Strong Cross-Pillar

Healthy-User Bias: Why the People Who Take the Supplement Were Already Going to Live Longer

Name: Healthy-User Bias: Why the People Who Take the Supplement Were Already Going to Live Longer
Creator: Realised
Published: 2026-06-01

Summary

The people who take the vitamin, stick to the regimen, get the screening, or show up to exercise are systematically different from those who don't — wealthier, more health-conscious, less sick to begin with — so "users of X live longer" observational findings often measure who chooses X, not what X does; the cleanest proof is that adherence to a sugar pill predicted survival almost as strongly as adherence to the real drug.

Why Strong

Strong Evidence. The bias is established epidemiologic methodology with a single, devastating empirical demonstration — adherence to a placebo predicting survival in a randomised trial, a result that cannot be explained by the treatment — plus a long, concordant graveyard of observational-to-RCT reversals (hormone therapy, vitamin E, beta-carotene) that the bias substantially explains, and confirmatory occupational-epidemiology data (the healthy-worker effect). The "not a reason for blanket cynicism" side is equally well-grounded: lower lifetime LDL survives both Mendelian randomization and statin RCTs. The only genuine uncertainty is apportioning how much of any given reversal is healthy-user bias versus other confounding or timing — and that uncertainty is the entry's point, not a weakness in its tier.

Practical takeaway

How to spot it, and what to do:
• Run the diagnostic question on every "users of X live longer" claim: could the kind of person who does X — rather than X itself — produce this association? If X is a deliberate, effortful, self-selected health behaviour (a supplement, a screening, a regimen), the answer is "very plausibly yes," and the claim is default-suspect.
• Ask whether a randomised trial exists, and which way it went. The whole graveyard (vitamin E, beta-carotene, hormone-therapy-for-CVD) is observational benefits that an RCT erased or reversed. No matching trial means the claim hasn't faced its hardest test.
• Use the magnitude heuristic. A roughly 1.5x-2x association in a wealth- and health-correlated behaviour is exactly the size healthy-user bias produces, so treat it as a hypothesis. A very large effect, or one confirmed by Mendelian randomization or an RCT, is hard to explain by selection alone.
• Watch for the biologically-implausible tell. When a behaviour is "associated with" a benefit it has no mechanism to cause (statins and fewer hip fractures), that incoherence is the bias announcing itself.
• Apply it to yourself, too. Realised recommends adherence-heavy behaviours (sleep regularity, exercise, supplements). The same effect operates within one person over time: the weeks you keep the protocol tend to be the weeks your life is together — sleep, money, mood, schedule all steadier — so don't over-credit the protocol for gains the surrounding stability also produced. Test it honestly (n_of_1_self_experimentation).

Evidence detail

Why This Entry Exists

The single most common shape of a weak health claim is "studies show people who do [healthy thing] have lower rates of [disease]." It sounds like evidence. Usually it is a portrait of the kind of person who does the healthy thing. Someone who remembers a daily supplement, adheres to a prescription, and turns up for screening is, on average, richer, more educated, more risk-averse, and less sick before the intervention ever touches them. Those traits drive longevity on their own. The behaviour gets the credit; the person earned it.

This is the healthy-user effect (and its cousins, the healthy-adherer and healthy-worker effects). It is the prime suspect behind the graveyard of observational "benefits" that died in randomised trials — vitamin E, beta-carotene, hormone therapy for heart protection, antioxidants. But it is not a licence to dismiss all observational data: some associations survive the strongest de-confounding tests precisely because the real effect dwarfs the bias. The honest move is a single diagnostic question — could the kind of person who does X, rather than X itself, produce this? — applied symmetrically.

This is a spoke under the evidence-literacy hub (rct_vs_observational_evidence). It owns the selection/adherence confounder. It defers the general RCT-vs-observational machinery to that hub, and funding-driven distortion to cui_bono_industry_funding_bias.

What bad advice this protects against, in all directions:
• "People who take [supplement] have lower disease rates, so take it." → mistaking a self-selected, healthier population for a treatment effect.
• "Adherent patients do better, so adherence is the cure." → adherence is mostly a marker of being the kind of person who does well; the placebo arm proves it.
• "This worker population is healthy, so the exposure is safe." → the sick were filtered out of the workforce; the comparison is rigged.
• Over-correction: "Healthy-user bias means all observational data is worthless." → false. Smoking, inactivity, and lifetime LDL survive RCT and genetic-randomisation scrutiny because the effect is far too large to be selection.

