Estimating preferences for a dermatology consultation using Best-Worst Scaling: Comparison of various methods of analysis

Abstract

Abstract Background Additional insights into patient preferences can be gained by supplementing discrete choice experiments with best-worst choice tasks. However, there are no empirical studies illustrating the relative advantages of the various methods of analysis within a random utility framework. Methods Multinomial and weighted least squares regression models were estimated for a discrete choice experiment. The discrete choice experiment incorporated a best-worst study and was conducted in a UK NHS dermatology context. Waiting time, expertise of doctor, convenience of attending and perceived thoroughness of care were varied across 16 hypothetical appointments. Sample level preferences were estimated for all models and differences between patient subgroups were investigated using covariate-adjusted multinomial logistic regression. Results A high level of agreement was observed between results from the paired model (which is theoretically consistent with the 'maxdiff' choice model) and the marginal model (which is only an approximation to it). Adjusting for covariates showed that patients who felt particularly affected by their skin condition during the previous week displayed extreme preference for short/no waiting time and were less concerned about other aspects of the appointment. Higher levels of educational attainment were associated with larger differences in utility between the levels of all attributes, although the attributes per se had the same impact upon choices as those with lower levels of attainment. The study also demonstrated the high levels of agreement between summary analyses using weighted least squares and estimates from multinomial models. Conclusion Robust policy-relevant information on preferences can be obtained from discrete choice experiments incorporating best-worst questions with relatively small sample sizes. The separation of the effects due to attribute impact from the position of levels on the latent utility scale is not possible using traditional discrete choice experiments. This separation is important because health policies to change the levels of attributes in health care may be very different from those aiming to change the attribute impact per se. The good approximation of summary analyses to the multinomial model is a useful finding, because weighted least squares choice totals give better insights into the choice model and promote greater familiarity with the preference data.