Prevalent User Bias

Prevalent user bias is a combined selection and confounding distortion in observational epidemiology, pharmacoepidemiology, and comparative effectiveness research that occurs when an observational study compares ongoing (prevalent) users of a pharmacological treatment against non-users or alternative treatment users, rather than restricting enrollment to incident new initiators at the moment treatment begins. Formalized by Wayne A. Ray (2003) in the New-User Design framework and expanded within the Target Trial Emulation methodology by Miguel A. Hernán and James M. Robins (2008, 2016), prevalent user bias arises from two distinct causal mechanisms: (1) depletion of susceptibles (a form of left-truncation selection bias analogous to incidence-prevalence bias in which patients who suffered early adverse events or intolerance already discontinued the drug before cohort entry), and (2) adjustment for causal intermediates (inappropriate baseline regression control for clinical covariates that have already been altered by prior exposure to the medication).

In a randomized controlled trial (RCT), three foundational events occur at the same chronological time point (t = 0, termed time zero): (1) eligibility criteria and baseline covariates are measured, (2) treatment strategy is assigned and initiated, and (3) follow-up for clinical outcomes begins. When an observational cohort study instead defines t = 0 as an arbitrary administrative date (such as a survey interview date or insurance database enrollment date) and classifies participants who have already been taking a medication for months or years as "exposed prevalent users," temporal alignment is broken, introducing two systematic errors.

1. Depletion of Susceptibles (Survivorship Selection on Early Hazard)

For most pharmacological agents, the hazard rate h(t) of treatment-related clinical events—such as venous thromboembolism, gastrointestinal hemorrhage, drug-induced hepatotoxicity, anaphylaxis, or pro-arrhythmic sudden death—varies over time since drug initiation. Typically, a biological subpopulation of genetically or physiologically susceptible patients experiences an acute surge in adverse events during the first weeks to months of therapy, after which those patients either die, suffer the endpoint, or discontinue the medication due to side effects.

Consequently, a cross-sectional pool of prevalent users who are still actively taking the drug 2 to 5 years after initiation represents a selectively filtered survivor cohort:

• They have already passed the early high-hazard vulnerability window without experiencing a fatal or treatment-terminating event.
• They tolerate the medication without symptomatic adverse effects and exhibit long-term healthy user bias and persistent adherence.

Comparing these prevalent survivors against untreated individuals (or against newly initiated patients on a comparator drug who are just entering their own early vulnerability period) conditions on a post-treatment survival collider, biasing the estimated hazard ratio downward (HR_prevalent < HR_incident) and concealing early drug toxicity.

2. Overadjustment and Collider Bias from Post-Treatment Covariates

To control for confounding by indication, observational investigators adjust for "baseline" clinical biomarkers and comorbidities measured at cohort entry (t = 0). However, for a prevalent user who began taking a drug two years prior to cohort entry (at t = −2 years), covariates measured at t = 0 (such as serum LDL cholesterol, systolic blood pressure, hemoglobin A_1c, or renal function) are post-treatment variables that have already been modified by two years of pharmacological action:

• Causal Intermediate Overadjustment: Suppose a lipid-lowering or antihypertensive drug reduces myocardial infarction (Y) by lowering LDL cholesterol (M) along the causal pathway A₋₂ → M₀ → Y. If an investigator enrolls prevalent users at t = 0 and adjusts for M₀ in a regression model, the analysis blocks the physiological pathway through which the drug prevents disease, biasing the true beneficial effect toward the null.
• Post-Treatment Collider Stratification: Conversely, if an unmeasured genetic or environmental risk factor U affects both the intermediate biomarker M₀ and the clinical outcome Y (A₋₂ → M₀ ← U → Y), conditioning on the post-treatment biomarker M₀ opens a non-causal collider path between prevalent treatment A and U, distorting the effect estimate.

To see how prevalent user enrollment distorts relative risk estimation, consider a medication that induces a two-phase hazard trajectory relative to non-use:

• Early Susceptible Phase (Year 1 after initiation, t ∈ [0, 1)): The drug triggers pro-thrombotic plaque destabilization in susceptible patients, yielding a true Year-1 hazard ratio of HR₁ = 1.50.
• Late Tolerant Phase (Years 2 to 6 after initiation, t ∈ [1, 6]): Once susceptible patients have been depleted, the hazard ratio among remaining persistent users drops to HR_2–6 = 0.75.

