Doubly-Robust Functional Average Treatment Effect Estimation

This paper introduces DR-FoS, a novel doubly-robust estimator for the Functional Average Treatment Effect (FATE) that ensures consistent estimation and valid simultaneous inference even when either the outcome or treatment assignment model is misspecified, demonstrating its effectiveness through simulations and a real-world application to the SHARE dataset.

The Gist

Imagine you are a doctor trying to figure out if a new diet helps people live longer. In the old days, you would just look at a single number: "Did they live to 80?" or "Did they live to 90?" That's a simple scalar outcome.

But in the modern world, we have smartwatches and medical sensors that track a person's health continuously. We don't just get one number; we get a whole curve of data over time—heart rate, sleep quality, and mobility levels changing every hour for years. This is called functional data.

The problem? The old statistical tools used to compare "Treatment A" vs. "Treatment B" break down when you try to apply them to these complex, wiggly curves. They get confused by the infinite amount of information and the messy real-world factors (like age, diet, or genetics) that influence the results.

Enter DR-FoS (Doubly-Robust Functional Average Treatment Effect). Think of this paper as the invention of a super-smart, double-checking navigator for analyzing these complex health curves.

Here is how it works, broken down with some everyday analogies:

1. The "Double-Check" Safety Net (Double Robustness)

Imagine you are trying to guess the average height of all the trees in a forest. You have two unreliable guides:

• Guide A knows the soil type but is bad at measuring trees.
• Guide B knows how to measure trees but is bad at understanding the soil.

In the past, if you used only Guide A, your answer would be wrong if the soil was weird. If you used only Guide B, your answer would be wrong if the trees were weird.

DR-FoS is like hiring both guides and building a system that says: "I will trust Guide A's soil data AND Guide B's measuring data. BUT, here is the magic: If Guide A turns out to be wrong, the system automatically switches to trusting Guide B. If Guide B is wrong, it trusts Guide A. As long as at least one of them is telling the truth, my final answer is correct."

This is called Double Robustness. It protects the researchers from making mistakes in their assumptions. If their model for "who gets the treatment" is slightly off, the "outcome model" saves them. If the "outcome model" is off, the "treatment model" saves them.

2. The "Wiggly Line" Problem (Functional Data)

Most statistics treat data like a single dot on a graph. But health data is a wiggly line (a function) that moves over time.

• The Old Way: Trying to flatten that wiggly line into a single average number loses all the nuance. It's like judging a whole movie by looking at just one frame.
• The DR-FoS Way: It treats the entire wiggly line as the object of study. It doesn't just ask, "Did the treatment help?" It asks, "How did the treatment change the shape of the health curve over time? Did it help in the morning but hurt at night? Did the effect grow stronger as the person got older?"

3. The "Confidence Blanket" (Simultaneous Confidence Bands)

When you look at a wiggly line, you want to know: "Is this curve really different from zero, or is it just random noise?"

• The Old Way: You check the line at 100 different points. If you check 100 points, you might accidentally find a "difference" just by luck (like flipping a coin 100 times and getting heads 10 times in a row).
• The DR-FoS Way: It creates a Confidence Blanket. Instead of checking points one by one, it wraps a fuzzy, transparent blanket around the entire curve. If the "zero line" (no effect) is outside this blanket for the whole duration, you can be 95% sure the treatment actually did something. It guarantees that the entire curve is statistically significant, not just random spikes.

4. The Real-World Test: The SHARE Study

The authors didn't just play with math; they tested this on real data from the SHARE study (a massive survey of European seniors).

• The Question: How do chronic conditions like high cholesterol or hypertension affect a person's quality of life and mobility over time?
• The Result: Using DR-FoS, they found that these conditions don't just cause a one-time drop in health. They create a slow, worsening decline in mobility and quality of life as people age. The "wiggly line" showed that the damage gets worse the longer you live with the condition.
• Why it matters: Because DR-FoS is so robust, they could trust these findings even though the data was messy and full of confounding variables (like different education levels or smoking habits).

Summary

DR-FoS is a new statistical tool that allows scientists to:

1. Analyze continuous curves of data (like health over time) instead of just single numbers.
2. Use a double-safety net so that if their assumptions about the data are slightly wrong, the answer is still correct.
3. Draw a confidence blanket around the whole curve to prove that the treatment effect is real and not just a fluke.

It's like upgrading from a black-and-white snapshot camera to a high-definition, 4K video camera with a built-in fact-checker, allowing us to see the true, dynamic story of how treatments affect our lives over time.

