A Nonparametric Welfare Analysis via a Randomized Conjoint Field Experiment: an Application to Water Quality Improvement and the Floating Settlements on Inlay Lake, Myanmar

Su Thet Hninn,, Shinji Kaneko, Keisuke Kawata, Yuichiro Yoshida
This study proposes a new approach to survey-based empirical welfare analysis, which combines a new design of the conjoint experiment of Hainmueller et al. (2014) and a non-parametric rational choice model. We focus on the welfare impact of a multi-attribute policy and report the identification result of the marginal component effect on the distribution of willingness-to-pay. As an illustration, the paper evaluates a water improvement policy package for the not previously analyzed floating
more » ... lyzed floating settlements on Inlay Lake, Myanmar. Our estimation result shows that the average surplus gain from the lake water quality improvement is at least as large as 5.9% of the average annual per capita income of those on the lake. Moreover, attributes such as toilet provision have a clear welfare effect. JEL Codes: Q53, Q56, Q58 The estimation of policy preferences is a central challenge in various policy studies including those in political science, economics, environmental science, and development studies. A difficulty in the estimation is multi-dimensionality: a policy typically has multiple attributes, for instance, an environmental policy consists of its target area, period, and cost. To evaluate a multi-attribute policy, the present paper connects a newly developed survey design and commonly used measurement in policy evaluation, namely, Hainmueller et al. (2014)'s design of the conjoint experiment and willingness-to-pay (WTP). Recently, a conjoint experiment and the associated analytical methods offered by Hainmueller, Hopkins, and Yamamoto (2014) (referred as an HHY conjoint here after) has become a popular survey experiment design to estimate preferences for multi-attribute policy 1 . Their approach non-parametrically estimates the causal effect of individual policy attributes on observable choice outcomes (e.g., rating and/or ranking among alternative policies). The advantage of the HHY conjoint is its design-based inference; while the causal inference in traditional conjoint experiments is based on explicit decision-making models (e.g., the random utility model; see Train 2000), HHY conjoint inference is not. It is instead, based on the survey design. All causal quantities of interest are defined and identified based on Rubin's potential outcome framework (Neyman 1923; Rubin 1974) , and their estimation result is then free from bias due to specification error in the decision-making model 2 . However, design-based approaches have had the common limitation that the approach cannot estimate the causal effect on "unobservable" outcomes, of which, probably the most relevant from policy implementation perspective is WTP. WTP is the maximum amount that people are willing to pay for policy implementation, which is a common measurement in policy evaluation among not only academic researchers but also policy practitioners (see Mishan and Quah 2007). Conventionally, WTP is defined over an explicitly-specified decision-making model, while the original inference of HHY conjoint does not assume any explicit specification. Conventional conjoint analyses are based on specific choice models that require parametric assumptions on the preference distribution. For wexample, the conventional random utility model assumes that the preference for observable characteristics is additive separable, the preference for unobservable characteristics follows the type-I extreme value distribution, and the preference parame-1 An increasing number of studies use the new randomized conjoint analysis to examine policy preferences. These include studies on international environmental agreements (Bechtel and Scheve 2013, Gampfer, Bernauer, and Kachi 2014, Bernauer and Gampfer 2015) and migration policies (Hainmueller and Hopkins 2015). 2 Hainmueller et al. (2015) provide evidence that the HHY conjoint performs well at replicating real-world behavior by comparison with natural experiments.
doi:10.15027/45447 fatcat:bsrvfnll6zemlc7bdyulj3vg3u