Evidence

1. The canonical receipt — adherence to a placebo predicts survival (Coronary Drug Project). In this randomised secondary-prevention trial in men with a prior heart attack, good adherers to the active drug clofibrate (took at least 80% of pills) had 5-year mortality of 15.0% versus 24.6% for poor adherers (P = 0.00011). That alone looks like the drug working. The kill-shot is the placebo arm: good adherers to the inert sugar pill had 15.1% mortality versus 28.3% for poor placebo adherers. Since the pill does nothing, the entire roughly 13-point survival gap is the healthy-adherer effect — adherence is a marker of being the kind of person who lives longer, not a treatment. (Coronary Drug Project Research Group, N Engl J Med 1980; the placebo-arm result is randomised-trial data not subject to confounding-by-treatment — essentially unarguable. Strong Evidence.)

2. The graveyard — hormone therapy for heart protection, the largest reversal. Observational cohorts, notably the Nurses' Health Study, found postmenopausal hormone replacement associated with roughly a one-third lower risk of coronary heart disease. The Women's Health Initiative randomised trial found the opposite: combined estrogen-progestin increased coronary events (about 29% higher CHD incidence) plus elevated stroke and breast cancer, and was stopped early. Hormone users in the cohorts were wealthier, leaner, more educated, and more health-engaged — a textbook healthy-user population. Both-ways caveat: a timing hypothesis partially rehabilitates the hormone story — women starting within 10 years of menopause or under 60 fared far better in WHI reanalysis than the 70-79 group — so the reversal is part healthy-user selection and part an age/timing mismatch between the younger cohort initiators and the older trial participants. Honest reading: multiple artefacts, not pure bias. (Rossouw et al., JAMA 2002; timing reanalysis Rossouw et al., JAMA 2007; reversal catalogued in Shrank et al. 2011. Strong Evidence.)

3. The graveyard — antioxidants didn't just fail, beta-carotene caused harm. Observational data suggested beta-carotene, vitamin E, and vitamin C lowered cancer and cardiovascular risk. Trials reversed the sign. The ATBC trial (Finnish male smokers, 20 mg beta-carotene/day) showed roughly 18% more lung cancers and about 8% more deaths in the supplemented arm. CARET (beta-carotene plus vitamin A) showed about 28% more lung cancers and 17% more deaths, and was stopped early. A meta-analysis across more than 109,000 subjects confirmed roughly a 24% increase in lung-cancer risk among smokers on high-dose beta-carotene. This is the strongest case against naive observational supplement claims: the bias did not merely inflate a null into a benefit, it inverted the truth — the supplement was net harmful, while the people who happened to eat beta-carotene-rich diets (and were healthier for unrelated reasons) appeared protected. (ATBC Study Group, N Engl J Med 1994; Omenn et al. CARET, N Engl J Med 1996. Strong Evidence.)

4. The named mechanism — three nested biases, defined. The standard physician-facing primer (Shrank, Patrick & Brookhart) formalises the family: (1) the healthy-user effect — people who adopt one preventive therapy also seek other preventive services and healthy behaviours (exercise, diet, seatbelts); (2) the healthy-adherer effect — among users, the adherent are systematically more health-engaged (the Coronary Drug Project demonstrandum); plus confounding by functional and cognitive status and by physicians selectively prescribing preventives to healthier patients. They flag a near-miss tell: statins associated with roughly a 23% reduction in hip fracture in meta-analysis — biologically implausible, and a healthy-user red flag, because statin users are simply healthier overall. This entry's diagnostic question comes from here. (Shrank, Patrick & Brookhart, J Gen Intern Med 2011 — canonical methodological reference, not primary effect data. Strong Evidence.)

