Background
Healthy user bias arises in observational studies when people who take up or continue an intervention, preventive or therapeutic, possess different underlying characteristics from those who do not. These characteristics form a multifactorial profile of the “healthy user,” comprising behavioural traits (e.g., exercising, non-smoking, attendance for health checks), psychological traits (e.g., optimism, proactive health orientation), demographic factors (e.g., higher education, socioeconomic status) and better functional or cognitive status; conversely, frailty and cognitive impairment reduce the likelihood of uptake or persistence.
Two related processes account for healthy user bias. The healthy initiator effect arises when healthier, more health-oriented individuals are more likely to start treatment. The healthy adherer effect occurs when those who remain adherent tend to be healthier, more functional, and less cognitively impaired over time than those who discontinue adherence. Both mechanisms produce systematic differences between exposed and unexposed groups that are unrelated to treatment effects.
As factors associated with “healthy users” are often poorly captured or unmeasured in observational data, substantial residual confounding frequently remains. This can lead to distorted estimates of treatment efficacy and safety in observational analyses. Healthy user bias has been documented across numerous therapeutic areas, including hormone therapy, statins, vitamin supplementation, and influenza vaccination. The reverse pattern can also occur when comparisons include individuals with historically unhealthy behaviours that elevate their baseline risk, such as comparing people who formerly smoked cigarettes and currently vape with people who have never smoked cigarettes, leading to potentially an incorrect attribution of risk to the exposure under study rather than to prior behaviours.
Healthy user bias is related to but distinct from other forms of selection bias. Unlike volunteer bias, which concerns who elects to participate in a study, healthy user bias arises from differences in individuals’ health behaviours in routine clinical practice, such as treatment uptake, lifestyle choices, and preventive care use. Healthy user bias primarily affects internal validity by distorting estimated treatment effects when users and non-users differ in ways that are not fully measured.
Example
Hormone therapy
A classic example of healthy user bias arises from the administration of HRT in women and the risk of coronary heart disease (CHD). Original observational data surmised that the use of HRT was associated with decreased risk among women who were on a course of HRT, and was later confirmed as the effect was ‘unlikely’ to be explained by confounding factors. However, subsequent RCTs failed to find this protective effect of estrogen and indeed found a significant risk of adverse effects. Large-scale RCT research also found an increased risk of CHD. This original effect may have been explained by the observational research failing to control for or adjust results by socio-economic status, a known moderator of CHD risk.
Vaccination
Observational studies have reported that seniors who receive the influenza vaccine appear to have substantially lower risks of death and hospitalisation during flu season compared to unvaccinated seniors. To explore whether this apparent benefit reflected the vaccine or underlying differences between groups, researchers analysed outcomes before, during, and after the influenza season in a large cohort of 72,527 seniors, contributing over 338,000 person-years of observation. They found that vaccinated individuals accounted for more person-time among those with chronic conditions (except dementia), indicating that vaccinated seniors were not randomly selected; they were generally healthier and more engaged with healthcare. They also found that vaccinated seniors showed significant reductions in mortality and pneumonia/influenza hospitalisation before the vaccine could have any biological effect. These reductions were consistent across different definitions of the pre-season period, reinforcing that the observed benefit was driven by differences rather than the vaccine itself.
Impact
In a cohort of more than 20,000 new statin users, Brookhart et al. found that adherent patients were substantially more likely to receive a range of preventive health services, including prostate-specific antigen testing and influenza and pneumococcal vaccination. There is no biological mechanism by which statin adherence should increase receipt of these services; instead, these associations potentially reflect healthier underlying behaviours, greater functional and cognitive capacity, and more consistent healthcare engagement among adherent individuals.
These patterns indicate that adherent patients had more favourable behavioural, functional, and cognitive profiles than non-adherers. Consequently, observational studies that fail to measure these characteristics may yield spurious or exaggerated protective associations, thereby compromising internal validity.
Preventive steps
Several approaches can help reduce healthy user bias in observational studies. Selecting comparison groups with similar patterns of healthcare use and preventive behaviours—such as through new-user and active-comparator designs—can minimise baseline differences between users and non-users. Analyses should adjust for measured confounders using appropriate methods and incorporate markers of frailty, functional status, cognitive impairment, or prior use of preventive services, as these help capture underlying health differences.
Evaluating associations during periods when the intervention is unlikely to have an effect (before the influenza season when studying influenza vaccines) can help identify residual bias. Sensitivity analyses and negative-control outcomes or exposures can further reveal whether unmeasured health-seeking behaviours are distorting results. Drawing a causal diagram (DAG) at the design stage can also help clarify where unmeasured health behaviours or conditioning on adherence may introduce bias and guide the choice of comparison groups and analytic strategies. Ultimately, well-conducted randomised controlled trials remain the most robust way to eliminate healthy user bias, and should be used where feasible.
