The functional average treatment effect

Shane Sparkes, Erika Garcia, Lu ZhangJournal of Causal Inference 2024 · DOI

TL;DR

This paper introduces the "functional average treatment effect" (FATE) as a robust new estimand for causal inference, particularly useful when traditional assumptions like mean exchangeability are hard to meet. They develop novel estimation and inference methods, including the "Hoeffding bootstrap" and a special class of "U random variables," which simplify causal analysis and can even lend causal interpretations to standard statistical procedures.

What Problem Does This Paper Address?

In causal inference, a core challenge is estimating the effect of a treatment or intervention when we can only observe one outcome per individual (the "fundamental problem of causal inference"). Traditional methods, like those based on the Neyman-Rubin model, often rely on strong assumptions such as "mean exchangeability" (where the average outcome under treatment is the same as the average outcome under control, conditional on covariates) and "positivity" (where all individuals have a non-zero probability of receiving any treatment). These assumptions can be difficult to satisfy in observational studies, especially when dealing with high-dimensional confounders or when the distribution of outcomes differs significantly between treated and untreated groups. This paper aims to provide a more robust framework for causal inference that bypasses some of these stringent requirements.

Methods

The paper's key contributions lie in several novel methodological developments:

  • Functional Average Treatment Effect (FATE): They propose using the "functional average" of an outcome variable as a new estimand. This is essentially a uniform average over the support of the outcome, making it less sensitive to specific probability distributions and thus more robust to confounding.
  • Hoeffding Bootstrap: To address the challenge of statistical inference for FATE, they introduce a new bootstrap method. This method is designed to be conservative and robust, working even under complex and unknown dependency structures, unlike traditional bootstrap approaches.
  • U Random Variables: They define and explore a class of "U random variables" where the expected value equals the functional average. This class simplifies estimation and allows standard statistical procedures like linear regression and ANOVA to be interpreted causally under milder conditions.
  • Estimators: They examine existing estimators like the sample mid-range and propose new plug-in estimators for the functional average.

Datasets

The paper uses the National Health and Nutrition Examination Survey Data I Epidemiologic Follow-up Study (NHEFS) to investigate the causal effect of exercise on blood pressure in adult smokers.

Key Insights and Ideas

  • Robustness to Confounding: The functional average is inherently more robust to confounding than traditional mean-based estimands because it focuses on the range of possible outcomes rather than their specific probabilities. This means that even if the probability distributions differ significantly between groups, the functional average can still be identified.
  • Causal Interpretation for Standard Methods: The introduction of "U random variables" is a significant breakthrough. It shows that familiar tools like linear regression can be used to estimate causal effects under surprisingly mild conditions, effectively bridging the gap between associational and causal inference.
  • The "No Free Lunch" Principle Still Applies: While FATE offers robustness, the paper acknowledges that efficient estimation and inference remain challenging, necessitating novel approaches like the Hoeffding bootstrap.
  • Beyond Mean Exchangeability: The paper demonstrates that causal inference is possible even when mean exchangeability is violated, as long as the support of the counterfactual distribution is preserved. This opens up new avenues for causal analysis in settings where traditional assumptions are untenable.

Key Results

  • Identification of FATE: The paper establishes conditions under which the functional average treatment effect is identifiable, even with unmeasured confounders, provided the support of the counterfactual distribution is preserved.
  • Hoeffding Bootstrap Performance: Simulations show that the Hoeffding bootstrap provides conservative and reasonably accurate confidence intervals for functional averages, outperforming traditional methods in certain challenging scenarios, especially those involving dependent data or non-smooth statistics.
  • U Variables Simplify Causal Inference: The U random variable framework allows for causal interpretations of linear regression and ANOVA, requiring only the preservation of support and U status in the error distributions, which are often met in practice.
  • NHEFS Application: The analysis of the NHEFS data suggests that exercise has a plausible, though not statistically significant, effect on blood pressure in adult smokers, with estimates of the functional average change being close to zero.

Limitations and Future Work

  • Computational Intractability: While the functional average is robust, its direct estimation and the inference for it can be computationally challenging, especially for continuous outcomes or complex dependency structures.
  • Bootstrap Conservatism: The Hoeffding bootstrap, while robust, is noted to be conservative, meaning it might produce wider confidence intervals than necessary in some cases.
  • Theoretical Work on Extremes: The authors suggest that more theoretical work is needed on the relationship between the extremes of the support of empirical distributions and population distributions to further refine bootstrap methods.
  • Generalizability of U Variables: While U random variables simplify analysis, their applicability and the conditions under which they hold need further exploration.
  • Data Limitations: The NHEFS application was hampered by a relatively light right tail in the conditional SBP distributions and a small sample size, limiting the power to detect effects. The authors also acknowledge that the consistency assumption might be violated due to the subjective nature of exercise categories.

Summary updated

Abstract

Abstract This article establishes the functional average as an important estimand for causal inference. The significance of the estimand lies in its robustness against traditional issues of confounding. We prove that this robustness holds even when the probability distribution of the outcome, conditional on treatment or some other vector of adjusting variables, differs almost arbitrarily from its counterfactual analogue. This article also examines possible estimators of the functional average, including the sample mid-range, and proposes a new type of bootstrap for robust statistical inference: the Hoeffding bootstrap. After this, the article explores a new class of variables, the <m:math xmlns:m="http://www.w3.org/1998/Math/MathML"> <m:mi class="MJX-tex-caligraphic" mathvariant="script">U</m:mi> </m:math> {\mathcal{U}} class, that simplifies the estimation of functional averages. This class of variables is also used to establish mean exchangeability in some cases and to provide the results of elementary statistical procedures, such as linear regression and the analysis of variance, with causal interpretations. Simulation evidence is provided. The methods of this article are also applied to a National Health and Nutrition Survey data set to investigate the causal effect of exercise on the blood pressure of adult smokers.
BibTeX
@article{sparkes2024functional,
  author = {Shane Sparkes and Erika Garcia and Lu Zhang},
  title = {{The functional average treatment effect}},
  journal = {Journal of Causal Inference},
  year = {2024},
  doi = {10.1515/jci-2023-0076},
}
Figure