I'll Tell You Which Patient is Most Bothered About Waiting Time: Estimating Individual Preferences from Discrete Choice Experiments

Abstract

Introduction: Discrete choice experiments (DCEs) are increasingly used to value attributes of health and health care. Generally these cannot be used to estimate individual patient preferences robustly for two reasons. First, 'pick one' tasks are relatively inefficient at eliciting preferences. Second, statistical designs for non-linear choice models previously were not widely available. However, adding best-worst choice tasks to DCEs that use statistically efficient designs can overcome this limitation. We illustrate this for the case of a DCE conducted among UK NHS patients awaiting a secondary care dermatology appointment. Patients increasingly can choose hospital consultant-led care with a longer waiting time or treatment from a GP with a special interest (GPSI) with a shorter waiting time. Little is known about individual patients' relative preferences for waiting time versus doctor expertise but these individual preferences (rather than averages across patients) will determine the acceptability of the policy. Our aim was to estimate the effects of introducing patient choice within dermatology secondary care services by estimating individual patient preferences for key attributes of a dermatology consultation. Methods: We conducted an archetypal analysis and latent class analysis of a DCE using best-worst scaling (BWS) for 138 UK NHS patients. Key outcome measures were willingness to wait (in months) to receive an appointment defined by higher levels of three 'process-based' attributes - doctor expertise, convenience of attending and thoroughness of care - for subgroups of patients identified by archetypal and latent class analysis. Results: We identified three patient archetypes: patients in archetype 1 (n=56) strongly preferred a thorough consultation, being willing to wait longer than current guidelines permit (3 months) to have this. Patients in archetype 2 (n=42) valued being seen by a consultant-led team and were willing to wait longer than 3 months. Only patients in archetype 3 (n=40) strongly preferred shorter waiting times. We had difficulty identifying a stable 3-class latent class analysis solution. In the 2-class solution, class 1 (n=98) patients focused on thoroughness and class 2 patients (n=40) strongly preferred zero waits. Interestingly, the two taxonomic methods did not always identify the same patients as the ones who valued waiting time. Characteristics of patients like age leaving full-time education and self-reported symptom severity were found to be significant predictors of archetype and class membership (and hence willingness to trade). Conclusions: Without data on thoroughness of individual doctors and given likely waiting times for hospital-led care, patients with severe self-reported symptoms are the ones most likely to choose GPSI-led care. Archetypal analysis expresses each patient's preferences as a weighted average of one or more actual (archetypal) patients in the data, which is a taxonomic method that is intuitive to clinicians and policy makers. Future work will use individual patient preference estimates from BWS to investigate further the relative merits of different taxonomic methods.