5. The occupational twin — the healthy-worker effect. Employed cohorts show all-cause mortality roughly 25% below the general population, because the severely ill and disabled are filtered out of employment (healthy-hire effect) and sick workers leave (healthy-survivor effect). The consequence: a study comparing exposed workers to the general public can make a genuinely hazardous exposure look harmless or even protective. First noticed by William Ogle in 1885. This is the same selection logic in a different domain — the comparison group is contaminated by who gets selected in, not by the exposure — which shows the bias generalises beyond pills to any "people who do or are X" grouping. (McMichael, J Occup Med 1976; reviewed Li & Sung, Occup Med 1999. Strong Evidence.)

6. The counterweight — associations that survive because the effect dwarfs the bias. Lower lifetime LDL cholesterol causing less coronary disease is confirmed by both Mendelian randomization — genotype randomised at conception, immune to healthy-user confounding (Ference et al. estimate naturally lower LDL gives roughly three times the per-unit risk reduction of late-life statin therapy) — and by statin RCTs. Smoking and physical inactivity likewise show mortality effects far larger than any plausible healthy-user inflation, confirmed across designs. The lesson is not "observational data is worthless" but the magnitude heuristic: a doubling of risk in a behaviour entangled with wealth and health-consciousness is suspect; a 20-fold effect (smoking and lung cancer), or one confirmed by a design that does not share the confounder, is real. (Ference et al., J Am Coll Cardiol 2012; Mendelian-randomization framing Davies et al., BMJ 2018. Strong Evidence.)

Mechanism

The statistical core is confounding by selection. An observational comparison of "did X" versus "didn't X" is only fair if the two groups are alike in everything except X. They almost never are when X is a deliberate health behaviour, because the decision to do X is itself caused by traits — affluence, conscientiousness, baseline health, access to care — that independently cause better outcomes. Those traits are the confounders. They sit upstream of both "takes the supplement" and "lives longer," manufacturing an association with no causal arrow from the pill to the survival.

The healthy-adherer layer is subtler and is what the Coronary Drug Project exposes. Even within a group of people prescribed the same thing, the ones who actually take it are different: adherence to a daily regimen is a behavioural marker of organisation, stability, and engagement with one's own health. Those same qualities predict survival regardless of what the pill contains. That is why adherence to a placebo "worked" — the pill was inert, so the only thing adherence could be measuring was the person.

The healthy-worker version is the same machine pointed at a population rather than a behaviour: employment selects in the well and selects out the sick, so the reference group is pre-filtered to be healthier than the general public, and any exposure measured against it looks safer than it is.

Randomisation defeats all three because it assigns X by chance rather than by who chooses it — breaking the link between the trait and the exposure, and balancing even the confounders no one thought to measure. Mendelian randomization achieves the same de-confounding through nature: a gene variant you inherited at conception was not chosen by your health-consciousness, so it isolates the effect of the thing the gene influences. When an association holds up under one of these, the healthy-user explanation is off the table.

Risks And Contraindications

The failure mode here is over-correction into nihilism — using "healthy-user bias!" as a universal solvent to wave away any observational finding you find inconvenient, regardless of its effect size or whether an orthogonal design confirms it. That is the mirror error of naively trusting observational data, and just as unscientific. The counterweight is load-bearing: lower lifetime LDL, smoking harm, and exercise benefit survive precisely because the effect dwarfs the bias and independent designs (RCTs, Mendelian randomization) agree. Selective skepticism — invoking the bias only against claims you dislike while accepting weaker evidence for claims you favour — is the same weaponised-doubt trap the hub warns about.

A second trap is conflating distinct artefacts. The hormone-therapy reversal is partly healthy-user selection and partly the timing/age mismatch between younger cohort initiators and older trial participants. Calling it "pure healthy-user bias" overstates the case; name the confounds honestly.

Scope discipline: this entry owns selection and adherence confounding ("who does X is different"). It does not re-argue RCT-versus-observational design in general (that is the hub), funding-driven distortion (cui_bono_industry_funding_bias), or regression to the mean — a separate artefact entirely.

Controversy

Nature of the dispute: not whether the bias is real — the placebo-adherence result settles that — but how much of any given observational-to-trial reversal it explains, and how far that justifies discounting observational evidence generally.

Position A — default-suspect. Healthy-user and healthy-adherer bias is real, large, and the prime suspect whenever an observational "people who do healthy thing X live longer" finding evaporates or reverses in a trial. People who take supplements, adhere to regimens, get screened, and exercise are systematically wealthier, more health-conscious, less sick, and better at follow-through, and these traits — not the intervention — drive much of the survival gap. Default suspicion of single-arm observational benefit claims is warranted.

