Alternative Approaches to Deriving Welfare Estimates in Discrete Choice Experiments
University of Aberdeen - Health Economic Research Unit
University of Aberdeen
University of Aberdeen
iHEA 2007 6th World Congress: Explorations in Health Economics Paper
Rationale: Whilst Discrete Choice Experiments (DCEs) often apply conditional logit model (CLM) for multiple-choice experiments, the importance of considering alternative approaches has been recognised, including nested (NLM) and mixed logit (MLM) models. If alternatives are not perfect substitutes as stated with CLM, a partial solution is to use the NLM, where options are grouped in nests. The assumption of perfect substitution is still valid within nests, but not between nests. Another generalisation of CLM is MLM, introducing heterogeneity across individuals and allowing for multiple observations for each respondent.
Objectives: To consider welfare estimates when making different assumptions about the pattern of substitution across alternatives.
Methods: A DCE looking at preferences for extending the role of the pharmacists was used. Each choice offered three options: the current scenario; a novel community pharmacist and general practitioner review medicines (CPGP); a GP only medicines review (GP). Four substitution patterns were investigated: all options equal substitutes (pattern1); CPGP closer substitute to GP than current (pattern2); GP closer to current than CPGP (pattern3); CPGP closer to current than GP (pattern4). Data were analysed using conditional, nested and mixed logit models (with a flexible substitution pattern). Consideration is given to the impact of these models on goodness of fit, marginal rates of substitution and welfare estimates. Results are compared across 3 groups (control, intervention all and intervention still receiving the treatment).
Results: Nested and mixed logit models resulted in better fits than the conditional logit model. Compared to the conditional, for the nested Logit model, pattern2 had the best fit for the control and intervention all groups, whereas pattern4 was the best for the intervention still receiving the treatment group. Welfare estimates differed across models. For example, when moving from a very poor to a very high chance of receiving best treatment respondents were willing to pay: £30, £32 and £38 (conditional logit); £32, £29 and £18 (preferred nested models) and £29, £36 and £39 (mixed logit). WTP of moving from the current to GP and pharmacist collaboration in reviewing drugs and from current to an alternative GP only review were estimated. For example, in the first case respondents were willing to be refunded: £17, £20, £19 (conditional logit); £5, £2, or pay £50 (preferred nested models); £16, £20, £20 (mixed logit).
Conclusion: The importance of identifying patterns of substitution between alternatives when estimating welfare gains is discussed. The findings showed that different substitution patterns apply to different groups and welfare estimates differ across different econometric techniques. This has important implications on the contribution of DCEs to decision making.
Keywords: Discrete choice experiments, welfare estimates, methodology
Date posted: June 16, 2007