OBJECTIVES:  The objectives of this study were to investigate whether medication adherence and persistence differed across: (1) responders vs. non-responders; and (2) early vs. late responders to a survey assessing medication and health beliefs. METHODS: A survey assessing medication and health beliefs was mailed to patients who filled a qualifying index prescription for one of five chronic conditions at one national and two regional retail pharmacies in 2008. Adherence and persistence to the index drug class was measured using pharmacy claims data over 12 months. A multivariate generalized linear model with a negative binomial distribution and log-link function was used to determine the significant predictors of adherence. Kaplan-Meier estimates of survival (persistence) curves were used to assess the time to discontinuation, and the multivariate Cox proportional hazards model was used to identify significant predictors of non-persistence. To assess differences between early vs. late responders, medication beliefs were compared across timing quartiles based on survey response time (date signed minus date mailed). RESULTS: A final survey response rate of 24.25% was achieved. Survey non-responders had statistically significantly lower medication adherence than responders. As assessed by the log-rank test, the Kaplan-Meier estimate of the persistence curve for responders was statistically significantly higher (indicating better persistence) than the one for non-responders These results were confirmed by the multivariate Cox proportional hazards model for time-to-discontinuation outcome. Compared to early survey responders, late responders reported less perceived need (p=.003), more medication concerns (p=.031), less prescription-medication knowledge (p=.008), and less trust in their prescribing physician (p=.002).  CONCLUSIONS: Our study found evidence of non-response bias in a survey assessing patients' medication and health beliefs. Non-response bias by adherence status can result in biased estimates of medication beliefs. We recommend using multiple survey levers to increase response rate among non-adherent patients to reduce the potential for non-response bias.