Healthy Adherer Bias: When Persistence Looks Like Pharmacology
Anas H. Alzahrani, MD PhD MPH
Department of Preventive Medicine and Public Health
Faculty of Medicine, King Abdulaziz University
Researchers often compare patients who stay on treatment with patients who do not, then act surprised when the persistent group lives longer, gets admitted less, and appears more responsible in every way. The surprise is charming. The inference is usually not.
Healthy adherer bias happens when patients who adhere to treatment also carry a package of prognostic advantages that the study only partly measures: better functional status, more stable routines, higher health literacy, stronger social support, better access, lower frailty, or simply a greater willingness to do difficult health-related things on time.
The Core Mistake
Adherence is not just a treatment-delivery variable. It is also a behavioral marker. Once a paper compares persistent users with discontinuers, it may be comparing different patients rather than different pharmacology.
Decision rule:
If adherence, persistence, refill regularity, or visit attendance defines the contrast, assume bias is in play until the paper shows how prognosis, time-varying illness, and health-seeking behavior were handled on purpose.
Or less politely: patients who reliably swallow pills are often doing several other beneficial things that never make it into the regression table.
Why This Bias Is So Persuasive
The exposure feels concrete
“Adherent” sounds operational and measurable, which makes the resulting estimate look cleaner than it really is.
The effect often points in a believable direction
Better adherence producing better outcomes is easy to believe, so implausibly broad benefit can pass as common sense.
The missing variables are annoyingly human
Motivation, cognition, support, competing priorities, and frailty rarely sit in neat columns, but they shape both adherence and prognosis.
Comparison Matrix: Three Easy Biases to Confuse
| Problem | What drives the distortion | Classic signal | What reviewers should ask |
|---|---|---|---|
| Confounding by indication | Sicker or higher-risk patients are more likely to receive treatment. | Treatment appears harmful because it is given to people already closer to the event. | How were baseline severity, disease activity, and treatment choice modeled at time zero? |
| Healthy user bias | People who initiate a preventive treatment are already healthier or more prevention-oriented. | New users of a preventive therapy look broadly protected, even for outcomes weakly related to the drug. | Was the initiator group compared with an active comparator living in a similar care pathway? |
| Healthy adherer bias | Patients who persist, refill, or attend follow-up differ behaviorally and prognostically from those who do not. | Adherent patients look better on many outcomes, sometimes including outcomes the drug should not affect much. | Did the analysis treat adherence as a time-varying process and address the reasons people stopped, switched, or disappeared? |
A Clinical Example That Looks Better Than It Deserves
Imagine an observational study after hospital discharge for myocardial infarction. Patients who remain highly adherent to statins over 12 months have lower mortality than patients who become nonadherent. That direction is plausible. The problem is that the observed contrast is doing at least three jobs at once.
1. Real treatment effect
Statins genuinely reduce cardiovascular risk. Some part of the adherence contrast could reflect biology.
2. Selection on evolving prognosis
Patients stop treatment because of frailty, adverse effects, polypharmacy burden, cost, cognitive decline, new cancer, worsening symptoms, or plain life chaos. Many of those same forces raise mortality.
3. Adherence as a behavior proxy
The persistent group may also exercise more, attend cardiac rehab, refill other medications, and keep follow-up appointments. The paper may call that residual confounding. The problem is usually larger than residual.
This is why an impressive adherence hazard ratio is not self-validating. Sometimes it is partly a drug effect. Sometimes it is partly a prognosis filter. Often it is both, tangled together.
The Placebo Clue Researchers Keep Forgetting
One of the cleanest warning signs comes from randomized settings: even adherence to placebo can track better outcomes. That does not mean placebo is secretly pharmacologically ambitious. It means adherence often marks patient characteristics and circumstances that matter for prognosis.
Takeaway:
If staying on an inert pill predicts better outcomes, then adherence itself is carrying information about the patient. Any nonrandomized adherence comparison deserves immediate suspicion.
Reviewer Red-Flag Table
| What you see | Why it should slow you down | Better response |
|---|---|---|
| Adherent patients show lower risk for almost everything | Broad benefit across unrelated outcomes often means healthy behavior is leaking into the estimate. | Check negative control outcomes and ask whether the effect looks implausibly universal. |
| The paper adjusts only for baseline covariates | Stopping and persisting are driven by time-updated health changes, not just baseline characteristics. | Ask how evolving symptoms, toxicity, hospitalizations, and access disruptions were handled over follow-up. |
| Nonadherence is treated like a simple subgroup label | That ignores the fact that people enter the nonadherent group for reasons tied to prognosis. | Demand a protocol-style definition of adherence and a method that respects time-varying selection. |
| Early divergence after treatment start is dramatic | Immediate benefit can be more behavioral sorting than mechanism, especially for long-latency outcomes. | Examine outcome plausibility, grace periods, and whether discontinuation reflects early illness rather than effect. |
| Authors say “we adjusted for medication possession ratio” and move on | Measuring adherence is not the same as neutralizing the bias it carries. | Ask whether adherence is the exposure, a mediator, or a selection process. Those are not interchangeable jobs. |
Design Moves That Actually Help
- Start with a clear estimand. “Effect of assignment” and “effect under sustained adherence” are different questions and deserve different analyses.
- Use active comparators when possible so initiation and persistence happen inside more comparable care pathways.
- Capture time-varying clinical status, toxicity, hospitalizations, access disruptions, and competing treatment changes instead of pretending baseline covariates can do the whole job.
- For sustained-adherence questions, use methods designed for time-varying adherence and informative censoring rather than ordinary regression on adherers versus nonadherers.
- Interrogate negative controls. If adherence predicts outcomes the drug should barely influence, your estimate is carrying more behavior than biology.
What This Means for AI-Assisted Methods Review
Healthy adherer bias is exactly the sort of problem that slides past shallow automation. The variables are present. The model fit looks neat. The conclusion sounds clinically sensible. Yet the study may still be comparing organized lives with disorganized ones and calling the gap a treatment effect.
That is useful territory for Aqrab: manuscripts where the failure is not a missing p-value but a confused causal story. If you want a faster critique of an adherence analysis, comparative-effectiveness paper, or reviewer response draft, start with Aqrab Try. If you want the methodology logic behind those critiques, visit /developers.
Keep reading
Don't stop at one method.
Good methods judgment comes from contrast. Read the neighboring guides, see where the assumptions diverge, and avoid treating every observational problem like it needs the same hammer.
Index Event Bias: When Your Cohort Already Selected the Wrong Comparison
A practical guide to index event bias for clinical researchers. Covers recurrence-risk paradoxes, conditioning on the first event, secondary prevention cohorts, and what reviewers should demand before trusting protective-looking associations inside diseased cohorts.
Surveillance Bias: When One Group Gets More Chances to Become a Case
A practical guide to surveillance bias for clinical researchers. Covers differential testing, follow-up intensity, diagnosis-based outcomes, and what reviewers should demand before trusting higher event rates.
Channeling Bias: When the Newer Treatment Inherits the Easier Patients
A practical guide to channeling bias for clinical researchers. Covers preferential prescribing, formulary-era drift, specialist selection, and what reviewers should demand before trusting observational comparisons of newer therapies.