ESSAYS ON IMPROVING THE ECONOMETRIC ESTIMATION OF WETLANDS VALUES VIA META ANALYSIS DISSERTATION Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduat[.]
ESSAYS ON IMPROVING THE ECONOMETRIC ESTIMATION OF WETLANDS VALUES VIA META-ANALYSIS DISSERTATION Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the Graduate School of The Ohio State University By Ding-Rong Chen, M.A Graduate Program in Agricultural, Environmental, and Development Economics The Ohio State University 2010 Dissertation Committee: Professor Alan Randall, Advisor Professor Tim Haab Professor Abdul Sam Copyright by Ding-Rong Chen 2010 ABSTRACT Historically, wetlands have been considered wastelands and can only be improved by drainage and then converted into production lands This situation has changed since the 1970’s due to the public becoming aware of the contributions of wetlands through the provision of various functions, which include habitats for species, protection against floods, water purification, amenities, recreational opportunities, and etc Since these functions not only bring significant improvements to the ecological system but can also further the living qualities for human beings, a shift from conversion to conservation policy has become clear to the U.S government in the use of wetlands In the past few decades, a large number of original site-specific wetland studies were conducted to provide value estimates of wetlands for policy makers’ reference to balance the benefits and costs of wetlands conservation so that they can determine an efficient allocation of scarce public resources to wetlands projects However, given the increasing costs of the original site-specific benefit studies, public agencies have expressed a strong interest in generalizing the findings already in the literature, in the hope of cheaper and quicker first-cut benefit approximations for non-valued policyrelevant sites This idea has stimulated the use of meta-analysis for benefit transfer (BT) Meta-analysis is a popular tool for wetland valuation since it can synthesize empirical results across a number of studies that address related research questions In the literature, the goals of most existing U.S wetland meta-analysis studies are aimed to distill the systematic trends from previous site-specific studies as well as to provide BT predictions through their estimated meta-equations Examples of studies within this framework are Brouwer et al (1999), Woodward and Wui (2001), Borisova-Kidder (2006), and Brander et al (2006) In this study, we focus our goal on improving the econometric estimation of BT through solving several issues we encountered while ii dealing with our wetland meta-dataset The potential issues are raised in the following three essays and the corresponding solutions are also proposed in each essay In essay 1, we conduct a meta-analysis for U.S wetlands by analyzing 72 observations collected in Borisova-Kidder (2006) to identify the important determinants of wetland values Since the 72 observations in Borisova-Kidder (2006) are stemmed from a number of published studies that applied different methodologies to elicit wetland values, whether pooling these multi-source observations will undermine the consistency and efficiency of model parameters is also examined in this essay To avoid the possible asymptotic issue in our small dataset, this essay examines the effect of valuation method in our wetland meta-analysis through bootstrapping Hausman’s test to see if pooling observations is adequate in the wetland meta-analysis The major findings in essay can be summarized as follows First, an increase in household income will lead to higher wetland values Second, saltwater marsh and prairie pothole, compared to freshwater marsh, were shown to have a negative relationship with wetland value Third, most of the wetland services except water quality and amenity were shown to have no significant influence on wetland value Fourth, among valuation methods only energy analysis was shown to have a significant effect on wetland value Last, we cannot find any evidence for wetland value being significantly affected by its region Other geographical factors, e.g distance from population center, may have impact For the pooling effect, our test results suggest that the value estimates (observations) 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 essay 2, we aim to improve the efficiencies of both the model parameter estimation and the BT predictions in our wetland meta-equation by overcoming several issues in our wetland meta-dataset In fact, the dataset used in essay confronts us with the following two difficulties First, the number of observations may be smaller than ideal for estimating effects on valuation methodology, wetland type, services, and etc Second, this dataset is relatively noisy In order to tackle the difficulties, we propose to augment our study with information from a second dataset in a published wetland meta-analysis conducted by Brander et al (2006) for introducing additional information to our thin iii dataset However, the analytical method in essay 1, while a common practice of wetland valuation in the literature, is impeded by some technical limitations in our context and suffers from inability to meet our objectives Therefore, in this essay, we turn to a Bayesian modeling framework to allow information monotonically extracted from the estimated meta-equation in Brander et al to enter our study through informed priors to carry out our U.S wetland meta-analysis The results of essay showed that models with the added information borrowed from Brander et al score higher logged marginal likelihood values than their respective counterparts This suggests that such added information can effectively improve our