WTP calculation accuracy using stated preference survey data: Comparison between logistic regression and neural network methods

Abstract

Valuing environmental recreational sites, which is also referred as non-market valuations, is widely used in environmental economics studies. Survey datasets are used to assess the price value of the site to an individual, which is referred as maximum willingness to pay (WTP), in order to understand demand for the sites. WTP calculations are used to decide on recreational site alterations, therefore an accurate WTP calculation is very important on the decision. The most common type for recreational site surveys is stated preference, where respondents are given two or more alternatives to choose one option out of them. Logistic regression is used in binary choice datasets because linear models don't work with binary outputs. Thus, traditional logit models are used most commonly with random utility models for stated preference discrete choice datasets. Neural networks on the other hand are improved logistic regression models, they are combination of logistic regressions with a learning process. Therefore, in order to calculate more accurate willingness to pay estimates, neural network models which are improved logistic regressions, could be used instead of simple traditional logit models. With this research, we wanted to test the superiority of neural networks on traditional logit in terms of willingness to pay estimation accuracy. In order to compare WTP calculation accuracies, we calculated out-of-sample prediction accuracies of the two methods and compared marginal effects and coefficients of covariates. Marginal effects of covariates would acquire us the effect of covariates on the probability of choosing one alternative, therefore they indicate the variable importance on WTP calculations. Coefficients on the other hand are used in WTP calculations because they represent the strength of particular covariate on the site choice. We examined two different articles that use stated preference survey datasets which are conducted for assessing the demand for recreational sites