In this dissertation, we focus our goal on improving the econometric estimation as well as the calculation of benefit transfer (BT) predictions. In the first essay, we conduct a meta-analysis for U.S. wetlands by analyzing 72 observations collected in Borisova-Kidder (2006). Since the value estimates in 72 observations are generated from various valuation methods, whether they can be pooled in the estimation of meta-equation is further examined in this essay through bootstrapping Hausman’s test. Our results indicate that the value estimates produced by different valuation methods, although they might not be directly comparable, can be pooled in the estimation of meta-equation without undermining the efficiency and consistency of model parameters.
In the second essay, we propose to improve the efficiencies of both the model parameter estimation and the BT predictions in our wetland meta-equation by introducing additional information contained in Brander et al.’s 215-observation global dataset. To avoid our goal of studying U.S. wetland values being undermined by those non U.S. wetland values, we turn to a Bayesian modeling framework to allow the additional information to enter our model estimation through informed priors. The results indicate that models with the additional information score higher logged marginal likelihood values than their respective counterparts. This suggests that our model performance is improved by increasing the probability to observe the sample points in our dataset. Besides, the results also indicate that the accuracy of model forecasting is improved by 5%. Moreover, it is also noted that the 90% confidence intervals for models with the additional information have been narrowed down by 54%.
In the third essay, we apply Bayesian model averaging (BMA) technique to account for model uncertainty in the derivation of the effects of estimated coefficients as well as the calculation of BT predictions by utilizing information from all possible models, each weighted by its posterior model probability. In this way, we can free ourselves from making inference and calculating BT predictions based on a single model. Thus, non-diluted effects of estimated coefficients as well as more meaningful and representative BT predictions can be obtained. The results indicate that for variables receiving higher posterior effect probabilities, the size of their estimated coefficients will be close to those in the FULL model (with a full set of variables) estimated by OLS. For variables with lower probabilities, their estimated coefficients are smaller compared to their respective counterparts. Moreover, the results also indicate that BMA technique, compared to the FULL model, not only provides better out-of-sample predictive performance, but also narrows down the 90% confidence internal by 70% in BT predictions.