OBJECTIVES Patient preferences have implications for treatment decision making, treatment adherence and follow-up care.  This study aimed to highlight, using metastatic breast cancer (mBC) as an example, a method to elicit preferences and, of particular novelty, examine individual differences of those preferences. METHODS Using mixed methods, a qualitative study (n=10) of patients with mBC informed the development of the preference survey (a cross-sectional Internet survey administered to women with mBC).  Survey participants (N=181) completed a conjoint exercise that included a series of choice questions. Each choice question included a pair of hypothetical treatments that were presented in terms of eight safety attributes, single attributes for effectiveness, dosing regimen, and quality of life .  Survey choice data were analyzed using hierarchical Bayesian logistic regression models.  Predicted values from this model were then analyzed to understand individual differences in patient preference. RESULTS Qualitative interviews identified the most relevant side effects to include in the choice task (e.g., alopecia, nausea/vomiting, etc.) and reinforced the importance of quality of life when making treatment decisions. In the survey data, treatment effectiveness was most strongly associated with treatment preference, followed by alopecia, fatigue, neutropenia, and quality of life. Predicted values from the choice model enabled preference comparisons across treatment experience subgroups (e.g., 6+ rounds of chemotherapy vs. less). Preference strength for individual attributes, e.g., side effects was correlated with various demographic and health history variables, though only modest associations were detected (Pearson rs<0.25).  CONCLUSIONS Understanding patient preferences provides opportunities for improved care and outcomes. Combining qualitative and quantitative methods in this study allowed for specificity of preferences and generalizability (albeit limited).  Patient preferences derived across the sample informed  predicted values from the choice models that can also be used for comparing preferences across subsamples and identifying factors that may be associated with certain preferences.