OBJECTIVES: Well-defined attributes are a cornerstone of choice-format conjoint studies (discrete-choice experiments).  However, little is known about how ill-defined attributes affect the quality of stated-choice data.  The objective of this study was to evaluate choice responses when attribute levels in the choice options do not align with a respondent’s personal current reference condition.  METHODS: A total of 193 US adults with self-reported epilepsy who reported taking antiepileptic drugs (AEDs) to treat their seizures completed a web-enabled, choice-format conjoint survey.  Respondents responded to a series of choice questions to assess their willingness to add another AED to control their seizures.  Consistent with endpoints measured in a clinical trial, the efficacy attribute was defined as a 100%, 75%, 50% or 25% reduction in the number of seizures relative to the respondent’s current reference condition: each respondent’s own number of seizures over the last 3 months.  Other treatment outcomes included short-term and long-term side effects, difficulty urinating, weight change, dosing frequency, and personal cost of the add-on AED. RESULTS: Respondents’ current treatment generally achieved good seizure control.  Only one-third of the respondents reported having any seizures and 20% reported having only 1 to 3 seizures. Nevertheless, aggregate preference estimates for the 100%, 75%, 50% or 25% seizure-reduction levels were correctly ordered, statistically different from each other (with one exception), and were of sensible magnitudes relative to other treatment outcomes.  Statistical tests indicated that there was no difference in preferences between respondents with < 4 seizures and those with 4 or more seizures. CONCLUSIONS: We have no information about how patients with few or no seizures evaluated meaningless percentage reductions in their reference condition.  Their reinterpretation of the attribute levels, however, resulted in the same preference estimates as those for whom the levels were consistent with their experience.  Thus, our respondents may just fix our mistakes.