claims a one-way arrow points the wrong way.
- Correlation with no causation in either direction — when the association is coincidental or non-causal, there is no arrow to reverse. Reverse-causation asserts a real causal link whose direction was misread.
- Genuine forward causation — when A really does cause B, the concept does not fire. Asserting reverse-causation requires positive evidence (often temporal) that the effect precedes the supposed cause.
Structure
Relationships
- confounding — sibling causal-inference failure, distinguished by mechanism: confounding needs a third variable (A ← C → B) and is fixed by adjustment; reverse-causation needs none (B → A misread as A → B) and is fixed by temporal/instrumental evidence. Confounding’s entry already lists reverse-causation as a frequently-confused neighbor.
- feedback-loop — the bidirectional foil: feedback is genuinely two-way, reverse-causation is one-way-the-wrong-way. The diagnostic is “the other direction” versus “both directions.”
- selection-bias — fellow member of the “real association, wrong causal inference” family; the two often co-occur in observational medicine (e.g. the “sick quitter” pattern).
Examples
Rose, G. & Shipley, M. J., "Plasma lipids and mortality: a source of error" (The Lancet, 1980, vol. 315, no. 8167, pp. 523–526) · medicine-and-health
Rose, G. & Shipley, M. J., "Plasma lipids and mortality: a source of error" (The Lancet, 1980, vol. 315, no. 8167, pp. 523–526) · medicine-and-health
Observational studies in the mid-twentieth century repeatedly found an association between low plasma cholesterol and elevated cancer mortality, and the association was read in the obvious direction — that low cholesterol somehow raised cancer risk. Rose and Shipley demonstrated that the arrow ran the other way: preclinical, not-yet-diagnosed cancer depresses plasma cholesterol in the months and years before diagnosis. The low cholesterol was an early effect of the latent disease, not a cause of it. Once the analysis excluded deaths in the early follow-up period — the window in which undiagnosed cancer was already lowering cholesterol — the spurious association attenuated.Inference: A clean reverse-causation case with no third variable required. The cholesterol–cancer association is real; only its imputed direction was wrong (cancer → low cholesterol, not low cholesterol → cancer). This is the protopathic pattern — an early manifestation of the disease mistaken for a cause — and the corrective is temporal: establish what preceded what, here by excluding the early-follow-up window where the effect-masquerading-as-cause was operating.
Levitt, S. D., "Using Electoral Cycles in Police Hiring to Estimate the Effect of Police on Crime" (The American Economic Review, 1997, vol. 87, no. 3, pp. 270–290) · economics
Levitt, S. D., "Using Electoral Cycles in Police Hiring to Estimate the Effect of Police on Crime" (The American Economic Review, 1997, vol. 87, no. 3, pp. 270–290) · economics
A naive reading of cross-sectional data finds that cities with more police per capita have more crime, which could be misread as “police cause crime.” The arrow runs the other way: high-crime cities respond to their crime by hiring more police. Estimating the true causal effect of police on crime therefore requires variation in police staffing that is not itself a response to crime. Levitt used electoral cycles — mayors and governors tend to expand police forces in election years for reasons unrelated to the current crime rate — as an instrument that moves police hiring without being driven by crime, and found that police do in fact reduce crime, the opposite of the naive directional reading.Inference: The textbook reverse-causation / simultaneity problem, with the canonical corrective. The observed police–crime association is real; the assumed direction (police → crime) is the inversion of one true arrow (crime → police hiring). Because the two variables are simultaneously determined, no amount of “controlling for” recovers the effect — only an instrument that perturbs the supposed cause independently of the outcome resolves the direction.
Baumeister, R. F., Campbell, J. D., Krueger, J. I., & Vohs, K. D., "Does High Self-Esteem Cause Better Performance, Interpersonal Success, Happiness, or Healthier Lifestyles?" (Psychological Science in the Public Interest, 2003, vol. 4, no. 1, pp. 1–44) · psychology
Baumeister, R. F., Campbell, J. D., Krueger, J. I., & Vohs, K. D., "Does High Self-Esteem Cause Better Performance, Interpersonal Success, Happiness, or Healthier Lifestyles?" (Psychological Science in the Public Interest, 2003, vol. 4, no. 1, pp. 1–44) · psychology
The self-esteem movement built decades of educational programming on the premise that high self-esteem causes good outcomes — better grades, career success, healthier behavior — so raising children’s self-esteem should raise their achievement. Reviewing the evidence, Baumeister and colleagues argued the assumed arrow runs largely backwards: the modest self-esteem–performance correlation is better explained by good performance raising self-esteem than by self-esteem raising performance. High self-esteem, on their reading, is substantially an effect of doing well — they describe it as closer to an epiphenomenon of success than a driver of it — rather than the cause the movement assumed.Inference: A clean reverse-causation case, and deliberately not a feedback-loop one. The review does not conclude that self-esteem and achievement each strongly cause the other (which would make it a feedback-loop); it concludes the assumed-cause path (self-esteem → achievement) is weak or absent while the reverse path (achievement → self-esteem) is the supported one. The association is real and needs no third variable — so it is not confounding either; only the imputed direction was wrong. The diagnostic that exposes it is the same as in any reverse-causation case: ask whether the supposed cause could be the outcome’s product, then look for evidence of temporal order or exogenous variation. The expensive practical consequence — interventions that boost self-esteem directly produce little achievement gain — is exactly what a reversed arrow predicts.
Shaper, A. G., Wannamethee, G., & Walker, M., "Alcohol and mortality in British men: explaining the U-shaped curve" (The Lancet, 1988, vol. 332, no. 8623, pp. 1267–1273) · medicine-and-health
Shaper, A. G., Wannamethee, G., & Walker, M., "Alcohol and mortality in British men: explaining the U-shaped curve" (The Lancet, 1988, vol. 332, no. 8623, pp. 1267–1273) · medicine-and-health
The “U-shaped curve” linking alcohol to mortality — where non-drinkers appear to have higher mortality than moderate drinkers — was widely read as evidence that moderate drinking is protective. Shaper and colleagues showed that the apparent-non-drinker group was contaminated by “sick quitters”: men who had stopped drinking because they were already ill. Their elevated mortality was caused by the pre-existing illness, which had also caused them to abstain; the abstention was an effect of ill health, not a cause of mortality. Reclassifying lifelong non-drinkers separately from ex-drinkers flattened much of the supposed protective effect.Inference: A reverse-causation pattern entangled with selection. The arrow within the abstainer group runs illness → quitting (not abstaining → death), and the abstainer category is also a selected one. The case shows why reverse-causation and selection-bias are catalogued as neighbors — they co-occur in observational medicine, and disentangling them requires separating “never exposed” from “stopped because already affected.”