AVERAGED PREDICTION MODELS (APM): IDENTIFYING CAUSAL EFFECTS IN CONTROLLED PRE-POST SETTINGS WITH APPLICATION TO GUN...

Area of Science:

• Causal inference
• Econometrics
• Policy analysis

Background:

• Controlled pre-post designs are common for policy impact evaluation.
• Disagreements persist regarding causal model specification in these designs, especially for gun policy effects on crime.
• Existing methods often lack consensus on the "correct" model.

Purpose of the Study:

• To propose a general identification framework unifying various causal models used in practice.
• To offer a data-driven procedure for selecting the most robust causal model.
• To address model specification debates in policy impact studies, particularly concerning gun policies and crime.

Main Methods:

• Developed a unified identification framework that generalizes Difference-in-Differences and other models.
• Employed models to predict untreated outcomes and adjusted treated group predictions using comparison group errors.
• Proposed a data-driven selection procedure based on model robustness to assumption violations, using pre-period prediction errors.

Main Results:

• The proposed framework nests common designs like Difference-in-Differences.
• The model selection procedure averages over candidate models, weighted by their robustness.
• The methodology is feasible using only pre-intervention data, offering a solution to model specification debates.

Conclusions:

• The proposed framework provides a robust method for causal inference in policy evaluation.
• Choosing models based on robustness, rather than a single "correct" specification, resolves common debates.
• The approach was applied to Missouri's 2007 gun policy change and an R package (apm) is available for implementation.