Research
Published Works
How the Reformulation of OxyContin Ignited the Heroin Epidemic
We attribute the recent quadrupling of heroin death rates to the August 2010 reformulation of an oft-abused prescription opioid, OxyContin. The new abuse-deterrent formulation led many consumers to substitute an inexpensive alternative, heroin. Using structural break techniques and variation in substitution risk, we find that opioid consumption stops rising in August 2010, heroin deaths begin climbing the following month, and growth in heroin deaths was greater in areas with greater prereformulation access to heroin and opioids. The reformulation did not generate a reduction in combined heroin and opioid mortality: each prevented opioid death was replaced with a heroin death.
William N. Evans, Ethan M. J. Lieber, Patrick Power; How the Reformulation of OxyContin Ignited the Heroin Epidemic. The Review of Economics and Statistics 2019; 101 (1): 1–15.
Working Papers
The Right to Counsel at Scale
We assess how the Right to Counsel affects housing stability. The Right to Counsel ensures that low-income tenants facing eviction have access to free legal representation. We exploit the recent adoption of this policy in some, but not all, zip codes in Connecticut. We show that legal representation improves court housing outcomes for those currently housed but adversely effects those currently unhoused. We use linear regression analysis for the intent-to-treat and IV estimates. We confirm our results using fine-tuned large language models and cluster regularized neural networks. We also provide insight about the type of tenants most likely to respond to the policy and how lawyers' strategies affect their clients housing outcomes.
Instrumental LLMs
In many applied microeconomic contexts, the underlying data is text (Health, Education, Housing). Causal inference in this setting has typically proceeded by hand selecting numerical representations of the text and estimating the corresponding conditional expectation function assuming that treatment or the instrument is locally randomly assigned. Recent developments in Natural Language Processing/AI though have an introduced alternative ways to produce causal estimates from text. In this paper we (1) clarify the general framework for using fine-tuned large language models for causal inference and (2) highlight their relative strengths in the setting of IV with preferential treatment.
Regularizing the Forward Pass
In certain applied microeconomic settings, it's typical to view one's dataset as the realization of a stratified cluster randomized control trial: treatment is assigned at the cluster level (such as zip code), and controls vary at both the individual and cluster level. Locally, this makes it more likely that observation will be from the same cluster which can increase the variance for estimators which don't account for the clustered nature of the data. We introduce a framework for partialling out nonparametric cluster effects in a way that generalizes least squares and is inherently compositional even under regularization. We provide a python library based on JAX: rfp