Deriving welfare measures in discrete choice experiments: a comment to Lancsar and Savage (2).

Lancsar and Savage (Deriving welfare measure from discrete choice experiments: inconsistency between current methods and random utility and welfare theory. Health Econ 2004; this issue) point out that the methods currently used to derive measures of welfare in the health economics literature are often inappropriate, and recommend the use of the expected compensating variation formula derived by Small and Rosen [1], especially when data are collected through discrete choice experiments. This is an important issue and the contribution of Lancsar and Savage (2004) may be a significant step to correct the misuse of welfare measures. Unfortunately, the work of Lancsar and Savage (2004) is marred by some errors and imprecisions which may be misleading. Indeed, the authors go too far in their criticism of the welfare measures currently used in the health economics literature and fail to point out some limitations of the Small and Rosen [1] method, which they advocate. Overall, I believe that Lancsar and Savage (2004) provide an unbalanced and poorly substantiated view of the problem. This note reconsiders the computation of measures of welfare in discrete choice problems and explores the relation between the marginal willingness to pay and the concepts of compensating variation (for an individual) and expected compensating variation (for a population). Moreover, some econometric issues related to the estimation of these welfare measures are also discussed. This aspect was ignored by Lancsar and Savage (2004), but has a critical role in the choice of the appropriate welfare measure to use in a particular application. In contradistinction to what is claimed by Lancsar and Savage (2004), it is argued that the marginal willingness to pay, the compensating variation and the expected compensating variation, can be important instruments to measure welfare changes. Like in Lancsar and Savage (2004), only the case of the multinomial logit model is considered, but the results can be easily extended to other settings.