Machine learning for resource economics : a review of modern computational statistics with empirical applications in fisheries management

Abstract

Machine learning increasingly permeates our everyday lives, from artificial intelligence suggesting how we complete a text message to big data selecting creepily relevant ads to show us as we browse the web. While science and technology researchers have pushed these methods forward and private companies have embraced their power in significant changes to their processes, the field of economics has largely watched them go by. Despite the credibility revolution and increased focus on estimating causal effects, the econometric techniques in use today are largely identical to the ones used three decades ago. This thesis contributes to the growing field of literature at the intercept of machine learning and economics by exploring whether modern computational statistics methods can provide practical value to resource economists. I answer the following research question: Can integrating machine learning methods into econometric models improve upon traditional methods and add value in solving resource economics problems? To answer this question, I review the machine learning literature on causal analysis to find that machine learning methods solve certain types of problems in unique ways that traditional methods cannot. To test the benefit of these new methods in a resource economics setting, I apply machine learning to a fisheries problem based on the Costello, Gaines, & Lynham (2008a) article, Can Catch Shares Prevent Fisheries Collapse?, and analyse performance in a first-stage estimation task for propensity score matching. The results show machine learning can improve performance for prediction-based econometrics tasks under certain conditions. Shrinkage-based methods like Lasso regression proved to substantially improve model fit for datasets with moderate variance, while performing in-line with traditional methods when this condition didn’t hold. While more flexible methods like Random Forest performed extremely well fitting the data, they captured significant levels of noise by overfitting, challenging the external validity of their predictions. Machine learning identified and modelled valid selection bias that traditional methods could not – demonstrating value in solving practical resource economics problems. The impact of first-stage overfitting on the final causal model was unclear and presents an important area for further research, but the overall findings support the application of machine learning methods for robustness analysis on prediction tasks in resource economics.