Patient-reported outcome (PRO) instruments such as the WOMAC Index of Osteoarthritis provide scores for outcomes that cannot be measured objectively.  However, such scores do not necessarily indicate patients’ perception of the relative importance of the domains and scales used in such instruments. OBJECTIVES: To obtain valid estimates of patient preferences for osteoarthritis domains and potential adverse outcomes.  METHODS: Residents of the United Kingdom aged 45 years or older with a self-reported physician diagnosis of osteoarthritis (OA) completed a web-enabled conjoint-analysis or discrete-choice experiment survey.  Patients evaluated 10 pairs of hypothetical OA treatment-outcome profiles. Each hypothetical OA treatment was defined by 4 benefit attributes derived from the WOMAC instrument (ambulatory pain, resting pain, stiffness, and daily activities) plus 3 adverse-event risk attributes (bleeding ulcer, stroke, and coronary infarction). Choice-model parameter estimates were obtained using random-parameters logit. RESULTS: A total of 294 subjects provided usable data. Valid parameter estimates should indicate that better outcomes are preferred to worse outcomes.  A main-effects choice model yielded correctly ordered, risk-level, preference-weight estimates for the 3 adverse-event attributes, but disordered severity estimates for all 4 PRO attributes.  Implausible combinations of ambulatory-pain and daily-activity levels may cause patients to focus on only one of the attributes.  A model that excludes the daily-activity attribute and includes a dummy variable to account for  implausible combinations yields correctly ordered preference-weight estimates. An improvement in ambulatory pain from 25 mm to 0 mm on a 100 mm scale is approximately 1.7 times as important to patients as the same improvement in resting pain.  CONCLUSIONS: Confounding among PRO domains can interfere with estimating valid preference weights and limit researchers’ ability to populate a model based on PRO outcomes with a full set of preference weights Controlling statistically for implausible outcome combinations  successfully corrects the effect of the confounding and yields plausible trade-off values.