Patrick Power

Applications of Deep Learning to Microeconomics

Motivation

Since my undergraduate econometrics course, I've been interested in causality - how do the potential outcomes relate to linear models?

In graduate school, I read Richard Sutton's 'The Bitter Lesson', and became interested in methods that scale/improve with greater compute

From reading Evicted my first year of graduate school, joining Winnie Van Dijik's Homelessness Reading group, and volunteering at Project Weber Renew, I developed an interest in the Economics of Housing & Homelessness (which has a lot of interesting Causal Inference questions)

I've had the opportunity to teach all of this in a seminar course which has been a lot of fun: EC970

Overview

Instrumental LLMs

The Right to Counsel at Scale

Regularizing the Forward Pass

Overview

The vast majority of applied microeconometric work involves averaging local predictions. In the following chapters, we extend this approach to (a) text data and (b) stratified clustered randomized control trials. We apply these methods to better understand the impact of providing free legal representation to tenants who are facing eviction.

OUtline of the Talk

Slides

Instrumental LLMs

Clarify how we think about causality with LLMs and highlight their usefulness in the setting of IV with preferential treatment

Slides

Regularizing the Forward Pass

Highlight the statistical challenge of estimating causal effects with stratified clustered random assignment and show how neural networks trained via regularized bilevel gradient descent might be well suited in this context.

Slides

The Right to Counsel at Scale

Using the above methods, estimate the impacts of providing tenants facing eviction with free legal representation. Primarily focused on housing instability

Python Libraries

trics

rfp