Bridging binarization: causal inference with dichotomized continuous exposures

Area of Science:

• Causal Inference
• Biostatistics
• Epidemiology

Background:

• Average treatment effect (ATE) is typically defined for binary exposures.
• Continuous exposures are often dichotomized, raising statistical concerns.
• Existing methods for continuous exposures lack clear interpretation.

Purpose of the Study:

• To validate binarization as a statistically sound method for continuous exposures.
• To clarify assumptions and interpretation of binarized causal effect estimators.
• To introduce a novel parameter for more relevant causal questions.

Main Methods:

• Equivalence proof between binarized ATE and modified treatment policies.
• Demonstration of assumed relative self-selection preservation.
• Introduction of a new benchmarked target parameter.

Main Results:

• Binarization is statistically equivalent to specific modified treatment policies.
• Clarified assumptions underlying binarization and interpretation of estimators.
• Proposed a new parameter addressing more relevant causal questions.

Conclusions:

• Binarization is a valid approach for causal inference with continuous exposures.
• Understanding and stating assumptions is crucial for proper interpretation.
• The new parameter offers a more relevant benchmark for causal analysis.