Healthy Adherer Bias

In plain language

Healthy adherer bias means the people who keep taking a medicine are often healthier and more organized in many other ways too. They may keep appointments, get screenings, exercise more, and follow other health advice. If they have fewer bad outcomes, the study can wrongly credit the drug instead of the broader pattern of healthy behavior.

Healthy adherer bias

is the adherence-specific member of the healthy-user family. It appears after treatment starts, when the patients who keep refilling or taking a drug differ from those who stop, gap, or take it inconsistently. The difference is not only pharmacologic exposure. Adherers tend to attend visits, follow preventive advice, use seat belts, maintain routines, have better access, and avoid behaviors that raise risk. A comparison of adherent versus nonadherent patients therefore estimates a mixture of treatment exposure and the hidden traits that made adherence possible.

Core conceptual distinction

Healthy user bias is often about who starts treatment; healthy adherer bias is about who persists after starting. That timing matters. Baseline confounding can be attacked with active-comparator new-user design and pre-index covariates. Healthy adherer bias enters through post-index behavior, so naive adjustment for adherence or censoring at nonadherence can create a biased per-protocol estimate. The classic warning sign is placebo adherence: people who faithfully take placebo have lower mortality, which cannot be caused by an inert pill. The same warning appears in statin studies where adherers have fewer accidents.

Pros, cons, and trade-offs

- Adherence as behavior marker vs adherence as causal exposure. PDC, MPR, and persistence are meaningful quality and implementation measures, but they are dangerous as causal exposures. If the question is "does taking more drug help?", adherence must be separated from the traits that predict adherence. - Intention-to-treat-style contrast vs per-protocol/as-treated contrast. An initiation-based contrast preserves the initial treatment decision and is less exposed to post-baseline adherence selection. Per-protocol contrasts are closer to the biological effect of sustained exposure but require censoring weights, protocol definitions, and falsification diagnostics. - Landmark adherence vs future adherence. Measuring adherence during a fixed early landmark window can avoid some immortal time, but the landmark cohort excludes early failures and deaths. Using full follow-up adherence as exposure is usually invalid because future survival is required to become adherent. - Negative controls vs proxy adjustment. Preventive visits, screenings, and prior adherence to unrelated chronic therapies are useful proxies, but only negative-control outcomes can show that an adherer advantage persists on outcomes the drug cannot affect.

When to use

Diagnose healthy adherer bias whenever adherence, persistence, PDC, MPR, dose intensity, or treatment continuation is used as an exposure or as a censoring rule in chronic, preventive, cardiovascular, vaccine, osteoporosis, diabetes, oncology maintenance, or specialty drug studies. It is especially important for mortality, cardiovascular events, fractures, hospitalizations, and cost/HCRU endpoints because those outcomes are strongly affected by the same access and behavior patterns that predict adherence.

When NOT to use - and when it is actively misleading

- Do not report adherent-vs-nonadherent benefit as a causal drug effect without a design for post-baseline selection and at least one negative-control outcome. - Do not adjust for follow-up adherence as if it were an ordinary confounder in an initiation study; it can be a mediator, collider, or marker of unmeasured health behavior. - Do not define exposure using future PDC over the entire outcome window. That gives adherers guaranteed survival and observation time. - Do not assume censoring at nonadherence fixes the problem. It creates informative censoring unless adherence predictors and outcome predictors are fully captured and weighted.

Data-source operational depth

Claims are strong for pharmacy refill adherence but weak for actual ingestion, samples, inpatient drug use, and reasons for stopping. Medicare Advantage or pharmacy carve-outs can make nonadherence appear as missing fills. EHRs can add clinician notes, side effects, labs, and patient-reported nonadherence, but documentation is visit-driven and patients with more visits look more adherent and more observable. Registries may capture protocol adherence and disease severity but often miss external fills. Linked claims-EHR-vital records are best for adherence, outcomes, and mortality, but linkability itself can select more engaged patients.

Worked example

Scenario

A claims study asks whether high adherence to statins reduces one-year hospitalization after initiation. The analyst classifies patients by PDC in the first 180 days and then compares hospitalizations from day 181 to day 365.

Dataset

Illustrative landmark comparison among statin initiators who survived and remained observable through day 180.

Steps

• The high-adherence group has fewer hospitalizations, but also had many more preventive visits before follow-up.

• The high-adherence group has fewer accidents after the landmark, an outcome statins should not prevent.

• The accident result is a falsification signal that adherence is marking broader health behavior.

• A causal per-protocol analysis would need pre-specified adherence strategy, landmark or clone-censor-weight design, censoring weights, and negative-control reporting.

Result

The apparent 6 percentage-point hospitalization advantage cannot be interpreted as a pure statin effect because the 4.5 percentage-point accident advantage indicates residual healthy adherer bias.

Runnable example

import pandas as pd
import statsmodels.formula.api as smf

def adherence_negative_control(landmark_df):
    # landmark_df has one row per patient alive/observable at the landmark.
    # Required columns: high_adherence, outcome_event, negative_control_event,
    # age, cci, prior_hcru, prior_preventive_visits
    models = {}
    for y in ["outcome_event", "negative_control_event"]:
        models[y] = smf.logit(
            f"{y} ~ high_adherence + age + cci + prior_hcru + prior_preventive_visits",
            data=landmark_df,
        ).fit(disp=0)
    return models