Pilot study of a hierarchical Bayes method for utility estimation in a choice-based conjoint analysis of prescription benefit plans including medication therapy management services
Consumers face an array of multiattribute prescription benefit insurance programs that include different access points (retail, supermarket, Internet, etc) and levels of pharmacist interaction (including medication therapy management services [MTMSs]). Because of this, there is a need for more sophisticated information to drive prescription benefit plan design.Objectives
A pilot study to determine if choice-based conjoint (CBC) analysis with hierarchical Bayes (HB) estimation for individual level part-worths could provide a stable model for attribute preferences for prescription benefit insurance; to pilot test the addition of MTMSs to a prescription benefit management model; and to pilot and compare logit-based utility estimates to HB estimations in a conjoint market simulator.Methods
A mail-based survey was conducted using a random sample of 1500 residents of the United States. A CBC analysis instrument was developed to provide a single-stated choice from a selection of different prescription benefit plans. Choice tasks were varied based on the attributes: co-payment, pharmacy access, formulary, level of pharmacist interaction including MTMSs and monthly premium. Analysis included logit-based and HB estimation for utilities, and preference share market simulation testing.Results
The utility estimations from HB analysis were consistent with those seen in the logit-based analysis. A goodness of fit of 83% (root likelihood) was achieved in the HB utility estimations with only 4 choice tasks per respondents and the inclusion of MTM-like services. There was convergence on preference shares from the market simulation between the 2 estimation methods.Conclusions
The use of CBC analysis with HB estimation provided utilities similar to those estimated using aggregated logit-based methods, with the added benefit of respondent specific part-worth scores for each attribute level. A larger sample, changes in the instrument design, more panels (tasks) per respondent, and selection of conjoint methods may allow for more predictive information from market simulators.