Combining revealed and stated preferences to identify hypothetical bias in repeated question surveys: a feedback model of seafood demand: Timothy C. Haab, Bin Sun and John C. Whitehead
There are numerous valuation methodologies available to estimate the economic values of environmental changes for benefit-cost and other policy analyses. As this book has pointed out repeatedly, these methodologies can be roughly classified as revealed preference and stated preference approaches. The strengths of the revealed preference approach are the weakness of the stated preference approach, and vice versa. The joint estimation of revealed and stated preference data seeks to exploit the contrasting strengths of both approaches while minimizing their weaknesses. While there are benefits, experimental evidence also suggests some problems with the use of combined revealed and stated preference data. Among these problems, two crucial characteristics of the joint data tend to be ignored or misspecified. The first is hypothetical bias (Cummings et al. 1997; Whitehead et al. 2000). Hypothetical bias exists when stated preference questions are not incentive compatible, i.e. the respondent’s answer is not “independent of the use of a real or hypothetical referendum mechanism” (Cummings et al. 1997: 611). When hypothetical bias exists, answers to stated preference questions do not reflect the respondent’s actual behavior and predictions are inaccurate whether the stated preference data is used alone or in combination with revealed preference data. The second oft-overlooked problem is correlation among answers offered by the same respondent to different scenarios offered in sequential questions. Many revealed and stated behavior studies attempt to take advantage of economies of scale and extract as much information as possible from a single respondent by offering multiple questions. Correlation between responses seems obvious although many studies ignore such a connection, treating every response as independent. To incorporate correlation between responses in multiresponse surveys, panel data techniques (Whitehead et al. 2000; Whitehead et al. 2003) and first-order autocorrelation models (Boxall et al. 2003) have been used in estimation of combined revealed and stated preference data. The panel data models include an additional unobserved individual specific error in the estimation error, while the first-order autocorrelation model assumes the error term between responses for a
given individual is first-order serially correlated. However, both models ignore another connection between responses; the direct correlation between the answers. The answer to follow-up questions may be affected by the answer to previous questions.