Comparison of Preferences and Data Quality between Discrete Choice Experiments Conducted in Online and Face-to-Face Respondents

INTRODUCTION: Discrete choice experiments (DCE) are increasingly being conducted using online panels. However, the comparability of such DCE-based preferences to traditional modes of data collection (e.g., in-person) is not well established. In this study, supervised, face-to-face DCE was compared with its unsupervised, online facsimile on face validity, respondent behavior, and modeled preferences. METHODS: Data from face-to-face and online EQ-5D-5L health state valuation studies were compared, in which each used the same experimental design and quota sampling procedure. Respondents completed 7 binary DCE tasks comparing 2 EQ-5D-5L health states presented side by side (health states A and B). Data face validity was assessed by comparing preference patterns as a function of the severity difference between 2 health states within a task. The prevalence of potentially suspicious choice patterns (i.e., all As, all Bs, and alternating As/Bs) was compared between studies. Preference data were modeled using multinomial logit regression and compared based on dimensional contribution to overall scale and importance ranking of dimension-levels. RESULTS: One thousand five Online respondents and 1,099 face-to-face screened (F2F(S)) respondents were included in the main comparison of DCE tasks. Online respondents reported more problems on all EQ-5D dimensions except for Mobility. The face validity of the data was similar between comparators. Online respondents had a greater prevalence of potentially suspicious DCE choice patterns ([Online]: 5.3% [F2F(S)] 2.9%, P = 0.005). When modeled, the relative contribution of each EQ-5D dimension differed between modes of administration. Online respondents weighed Mobility more importantly and Anxiety/Depression less importantly. DISCUSSION: Although assessments of face validity were similar between Online and F2F(S), modeled preferences differed. Future analyses are needed to clarify whether differences are attributable to preference or data quality variation between modes of data collection.