Interpreting Experiments with Multiple Outcomes
(with Tom Cunningham)
Current version: Nov 2019
Abstract: When we observe the effect of an experiment on multiple outcomes, the interpretation will be sensitive to how those outcomes covary across units. We derive a number of results in a Gaussian model of multi-outcome experimentation - (1) the observed effect on one outcome will be a negative signal about the true effect on another outcome, under conditions that are likely to hold for many experiments; (2) in some cases, the inferred treatment-effect on an outcome can be decreasing in its own observed treatment-effect; (3) naive approaches to metric "surrogacy", when one metric is used to predict another, will be biased in the direction of the unit-level covariance, and naive causal estimates will suffer attenuation bias; (4) composite metrics, i.e. weighted averages of multiple outcomes, will often be shrunk by more than their components. Finally we show how to combine multivariate shrinkage with network effects and dynamic effects to yield a single matrix which maps outcomes of an experiment into the best estimate of the long-run aggregate impact of a policy.
Are High-Interest Loans Predatory? Theory and Evidence from Payday Lending
(with Hunt Allcott, Dmitry Taubinsky, and Jonathan Zinman)
Current version: March 2020
Abstract: It is often argued that consumer lending regulations can increase welfare, because high-interest loans cause "debt traps" where people borrow more than they expect or would like to in the long run. We test this using an experiment with a large payday lender. While the most inexperienced quartile of borrowers underestimate their likelihood of future borrowing, the more experienced three quartiles predict correctly on average. This finding contrasts sharply with priors we elicited from 103 payday lending and behavioral economics experts, who believed that the average borrower would be highly overoptimistic about getting out of debt. Borrowers are willing to pay a significant premium for an experimental incentive to avoid future borrowing, implying that they believe they are present focused. We combine these data with a novel sufficient statistic-based identification strategy to estimate a model of partially naive present focus. Using our estimated parameters, we carry out a behavioral welfare evaluation of common payday lending regulations. In our model, payday loan bans unambiguously reduce welfare, and limits on repeat borrowing generate at best small welfare gains.
Food Labelling: Effects on Demand and Supply of Nutritional Content
(with Nano Barahona, Cristobal Otero, and Sebastian Otero)
Currency Depreciations and Savings Decisions: Evidence from Household Savings in Armenia
(with Diego Jimenez and Aleksander Shirkhanyan)