Bias correction for welfare measures in non-market valuation
MetadataShow full item record
Welfare measures in non-market valuation can be biased by nonrandom sampling or nonlinear transformation of estimated parameters. This dissertation consists of two studies focusing on correcting each type of bias. The first study compares the performances of the delta method, jackknife, and bootstrap in correcting nonlinear transformation bias through Monte Carlo simulation. The simulated models include Poisson, logit, probit, and misspecified probit. The results indicate that the delta method and jackknife can correct the bias of welfare measures for the listed models. With distributional misspecification, the delta method outperforms the jackknife. The bootstrap result has more expected bias, but also the smallest mean square error, especially when the sample size is large enough. Furthermore, the delta method was modified by correcting the bias of estimated parameters beforehand, and then adding the first order bias term or replacing parameter estimates with corrected ones. The simulation results show that adding first order bias term can lower the MSE of willingness to pay(WTP) estimates at the cost of increasing their expected bias. The simulation results show that the delta method is also effective to estimate median WTP. The second study address the sampling bias decomposed as one part correlated with covariates and the other part correlated with the result variable. Non-response is common in surveys used in non-market valuation studies which can bias the parameter estimates and mean WTP estimates. One approach to correct this bias is to reweight the sample so that the distribution of the characteristic variables of the sample can match that of the population. Kernel Mean Matching (KMM) is used to produce resampling weights in a non-parametric manner. KMM’s performance is tested through Monte Carlo simulations under multiple scenarios and it is shown that KMM can effectively correct mean WTP estimates, especially when the sample size is small and sampling process depends on covariates. The results also reveal KMM’s robustness to skewed bid design and model misspecification.