SDS Seminar Series – Dr. Laura Hatfield

Mar

2024

2:00 pm – 3:00 pm

Virtual (view details)

Featured Speaker(s): Laura Hatfield

Cost: Free

Predict, Correct, Select: A New General Identification Strategy for Controlled Pre-Post Designs

Description

The Spring 2024 SDS Seminar Series begins on March 1st from 2:00 p.m. to 3:00 p.m. with Dr. Laura Hatfield (Health Care Policy (Biostatistics), Department of Health Care Policy, Harvard Medical School). This event is in-person.

Title: Predict, Correct, Select: A New General Identification Strategy for Controlled Pre-Post Designs

Abstract: Whether policies that expand access to firearms reduce or increase crime is a question of fierce debate. To study it, researchers may observe changes in crime among people exposed to a change in firearm law and compare these changes to those of an unexposed comparison group. With some counterfactual assumptions, this enables causal conclusions about the effects of gun laws. However, these empirical investigations have reached widely varying conclusions depending on the specifics of their methods. The policy debate is therefore stymied by disagreements over the "correct" causal model. In this talk, I describe a novel identification framework for controlled pre-post designs. We use models to predict untreated outcomes and correct the treated group's predictions using the comparison group's observed prediction errors. The crucial identifying assumption is that the treated and comparison groups would have equal prediction errors (in expectation) under no treatment. To select the best prediction model, we propose a data-driven procedure that is motivated by design sensitivity. We choose the prediction model that is most robust to violations of the identification assumption by observing the differential average prediction errors in the pre-period. Our approach offers a way out of the debate over the "correct" model by choosing the most robust model instead. It also has the desirable property of being feasible in the "locked box"' of pre-period data only and accommodates the range of prediction models that applied researchers employ. We use our procedure to select from a set of candidate models and estimate the effect on homicide of Missouri's 2007 repeal of its permit-to-purchase law.