The widespread use of Stated Preference methods in environmental valuation continues to evolve, but also continues to suffer from a number of methodological challenges. One of these is the inherent respondent uncertainty in Stated Preference studies. The causes of respondent uncertainty, as well as particular ways to handle it, have received considerable interest in the in Contingent Valuation (CV) literature. In recent years, the Choice Experiment (CE) method has been increasingly used for environmental valuation, but so far apparently no studies have investigated the issue of respondent uncertainty in Choice Experiments.

            In this study, we use data sets from two independent Choice Experiment surveys to evaluate different ways of handling respondent uncertainty in CE. We use random parameter error component logit models to evaluate three different approaches to adapt the recoding-of-answers method from the CV literature to CEs. This is based on respondents’ self-reported degree of uncertainty in choice. We evaluate the effect of eliminating uncertain answers as well as an asymmetric and a symmetric recoding approach.  We compare these results to results obtained from two other approaches, where we, instead of recoding the data, take uncertainty explicitly into account in the econometric specification of the model. In the first of these two models, we do this by accounting for variation in scale as a function of respondents’ self-reported level of uncertainty in each choice set. The different levels of respondents’ stated uncertainty is directly parameterized in the model through an exponential formulation of the indirect utility function. This is done in order to maintain a panel specification even though we use uncertainty reported after each choice set which implies that the individual respondent can switch between uncertainty-groups through the course of the choice sets. Since most CE surveys do not ask respondents to state their experienced uncertainty after each single choice set, but rather at the end of the choice set sequence, we also evaluate a second modelling approach. In this second model, we use results obtained in a previous paper to explicitly model the scale parameter as a function of utility difference in the choice set, as well as respondent specific variables found to be significant determinants of uncertainty.

We find that elimination of uncertain answers increases model performance, but the effect on Willingness to Pay (WTP) is insignificant for most attributes in both surveys. The effect of asymmetric recoding is to reduce the overall WTP, as expected a priori, but again there is some variation across attributes. The effect on WTP of the developed approach to symmetric recoding is almost negligible for all attributes in both surveys. We find, however, that the explicit modelling of scale variations with stated uncertainty levels or underlying variables are equally successful in terms of model performance and precision in WTP estimates, and that these approaches outperform the recoding approaches.