OBJECTIVES: Medication adherence has been linked to better health outcomes. Therefore, a comprehensive understanding of the predictors of adherence is essential for formulating adherence-improving strategies. Existing methods have not considered evaluation of heterogeneous impacts of predictors at different parts (quantiles) of the adherence distribution as defined by medication possession ratio. Using the novel econometric framework of unconditional quantile regression (UQR), this study assesses the heterogeneity of impacts of adherence predictors for an Alzheimer’s disease (AD) population. METHODS: This retrospective claims analysis identified AD patients from a large US health plan that initiated oral AD therapy (rivastigmine, donepezil, galantamine, or memantine) between 1/1/2006 and 12/31/2007. Baseline characteristics were assessed during the 6-month pre-index period; medication adherence was assessed during the 1-year post-index period. UQR was estimated at 10th, 20th, …, 90th quantiles. Predictors of adherence identified from the data included age, gender, indicator of mental health insurance coverage, region, commercial vs. Medicare insurance, log cost, comorbidity, and formulary tier for the AD medication. RESULTS: Baseline medication count was positively associated with adherence (p<0.05) in the upper half of the adherence distribution. Having mental health coverage is negatively associated with adherence in all but the 10th and 20th quantiles but the impact was substantially higher in the first half of the adherence distribution. Baseline (log) cost was positively associated with adherence in the 40th and upper quantiles of the adherence distribution. For patients in the 80th and 90th quantiles, the number of baseline office visits predicted lower adherence. Compared to patients from the East, patients from the South were less likely to be adherent in the 60th and 70th quantiles. CONCLUSIONS: The study results highlight the heterogeneity of impacts of various adherence predictors – a predictor may be statistically significant only in specific quantiles of the adherence distribution, and the impacts may vary substantially between quantiles.