« Back to Results

Applications of Machine Learning in Microeconomics for Public Policy

Paper Session

Friday, Jan. 4, 2019 10:15 AM - 12:15 PM

Atlanta Marriott Marquis, International 7
Hosted By: American Economic Association
  • Chair: Matthew Lang, University of California-Riverside

Cell Phone Coverage and Traffic Accidents: New Evidence Using Cell Phone Towers

Bree Lang
,
University of California-Riverside
Matthew Lang
,
University of California-Riverside
Jonathan Hersh
,
Chapman University

Abstract

We examine the relationship between the growth of 3G cellular phone coverage and traffic accidents in California between 2001 and 2011. Cellular coverage maps are only available from the FCC in 2015 and 2016, not in the mid-2000s when 3G coverage was introduced to the public. We link cellular coverage along a highway in 2016 to the location of antenna towers and then apply machine learning techniques to predict coverage between 2001 and 2011. Controlling for traffic volume, empirical results indicate an increase in the car accident rate when 3G cell phone coverage increases. The largest effects are found for accidents that are caused by drivers younger than than 30. There are no effects for accidents caused by drivers older than 50. In contrast to previous research examining the effects of changes in cell phone laws, we provide evidence that traffic accidents increase when 3G coverage increases.

Measuring the market impact of deep learning based valuations in common value auctions with endogenous entry (Jason Ansel, Matthew Harding, Michael P. Leung, Jessie Li)

Jason Ansel
,
Godaddy.com
Matthew Harding
,
University of California-Irvine
Jessie Li
,
University of California-Santa Cruz

Abstract

In most transactions humans play the essential role of reviewing the available information to determine the value of a good. Recent developments in machine learning allow however for the possibility of replacing humans with algorithms in many such tasks since algorithms are arguably both faster and more accurate at reviewing complex information. This paper evaluates the market impact of using valuations provided by a deep learning algorithm in a large scale randomized controlled trial conducted by GoDaddy, a large domain name registrar that facilitates the sale and resale of domain names. Using both reduced form evidence and a structural model of auctions with pure common values and endogenous entry, this paper quantifies the extent to which the use of machine learning in this context changes market behavior on both the extensive and intensive margins using a variety of measures such as completion of sale, bidding behavior, and bidding amounts. We find that adding the deep learning valuation increases the probability of sale, the sale price, and the number of auction participants.

JEL classification: C55; C57; D44
Keywords: auctions; randomized experiment; common value; endogenous entry

Imperfect Markets Versus Imperfect Regulation in United States Electricity Generation

Steve Cicala
,
University of Chicago

Abstract

This paper evaluates changes in electricity generation costs caused by the introduction of market mechanisms to allocate production in the United States. I use the staggered transition to markets from 1999-2012 to estimate the causal impact of liberalization using a machine learning-augmented differences-in-differences design on a comprehensive hourly panel of electricity demand and unit-level costs, capacities, and generation. I find that markets reduce production costs by reallocating output: Gains from trade across service areas increase by 20% based on a 10% increase in traded electricity, and costs from using uneconomical units fall 20% from a 10% reduction in their operation.

What Determines Labor Market Re-attachment? A Machine Learning Approach

Jonathan Hersh
,
Chapman University
Benedikt Herz
,
European Commission

Abstract

Job displacement continues to be one of the major concerns in the United States, especially because of the increasing perceived threat of automation. The effects of displacement are still not entirely understood by economists. In particular, a large body of literature in labor economics documents that the experience of displaced workers is very heterogeneous: while some workers remain unemployed for sustained periods or accept a job where they earn substantially less, others quickly accept a new job and often increase earnings relative to their pre-displacement job. Many reasons for this heterogeneity have been discussed. For example, some researchers claim that the willingness to move for a job is essential, others argue that the characteristics of the local labor market are a decisive factor while again others see an important determinant in occupation or industry changes. In this paper, we shed light on robust determinants of wage-losses and unemployment duration of displaced workers using machine learning (ML) methods. We use a survey on displaced workers in the United States that contains a rich set of demographic variables and other covariates. We use LASSO and random forest models to determine which demographic and local labor market covariates predict re-entry into the labor market. We then discuss to what extent these robust estimates are in line with previous literature. Our findings have important implications for US labor market policy.

The Impact of Local Tax Complexity on Firm Behavior

Matthew Harding
,
University of California-Irvine
Michael F. Lovenheim
,
Cornell University
Elizabeth Maloney
,
University of California-Irvine

Abstract

This paper examines a novel concept in the economics of taxation: how does the complexity of local tax policies impact firm behavior, prices, and therefore consumption of grocery items? Local taxes are extraordinarily complex: there often are many exemptions, and taxes can vary by the ingredients and package size of a good. Complying with such complex taxes is difficult, and the incidence of these compliance costs may fall on the consumer. That is, high compliance costs may lead to higher-priced goods. Any higher prices could be targeted to taxed goods, but there is no particular reason to suspect this is the case. If all goods are made more expensive when local sales taxes are more complex, these taxes may have unintended effects on consumer purchasing patterns and nutrition. We use machine learning algorithms to develop complexity measures for local taxes at the UPC level. These measures are then linked with household level grocery purchases in a large panel data set of over 100 million food purchase transactions. We will exploit variation over time and across space in the complexity of the sales tax code and adopt a machine learning strategy to estimate the effect of tax complexity on food prices (both overall and by category) as well as on purchasing patterns and nutritional bundles. We also account for the heterogeneous impact on different types of consumers, products and retailers.
Discussant(s)
Bree Lang
,
University of California-Riverside
Benedikt Herz
,
European Commission
Jonathan Hersh
,
Chapman University
Matthew Harding
,
University of California-Irvine
Steve Cicala
,
University of Chicago-Harris
JEL Classifications
  • C1 - Econometric and Statistical Methods and Methodology: General