Christian Baker and Richard W Evans, Economics
Introduction
In this project we develop a big data computational technique to solve optimal policy problems and use this method to perform an analysis of optimal sales taxation. This approach provides for the consideration of a much richer sales tax schedule than current theoretical models allow. These same tools can also be used for a wide array of economic policy research.
Theoretical research can be more fruitfully used as part of forming a policy recommendation if the tax policy prescription is practically implementable and does not depend on restrictive assumptions. The US federal tax system resembles a piece-wise linear function, and the resulting discontinuities and nonconvexities make it diffcult for existing methods to fully model such a policy. Our approach is able to handle more realistic policy functions, as well as allow for many degrees of population heterogeneity.
Methodology
This project is particularly novel in its use of big data. Big data refers to any repository of data that is either large enough or complex enough that distributed and parallel I/O approaches must be used. In this approach, the idea is to build up a large database of how different types of people react to a set of tax policies, and then analyze the impact of those policies. Most research involving big data involves the analysis of large quantities of empirical data. This project is a different application of big data in that it is used as a solution method to theoretical models rather than as an empirical method. Using big data in this way can greatly expand the power of theoretical modeling.
We used data from the consumer expenditure survey to calibrate consumption shares and minimum consumption. We have used U.S. markup data to calibrate the elasticity of sub- stitution between goods. We also used quantile data from the Bureau of Labor Statistics to calibrate the household income distribution using a generalized beta prime distribution.
The software to perform these computations has been written in Python, a very exible and scalable language. It is parallelized using MPI (Message Passing Interface), the standard for distributed memory programming. This allows it to scale very efficiently across many cores at BYUs Fulton Supercomputing Lab. The software is also sufficiently generalized to enable it to handle a broad class of optimal policy models, making it highly reusable.
Results
We were able to calculate the optimal tax rates for a sales tax that differs across categories of goods to a at tax across all goods. Figure 1 displays the most important results. Here we compare the loss in government revenue by employing an optimal at sales tax as opposed to an optimal sales tax which differs across good categories. Based on these results, for a given level of social welfare the government may face revenue losses of up to 35% by implementing a at tax instead of a differentiated tax, but often the losses are much smaller. An optimally differentiated tax policy allows the policy maker to maximize total revenue collection given a particular total utility level.
Conclusion
The revenue loss of a at sales tax might not be that large if one considers the information and enforcement requirements and costs associated with an optimally differentiated sales tax system. We interpret this as evidence that a broad-based at tax system might be a reasonable option for fundamental tax reform. We also find that there is only a small loss in revenue from exempting a class of goods such as services in the United States.
This research paper has been submitted for publication in the Journal of Public Economics.