model performance by increasing the probability to observe the sample points in our dataset Besides, our results also showed that the accuracy of model forecasting is improved by 5% in MAPEs when the added information has been introduced into our model Moreover, it is also noted that the 90% confidence intervals for models with the added information from Brander et al have been narrowed down by 54% In essay 3, we apply Bayesian modeling averaging (BMA) to mitigate the model uncertainty issue in the derivation of estimated coefficients as well as the calculation of BT predictions In essay 1, the results showed that there are more than 80% of explanatory variables in the model not significant at 90% level of significance This implicitly suggests that our data is generated with no strong relationship between wetland values and those insignificant explanatory variables The traditional solution for the above situation is to apply a sequence of statistical variable selection exercises to help select a single best model with most or all irrelevant variables being removed However, such a solution confronts us with the model uncertainty issue in the following two aspects First, there is no theory suggesting which variables should be retained in wetland metaregression model Thus, we are left with an arbitrary choice among various variable selection methods Second, basing inference on a single model is inherently risky unless other not quite so good (or non-selected) models receive no support from data Thus, in this essay, we apply BMA method to utilize information from all possible models, each weighted by its posterior model probability The results of essays showed that for variables receiving higher posterior effect probabilities, the size of their estimated coefficients will be close to those in the FULL iv 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, our results also showed that BMA technique not only provides better out-of-sample predictive performance than other single models selected by Adj-R2, Mallow’s Cp and stepwise regression, but also narrows down the 90% confidence internal by 70% in BT predictions obtained from the FULL model This is because BMA uses information from the insignificant variables but does not assign them excessive weight Naturally, we also would like to address some limitations in this study First, the prior calculation can be more precise if the variance-covariance matrix of estimated coefficients in Brander et al can be obtained This is because some prior information used in this study came from the combination of two or more variables in Brander et al.’s meta-equation and the derivation of their variances requires the summation of variances and co-variances over several variables Second, the model uncertainty empirically consists of at least two elements: specification uncertainty and functional form uncertainty In this study, we only address the first one However, if the latter one can be taken into account while deriving the effects of estimated coefficients as well as calculating BT values, we might be able to provide more meaningful and representative conclusions and suggestions v Dedicated to my family vi ACKNOWLEDGMENTS I have been blessed to have the support from many professionals throughout my graduate studies First, I wish to thank my advisor, Professor Alan Randall, for his invaluable guidance and advice His candid comments and sharp insights improved this dissertation a great deal It is his intellectual support that makes this dissertation possible My gratitude also goes to Professor Tim Haab and Professor Abdoul Sam for serving on my committee and continually providing econometric expertise to me This dissertation benefits greatly from their efforts and close review I also wish to thank Professor Eugene Jones for his constant encouragement, which means a lot to me and always cheers me up Finally, I would especially like to thank my parents and two sisters for always being there for me with their love and support vii VITA July 30, 1976 Born in Taipei, Taiwan 1999 B.S in Agricultural Economics, National Taiwan University 2001 M.S in Agricultural Economics, National Taiwan University 2005 M.A in Economics, The Ohio State University FIELDS OF STUDY Major Fields: Agricultural, Environmental, and Development Economics Studies in: Environmental Economics, Applied Econometrics, Bayesian Econometrics viii TABLE OF CONTENTS ABSTRACT ii ACKNOWLEDGMENTS vii VITA viii LIST OF TABLES x ESSAY 1.1 Introduction 1.2 Research Objective and Hypothesis 1.3 Data and Descriptive Statistics 1.4 Empirical Implementation 1.5 Conclusion 25 REFERENCES 31 ESSAY 36 2.1 Introduction 36 2.2 Model Description 38 2.3 Posterior Distribution, Model Space Selection, and Prior Refinement 40 2.4 Empirical Implementation 43 2.5 Hypothetical Policy Scenarios for Benefit Transfer Predictions 48 2.6 Conclusion 52 REFERENCES 56 ESSAY 58 3.1 Introduction 58 3.2 Bayesian Model Averaging (BMA) 60 3.3 Model Description 62 3.4 Empirical Analyses 65 3.5 Predictive Performance 71 3.6 Hypothetical Benefit Transfer Practice 73 3.7 Conclusion 74 REFERENCES 76 Appendix A 78 Appendix B 86 Appendix C 91 ix ... In the last few decades, this situation has changed since the public became aware of the contributions of wetlands through the provision of functions such as habitats for species, protection... prior information used in this study came from the combination of two or more variables in Brander et al.’s meta- equation and the derivation of their variances requires the summation of variances... due to the public becoming aware of the contributions of wetlands through the provision of various functions, which include habitats for species, protection against floods, water purification, amenities,