Essays in Methodological Improvements of Structural Economics

Abstract

My dissertation involves using computational tools to improve structural modeling in economics. Structural models concern the building of models whose parameters correspond to economic primitives. For example, the depreciation rate may be one parameter.

My first paper uses a machine learning technique to perform likelihood-free Bayesian inference on structural macroeconomic models. This method allows for both faster and more accurate Bayesian posteriors and the ability to estimate complex models like heterogenous agent models that often don't have closed-form likelihood functions. Compared to baseline techniques like Markov Chain Monte Carlo, my method is faster/more accurate, can handle more underlying macroeconomic models, and is easier to code up.

My second paper uses the same method to estimate structural models on graphs/networks. The current literature doesn't have a workhorse algorithm that can estimate structural models on graphs. Often it is required that the particular structural model has to be designed to have a closed-form likelihood function which is often unknown. Or, one can use "Reduced form techniques like exponential random graph models where the parameters don't correspond to economic or other primitives but only exist to determine the "shape" of the estimation function. My approach allows one to estimate structural models on networks as long as one can simulate networks from the structural model. This opens up the ability to estimate many more structural network models.

My final paper argues that better estimates can be made if one estimates macroeconomic models across countries rather than through time. The conventional way to estimate macroeconomic models is to pick one country like the US and ignore the cross-section of countries. I argue this is less efficient than using many countries and uses data less far back in time.

In conclusion, my work shows how economists and other social scientists can use tools to improve structural modeling accuracy and the number of models that can be estimated.