Essays in Information Technology and Productivity

Abstract

This dissertation studies the relationship between information technology and productivity in three domains. Chapter 1 examines a mobile software application that allows recipients of the Supplemental Nutrition Assistance Program (SNAP) to check their benefit balance. Recipients spend a large majority of benefits before halfway through a benefit deposit cycle. Using an event study, I show that the introduction of the application on average has small but significant impacts - around 5% - on the ability of recipients to extend the time frame over which they spend benefits within a cycle. These effects are higher for recipients who are new to SNAP, who are highest in the distribution of SNAP benefits, and who have the largest tendency pre-adoption to spend down quickly. The results are consistent with the impact of salience on consumer choice and offer evidence that such software tools may be a cost effective means to support policy goals. Chapter 2 examines the impact of machine learning on public sector productivity in practice. Partnering with Yelp and the City of Boston, we run an experiment to compare an inspector-curated list of restaurants to inspect (i.e. business-as-usual) to a pair of algorithm-created lists based on empirical predictions of which restaurants are most likely to have health code violations. Our goal is to understand the gains from - and barriers to - implementing predictive algorithms to improve the city’s ongoing inspection operations. We present four main findings. First, even simple algorithms greatly outperform business-as-usual; the city can identify 50% more violations using the same number of inspections. Second, one key barrier to implementing an algorithm in managerial contexts is compliance. In our sample, inspectors were only half as likely to comply with a directive to inspect a restaurant based on the algorithm relative to restaurants based on their own judgment. Third, beyond efficiency differences, the algorithm also has equity implications. For example, relative to the inspector-created list, algorithms were more likely to target ethnic restaurants and major chains. Fourth, based on these results, Boston has proceeded with implementing a modified version of the algorithm into their ongoing inspection process. Chapter 3 studies the theoretical impact of machine learning applied to the selection of public sector workers. Economists have become increasingly interested in studying the nature of production functions in social policy applications with the goal of improving productivity. Traditionally models have assumed workers are homogenous inputs. However, in practice, substantial variability in productivity means the marginal productivity of labor depends substantially on which new workers are hired--which requires not an estimate of a causal effect, but rather a prediction. We demonstrate that there can be large social welfare gains from using machine learning tools to predict worker productivity using data from two important applications - police hiring and teacher tenure decisions.