Position B — not a licence for blanket cynicism. The bias is not grounds to discard all observational data. Some associations survive precisely because the true effect dwarfs the bias and is confirmed by orthogonal designs that don't share the confounding — Mendelian randomization and RCTs both confirm that lower lifetime LDL causes less coronary disease, and that smoking and inactivity cause large mortality differences far bigger than any plausible healthy-user inflation. Reflexively discarding all observational data is as unscientific as naively trusting it.

Funding: the receipts here come from independent methodology and trial literature (the Coronary Drug Project, ATBC, CARET, WHI, the Shrank primer) — public/academic, none selling a product. The bias itself, notably, operates with zero industry funding: it is about who selects into the exposure, not who paid for the study, which is exactly why it earns a spoke separate from funding bias.

Realised Position: Realised treats "users of X live longer" observational claims as default-suspect for healthy-user and healthy-adherer confounding, and asks whether a matching RCT exists before crediting the behaviour — but it does not weaponise the bias into nihilism. Where the effect dwarfs the bias or an orthogonal design (RCT, Mendelian randomization) confirms it, the association stands. The operative move is the diagnostic question — could the kind of person who does X, rather than X itself, produce this? — not reflexive trust and not reflexive dismissal. Because Realised recommends adherence-heavy behaviours, it flags that the same bias operates within a single user over time, so improvements are not over-credited to the protocol.

Cross-Pillar Connections

This is a cross-pillar evidence skill — it governs how to read claims in every pillar where a behaviour is chosen: a sleep supplement, a diet pattern, a training habit, a meditation practice. Wherever an entry rests on "people who do this have better outcomes," this spoke is the check on whether the behaviour or the person earned the result. It anchors to the evidence-literacy hub (rct_vs_observational_evidence) and pairs with n_of_1_self_experimentation for the within-person version of the same trap.

What would change our mind

Falsifiability: explicit upgrade/downgrade criteria from source

The core phenomenon is effectively unfalsifiable by better data — the placebo-adherence result is randomised-trial evidence immune to the usual confounding, so "the bias is real and large" stands. What would update the balance:
• If adherence-adjustment methods (per-protocol estimators, negative-control-outcome calibration, active-comparator new-user designs) were shown to reliably recover the trial answer from observational data, we would soften "observational benefit claims are default-suspect" toward "suspect unless properly adjusted."
• If a re-analysis showed a flagship reversal — say hormone therapy — was mostly timing and age confounding rather than healthy-user selection, that specific receipt should be down-weighted and re-labelled.
• On the counterweight side, if Mendelian-randomization assumptions (no pleiotropy, no population stratification) were shown to be systematically violated for the LDL case, the "some associations survive" argument would need a different anchor — though it has several (statin RCTs, the sheer size of the smoking effect), so the both-ways structure survives regardless.

The entry is robust; the live uncertainty is apportionment per case, not the existence of the bias.

Industry bias note

Structural incentives the evidence base may reflect

Healthy-user bias is the engine behind a large share of the supplement and wellness industry's strongest-sounding evidence: "studies show people who take [X] have lower rates of [disease]." That cross-sectional/cohort design is precisely the one most contaminated by the bias — the people buying and remembering to take a daily supplement are wealthier, more health-conscious, and less sick at baseline. The commercial incentive is to present the unadjusted observational association and quietly omit that the matching trials found nothing (vitamin E, multivitamins for cardiovascular disease) or harm (beta-carotene in smokers).

This is the mirror image of funding bias (cui_bono_industry_funding_bias): there the distortion is who paid; here it is who selects into the exposure. The bias operates even with zero industry funding, which is why it earns its own spoke. The exploitation runs both ways — the same selection logic can make a hazardous exposure look safe (the healthy-worker effect shielding an industrial exposure from scrutiny), not only make a useless supplement look protective. For Realised's posture: when surfacing any "users of X live longer" claim, default to asking whether an RCT exists and what it showed, and treat unreplicated observational supplement benefits as Emerging or Experimental at best.

Sources (7)

Open in the Library: search, filter, every entry →