Whereas an incident new-user cohort captures both the Year-1 hazard spike (HR₁ = 1.50) and the later phase, a prevalent user cohort in which 85% of enrolled exposed participants at baseline have already been taking the medication for >1 year misses the Year-1 spike and estimates an apparent protective effect (RR_prevalent ≈ 0.60–0.70).

In 2008, Miguel A. Hernán, Andrea Alonso, Francine Grodstein, and James M. Robins resolved the famous discrepancy between the observational Nurses' Health Study (RR ≈ 0.60 for coronary heart disease with postmenopausal estrogen–progestin therapy) and the 2002 Women's Health Initiative (WHI) randomized trial (HR = 1.29 overall; HR = 1.81 in Year 1). Because Nurses' Health Study questionnaires were mailed only every 2 years, women who initiated hormone therapy between questionnaires and suffered an acute myocardial infarction within their first 6 to 18 months either died before the next questionnaire or stopped hormones after their heart attack; thus, most exposed person-time came from prevalent users who had already survived the early thrombotic risk window. When Hernán and colleagues restructured the observational data to compare incident new initiators against non-initiators, the observational hazard ratio during the first 2 years of therapy was HR = 1.42—matching the randomized WHI trial. Similarly, Wayne A. Ray (2003) showed that prevalent user studies of nonsteroidal anti-inflammatory drugs (NSAIDs) underestimated upper gastrointestinal bleeding risk by >50% because patients developing early peptic ulceration quit NSAID therapy within 30 days.

To eliminate prevalent user bias alongside confounding by indication and immortal time bias, contemporary biostatistics and regulatory bodies mandate the Active-Comparator New-User (ACNU) design (Lund et al., 2015):

1. Washout Period Enforcement: Both the index treatment arm (A = 1) and the comparator arm (A = 0) require a documented drug-free washout window (180 to 365 days with zero dispensings of either medication class) prior to the index prescription date (t = 0).
2. Therapeutic Equivalence of the Active Comparator: The comparator group consists of patients newly initiating a clinically alternative drug prescribed for the same indication and disease stage (for example, SGLT2 inhibitors vs. GLP-1 receptor agonists in type 2 diabetes).
3. Pre-Index Covariate Window and Time-Varying Confounding Adjustment: All propensity score covariates (L₀) are measured during the pre-initiation lookback window [−365, −1 days], and post-initiation treatment switching is adjusted via inverse probability of censoring weights.

How does prevalent user bias relate to incidence-prevalence (Neyman) bias?

Incidence-prevalence bias (Neyman bias) occurs when a study samples prevalent disease cases rather than incident cases, missing rapidly fatal cases. Prevalent user bias is the exposure-side mirror image: the study samples prevalent medication users rather than incident treatment initiators, missing patients who experienced early treatment-induced events or early intolerance.

Why is an active comparator preferred over a non-user comparator even in a new-user design?

Untreated non-users lack a natural clinical "time zero" date on which a prescription was written (often tempting analysts to assign an arbitrary date that causes immortal time bias) and differ in disease severity (confounding by indication) or health consciousness (healthy user bias) from treated patients. An active comparator provides an identical clinical decision date (t = 0) and balances baseline clinical indication.

1. Ray, W. A. (2003). "Evaluating medication effects outside of clinical trials: new-user designs." American Journal of Epidemiology, 158(9), 915–920. https://doi.org/10.1093/aje/kwg231
2. Hernán, M. A., Alonso, A., Logan, R., Grodstein, F., Michels, K. B., Willett, W. C., Manson, J. E., & Robins, J. M. (2008). "Observational studies analyzed like randomized experiments: an application to postmenopausal hormone therapy and coronary heart disease." Epidemiology, 19(6), 766–779. https://doi.org/10.1097/EDE.0b013e3181875e61
3. Lund, J. L., Richardson, D. B., & Stürmer, T. (2015). "The active comparator, new user study design in pharmacoepidemiology: historical foundations and contemporary application." Current Epidemiology Reports, 2(4), 221–228. https://doi.org/10.1007/s40471-015-0053-5
4. Hernán, M. A., Sauer, B. C., Hernández-Díaz, S., Platt, R., & Shrier, I. (2016). "Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses." Journal of Clinical Epidemiology, 79, 70–75. https://doi.org/10.1016/j.jclinepi.2016.04.014