An Automated Approach to Causal Inference in Discrete Settings.

Area of Science:

• Causal inference
• Econometrics
• Machine Learning

Background:

• Traditional causal inference often requires strong, untestable assumptions for point identification.
• Partial identification, offering bounds on causal effects, is theoretically sound but practically challenging to implement.
• Existing methods struggle with complex, real-world data issues like confounding, selection, and measurement error.

Purpose of the Study:

• To develop a general, automated numerical approach for deriving sharp bounds on causal effects in discrete settings.
• To overcome the practical difficulties of applying partial identification in idiosyncratic research scenarios.
• To provide a user-friendly tool for causal inference with incomplete or imperfect data.

Main Methods:

• Causal questions with discrete data are reformulated as polynomial programming problems.
• An algorithm employing dual relaxation and spatial branch-and-bound techniques is used to automatically derive bounds.
• The approach handles incomplete or mismeasured data by searching over admissible data-generating processes.

Main Results:

• The method automatically computes sharp bounds for causal effects, identifying point-identified solutions when possible.
• It provides continually refined, non-sharp bounds that guarantee coverage even if computation is interrupted.
• Simulations demonstrate robustness against confounding, selection, measurement error, noncompliance, and nonresponse.

Conclusions:

• The automated numerical approach offers a principled and practical solution for causal inference under weak assumptions.
• The autobounds Python package facilitates the application of these advanced bounding techniques.
• This method enhances the reliability of causal effect estimation in diverse applied research settings.