Fair Inference on Outcomes.

Area of Science:

• Statistical Inference
• Causal Inference
• Machine Learning

Background:

• Fairness is a critical concern in statistical inference, particularly when sensitive covariates may lead to discrimination.
• Existing methods often struggle to formally define and address discrimination in outcome prediction and treatment effect estimation.

Purpose of the Study:

• To propose a formal definition of discrimination in statistical inference based on causal pathways.
• To develop a method for learning fair outcome models using constrained optimization.
• To address complications in classical statistical inference arising from this fairness perspective.

Main Methods:

• Formalizing discrimination as the effect of sensitive covariates on outcomes via specific causal pathways.
• Employing constrained optimization techniques to learn fair statistical models.
• Leveraging recent advancements in causal and semi-parametric inference to handle statistical complexities.

Main Results:

• A novel framework for defining and detecting discrimination in statistical inference problems.
• A practical approach to learning fair outcome models that mitigate discriminatory effects.
• Identification of challenges in classical statistical inference and proposed workarounds.

Conclusions:

• The causal pathway approach provides a robust definition of discrimination in statistical inference.
• Constrained optimization offers a viable method for achieving fairness in outcome models.
• This work advances fair machine learning and statistical inference by integrating causal reasoning.