Behavioral profiles associated with adherence to adjuvant endocrine therapy in breast cancer: a retrospective population-based cohort study

Technical Summary: Behavioral Profiles Associated with Adherence to Adjuvant Endocrine Therapy in Breast Cancer

Problem Statement
Adjuvant endocrine therapy (AET) is a critical component of long-term care for hormone receptor-positive breast cancer, significantly reducing recurrence and mortality. However, adherence remains a persistent challenge, with approximately one-third of patients failing to meet adherence criteria within the first year. While clinical and sociodemographic factors are known to influence adherence, current interventions often yield mixed results. A primary limitation in existing literature is the inability to characterize how multiple behavioral factors combine within individual patients. Previous studies have relied on single-region cohorts, geographically aggregated estimates prone to ecological bias, or survey designs that impose additional burdens on patients. Consequently, there is a lack of high-resolution, individual-level data linking behavioral determinants to adherence outcomes across a national population.

Methodology
The authors conducted a retrospective, population-based cohort study utilizing a novel data linkage approach.

• Data Sources: The study linked two large-scale datasets via privacy-preserving, hash-based tokenization:
  1. BehavioralPulse™ (Surgo Health): A national dataset providing modeled individual-level behavioral propensity scores (e.g., health beliefs, perceived barriers, social support, trust in clinicians) and consumer-level sociodemographic attributes for US adults.
  2. Komodo Healthcare Map®: Longitudinal claims data (inpatient, outpatient, pharmacy) for US women diagnosed with invasive breast cancer between January 2016 and April 2025.
• Cohort Definition: The analytic cohort included 401,450 women who initiated standard oral AET (tamoxifen or aromatase inhibitors) following surgery for invasive breast cancer. Inclusion required continuous insurance coverage for 365 days post-initiation and evidence of at least two breast cancer diagnosis codes. Patients with ovarian function suppression or three or more surgeries were excluded.
• Outcome Measure: The primary outcome was medication adherence, defined as achieving a Medication Possession Ratio (MPR) of ≥80% during the first 365 days of therapy.
• Statistical Analysis:
  • Predictor Selection: Fifty behavioral variables were initially considered. To mitigate multicollinearity, variables with pairwise Pearson correlations >0.7 were iteratively removed, leaving a set of significant behavioral predictors. These were binarized at the population mean.
  • Modeling: A mixed-effects logistic regression model was employed to estimate adjusted odds ratios (aORs) for adherence. The model controlled for age, race/ethnicity, insurance type, geographic location (state/ZIP3), comorbidities (anxiety, depression, arthralgia, etc.), and consumer sociodemographics. A ZIP3-level random intercept accounted for geographic clustering.
  • Profile Identification: The study identified the 10 most common configurations of significant behavioral predictors to define patient profiles. Relative adherence odds for these profiles were estimated using Monte Carlo simulation.
  • Validation: Model performance was assessed via ROC-AUC on a held-out test set, and robustness was tested through sensitivity analyses (e.g., using z-scored variables, excluding covariates with missingness).

Key Results

• Adherence Rates: Of the 401,450 women, 280,595 (69.9%) achieved an MPR ≥80%.
• Behavioral Determinants: After adjusting for clinical and sociodemographic factors, eight behavioral factors were independently associated with adherence:
  • Positive Associations: Comfort with following medication instructions (aOR 1.15), instrumental social support (aOR 1.06), trust in doctors (aOR 1.04), and religiosity (proxied by interest in Bible-related materials; aOR 1.04).
  • Negative Associations: Concerns regarding sexual side effects (aOR 0.96), cost-related access barriers (e.g., inability to see a doctor due to cost; aOR 0.97), and unexpected high medical bills (aOR 0.97).
  • Unexpected Finding: Higher interest in health and fitness was associated with lower adherence (aOR 0.94), suggesting a complex relationship between health orientation and long-term medication persistence.
• Patient Profiles: The ten most common behavioral profiles accounted for >70% of the patient population. While individual behavioral associations were modest, the cumulative impact was substantial.
  • The most prevalent profile (22% of patients) was characterized by the absence of cost barriers, no concern about sexual side effects, and high social support.
  • Profiles with cost-related barriers showed consistently lower adherence propensity.
  • Crucially, different combinations of barriers could result in similar adherence risks (e.g., a profile with two cost concerns had a similar adherence risk profile to one without cost concerns but with a health-and-fitness orientation).
• Model Performance: The model demonstrated modest discrimination (ROC-AUC 0.59), and geographic clustering (ZIP3) contributed minimally to explained variance.

Significance and Claims
The paper claims that by linking individual-level behavioral data to large-scale claims data, it is possible to identify non-clinical drivers of adherence that persist even after controlling for traditional risk factors. The study's primary contribution is the identification of recurring behavioral profiles, demonstrating that patients with similar adherence risks may arrive at that state through substantively different pathways.

The authors argue that this heterogeneity explains why previous adherence interventions have produced inconsistent results; aggregate outcomes may appear similar, but the underlying drivers differ. Consequently, the paper suggests that adherence support strategies should move away from targeting an "average" patient or a single domain (e.g., financial assistance alone). Instead, interventions should be tailored to specific behavioral profiles, matching support mechanisms—such as financial navigation, side-effect management, or social resource allocation—to the specific combination of barriers faced by distinct patient segments.

Limitations
The authors acknowledge several limitations: the observational nature of the study precludes causal inference; behavioral scores are modeled probabilities rather than direct measurements; the cohort is limited to insured populations (excluding self-pay/uninsured); potential survivorship bias due to the exclusion of early mortality; and the use of a binary MPR threshold which obscures variation in persistence timing. Additionally, the model's modest discrimination suggests that unmeasured clinical factors (e.g., tumor stage) remain influential.