Causal inference on distribution functions

Zhenhua Lin, Dehan Kong, Linbo WangJournal of the Royal Statistical Society Series B (Statistical Methodology) 2021

TL;DR

This paper introduces a new framework for causal inference when outcomes are entire distribution functions, moving beyond traditional analyses focused on scalar or vector outcomes. They develop robust statistical methods to estimate causal effects in this complex, non-linear "Wasserstein space" and demonstrate its utility by analyzing the impact of marriage on physical activity patterns.

What Problem Does This Paper Address?

Causal inference, the science of determining cause-and-effect relationships, has traditionally focused on outcomes that are simple numbers (like blood pressure) or vectors of numbers (like gene expression levels). However, many real-world datasets are more complex, with outcomes naturally represented as entire distributions. For example, instead of just looking at the average physical activity of a group, we might want to understand how an intervention affects the entire distribution of activity levels. Existing causal inference methods struggle with these distribution-valued outcomes because they often live in non-linear spaces where standard statistical tools break down. This paper addresses this gap by developing a principled way to perform causal inference on outcomes residing in the Wasserstein space of probability distributions.

Methods

The core of this paper is the development of a novel causal inference framework for outcomes in the Wasserstein space. Key methodological contributions include:

  • Wasserstein Space: The authors leverage the Wasserstein distance, a metric from optimal transport, to define a meaningful "distance" and "average" for probability distributions. This allows them to treat distributions as data points in a non-linear space.
  • Causal Effect Map: Instead of a single causal effect value, they define a "causal effect map" that describes how an intervention shifts the entire distribution of outcomes. This map can be interpreted as a difference in quantiles or as a causal transport map.
  • Doubly Robust Estimators: To estimate these causal effects from observational data, they develop "doubly robust" estimators. These estimators are designed to be reliable even if either the model for the outcome or the model for the treatment assignment is misspecified, a standard technique in causal inference adapted to their new setting.
  • Cross-Fitting: To further improve robustness and handle complex functional data, they employ cross-fitting techniques, a method that partitions data to reduce estimation bias.
  • Asymptotic Theory: Rigorous theoretical analysis is provided to establish the statistical properties (like consistency and asymptotic normality) of their proposed estimators.

Datasets

  • National Health and Nutrition Examination Survey (NHANES) 2005-2006: This is the primary dataset used for the real-world application. It contains physical activity data from wearable devices collected from participants.

Key Insights and Ideas

  • Distributions as Outcomes: The most significant insight is the formalization of causal inference when the outcome itself is a distribution. This opens up a new avenue for analyzing complex data where traditional summary statistics might hide crucial information.
  • Wasserstein Space as a Natural Home: The Wasserstein space, with its optimal transport foundation, provides a mathematically sound and intuitively appealing framework for comparing and averaging probability distributions, making it ideal for this type of causal analysis.
  • Causal Effect Map Captures Richer Information: Moving beyond a single average causal effect, the "causal effect map" provides a more nuanced understanding of how an intervention affects the entire spectrum of outcomes, not just the mean or median.
  • Double Robustness is Key: Adapting doubly robust estimation to this new setting is crucial for reliable inference, especially given the complexity of functional data and potential model misspecification.

Key Results

  • Framework for Distributional Causal Effects: The paper successfully defines and develops methods for estimating causal effects on distribution functions in the Wasserstein space.
  • Application to Marriage and Physical Activity: Using NHANES data, they quantify the causal effect of marriage on physical activity patterns. The results suggest that marriage is associated with an increase in physical activity levels across various quantiles. For instance, the DR estimator indicated that marriage improves the median physical intensity level by 19.1 cpm (95% CI: [12.1, 26.1]).
  • Doubly Robust Estimators Perform Well: Simulation studies show that their doubly robust estimators (both standard and cross-fitted) exhibit desirable properties, including low bias and good coverage probabilities for confidence bands, especially when compared to simpler outcome regression or inverse probability weighting methods.
  • Data-Adaptive Methods Show Promise: The paper also explores data-adaptive methods, which are important when the true functional forms of the outcome and propensity score models are unknown, demonstrating their consistency.

Limitations and Future Work

  • Focus on 1D Wasserstein Space: The current framework is primarily developed for distributions on a 1D interval. Extending this to higher-dimensional spaces or other non-linear manifolds presents significant challenges and is a key area for future research.
  • Unmeasured Confounding: Like most causal inference methods, this framework relies on the assumption of no unmeasured confounding. Extending these methods to handle unmeasured confounding, perhaps using instrumental variable approaches, is an important next step.
  • Efficiency Theory: While double robustness is established, extending semiparametric efficiency theory to these distribution-valued outcomes is an open area for future work.
  • Machine Learning Integration: While the cross-fitting approach can accommodate machine learning models, further theoretical work might be needed to fully leverage advanced ML techniques in this setting.

Summary updated

Abstract

Abstract Understanding causal relationships is one of the most important goals of modern science. So far, the causal inference literature has focused almost exclusively on outcomes coming from the Euclidean space Rp. However, it is increasingly common that complex datasets are best summarized as data points in nonlinear spaces. In this paper, we present a novel framework of causal effects for outcomes from the Wasserstein space of cumulative distribution functions, which in contrast to the Euclidean space, is nonlinear. We develop doubly robust estimators and associated asymptotic theory for these causal effects. As an illustration, we use our framework to quantify the causal effect of marriage on physical activity patterns using wearable device data collected through the National Health and Nutrition Examination Survey.
BibTeX
@article{lin2021causal,
  author = {Zhenhua Lin and Dehan Kong and Linbo Wang},
  title = {{Causal inference on distribution functions}},
  journal = {Journal of the Royal Statistical Society Series B (Statistical Methodology)},
  year = {2021},
}
Figure