Abstract:Introduction: The environmental health effects have been studied for decades, where model-based methods have been standard for analyzing exposure-outcome associations. Methods: We propose a complementary approach to examine causal effects of environmental exposures on health outcomes based on insights from classical experimental design that rely on modern computing. We illustrate our approach using two examples, the effect of extreme hot temperatures on cardiovascular mortality, and the effect of parental smoking on children's lung function. Our approach proposes an analysis strategy with four transparent, distinct, and ordered stages: 1) The precise formulation of the causal question (and related assumptions) in terms of a hypothetical randomized experiment using a potential outcomes framework. 2) A design stage that attempts to reconstruct the ideal conditions for a randomized experiment (that is, with unconfounded exposure assignment). 3) A statistical analysis comparing the outcomes of interest in the exposed versus unexposed units of the hypothetical randomized experiment. 4) A conclusion about evidence for or against the hypothesized causal question. Results: When investigating the causal effect of extreme temperatures on cardiovascular mortality in Los Angeles, we find similar effect estimates using the standard regression approach and our alternative approach (e.g., the estimated difference in daily mortality is 0.05, comparing days with Tmax>30C to milder days). However, we find different conclusions when we examine the causal effect of parental smoking in a classical dataset on children's FEV-1 (the estimated causal effect equals to -0.34 [95%CI:-0.59; -0.09] using our approach (here, matched-sampling) vs. -0.09 [95%CI:-0.20; 0.03] using standard regression). Conclusions: Because of its conceptual links to (hypothetical) interventions, our approach can suggest the effects of policies for reducing environmental pollutants and thereby preventing diseases.

Rejoinder: Causal inference with misspecified exposure mappings: separating definitions and assumptions

Discussion of 'Causal inference with misspecified exposure mappings: separating definitions and assumptions'

Exposure effects are not automatically useful for policymaking

Causal Interpretation of Estimands Defined by Exposure Mappings

Causal definitions versus casual estimation: Reply to Valente et al. (2022).

Causal inference under over-simplified longitudinal causal models

Conflating marginal and conditional treatment effects: Comments on 'Assessing the performance of population adjustment methods for anchored indirect comparisons: A simulation study'

A Review of Causal Inference Methods for Estimating the Effects of Exposure Change when Incident Exposure Is Unobservable

Causal inference with a quantitative exposure

Comparing Causal Inference Methods for Point Exposures with Missing Confounders: A Simulation Study

Formulating causal questions and principled statistical answers

The Difference Between Causal Analysis and Predictive Models: Response to “Comment on Young and Holsteen (2017)”

Transporting established insights from classical experimental design, implemented using methods that rely on modern computing to address causal questions in environmental epidemiology

Causal Inference When Counterfactuals Depend on the Proportion of All Subjects Exposed

Rejoinder on Causal Inference Through Potential Outcomes and Principal Stratification: Application to Studies with ``Censoring'' Due to Death by D. B. Rubin

Recanting witness and natural direct effects: Violations of assumptions or definitions?

Estimating a Causal Exposure Response Function with a Continuous Error-Prone Exposure: A Study of Fine Particulate Matter and All-Cause Mortality

Estimators for longitudinal latent exposure models: examining measurement model assumptions

Causal Inference with Noncompliance and Unknown Interference

The built-in selection bias of hazard ratios formalized using structural causal models

A Unification of Exchangeability and Continuous Exposure and Confounder Measurement Errors: Probabilistic Exchangeability