Bridging binarization: causal inference with dichotomized continuous exposures

Summary

This methodological paper addresses a common practice in causal inference where researchers dichotomize continuous exposures to estimate average treatment effects, which are theoretically defined only for binary treatments. The authors establish that binarization is statistically valid by proving equivalence to specific modified treatment policies that impose exposure cut-offs while preserving relative self-selection (the ratio of probability densities). They introduce an alternative target parameter using the observed world as a benchmark, argue it addresses more relevant causal questions, and demonstrate the approach through simulation and application to California oil and gas well regulations and birth outcomes.

Regional applicability

This is a methodological contribution to causal inference statistics with no direct application to United Kingdom farming, soil health or food systems. The empirical application concerns environmental health policy in California. The statistical framework may be relevant to UK policy evaluation where continuous environmental or agricultural exposures are dichotomized for regulatory purposes, but direct transferability depends on whether UK researchers face analogous methodological questions.

Key measures

Average treatment effect (ATE), binarized ATE, modified treatment policies, relative self-selection ratios, birth outcomes in relation to proximity to oil and gas wells

Outcomes reported

The study demonstrates the statistical validity of binarization as a causal inference method by establishing equivalence between binarized average treatment effects and modified treatment policies with cut-offs and relative self-selection preservation. It introduces an alternative target parameter that uses the observed world as a benchmark and illustrates application to California oil and gas well exposure regulations and birth outcomes.

Topic tags