Technical Summary

Here is a detailed technical summary of the paper "Doubly-Robust Functional Average Treatment Effect Estimation" by Testa et al.

1. Problem Statement

The paper addresses a critical gap in causal inference: estimating the Functional Average Treatment Effect (FATE) when outcomes are not scalar variables but functional data (continuous functions observed over a domain, such as time or space).

• Context: Traditional causal inference methods (e.g., for Average Treatment Effects) are designed for scalar outcomes. However, modern applications (longitudinal health studies, epidemiology, neuroscience) generate data where the outcome $Y$ is a function $Y(t) \in C(\mathcal{T})$.
• Challenges:
  • Infinite Dimensionality: Functional data requires handling infinite-dimensional spaces.
  • Model Misspecification: Existing functional methods often rely on strong parametric assumptions (e.g., linear function-on-scalar regression) or lack robustness. If the outcome model or the treatment assignment (propensity score) model is misspecified, estimates become biased.
  • Inference: Constructing valid simultaneous confidence bands (SCBs) over the entire functional domain is difficult. Standard $L_2$ (Hilbert space) approaches measure average deviation but fail to control pointwise errors, which are essential for simultaneous inference.

2. Methodology: DR-FoS

The authors propose DR-FoS (Doubly-Robust Function-on-Scalar), a novel estimator for the FATE defined as $\beta = E[Y(1) - Y(0)]$.

Core Concept: Double Robustness

The estimator leverages the principle of double robustness, ensuring consistent estimation if either the outcome regression model ($\mu(a)(x) = E[Y(a)|X=x]$) or the propensity score model ($\pi(a)(x) = P(A=a|X=x)$) is correctly specified.

Estimation Procedure

1. Influence Function Construction: The estimator is derived using the influence function for the target parameter:
   $$\phi(D) = \gamma(1)(D) - \gamma(0)(D) - \beta$$
   where $\gamma(a)(D)$ is the case-corrected regression function:
   $$\gamma(a)(D) = \mu(a)(X) + \frac{\mathbb{I}(A=a)(Y(a) - \mu(a)(X))}{\pi(a)(X)}$$
2. One-Step Estimator: The DR-FoS estimator is the sample average of the estimated influence functions:
   $$\hat{\beta}_{DR-FoS} = \frac{1}{n} \sum_{i=1}^n \left( \hat{\gamma}(1)(D_i) - \hat{\gamma}(0)(D_i) \right)$$
3. Cross-Fitting: To avoid overfitting and relax Donsker-type assumptions, the authors employ cross-fitting. The data is split into $J$ folds; nuisance parameters ($\hat{\mu}, \hat{\pi}$) are trained on $J-1$ folds and evaluated on the held-out fold. The final estimate is the average of these fold-specific estimates.
4. Space of Functions: Crucially, the method operates in the Banach space of continuous functions $C(\mathcal{T})$ equipped with the sup-norm ($\|f\| = \sup_{t \in \mathcal{T}} |f(t)|$), rather than the standard $L_2$ Hilbert space. This choice is necessary to control the maximum deviation for simultaneous inference.

Inference (Simultaneous Confidence Bands)

The paper establishes that $\sqrt{n}(\hat{\beta}_{DR-FoS} - \beta)$ converges to a Gaussian Process (GP).

• Theoretical Basis: Under weak regularity conditions (including expected Hölder continuity of the outcome and nuisance functions), the estimator satisfies a functional Central Limit Theorem.
• Band Construction: Two methods are proposed for $(1-\alpha)$ simultaneous confidence bands:
  1. Critical Value Functions: Based on Liebl and Reimherr (2023), requiring estimation of a critical value function.
  2. Parametric Bootstrap: The authors adopt this approach (Pini and Vantini, 2017) for implementation, as it requires fewer assumptions on the covariance structure. It involves resampling from the estimated Gaussian process to determine quantiles.

3. Key Contributions

1. Novel Estimator: Introduction of DR-FoS, the first doubly-robust estimator specifically designed for functional outcomes in observational studies.
2. Theoretical Guarantees:
   • Proof of consistency under double robustness.
   • Proof of asymptotic normality for finite-dimensional projections.
   • Proof of convergence to a Gaussian Process in the sup-norm topology, enabling valid simultaneous inference.
3. Relaxed Assumptions: Unlike prior work (e.g., Liu et al., 2024) that requires strong parametric assumptions (linear outcomes, logistic propensity) and Hilbert space settings, DR-FoS works with flexible machine learning estimators (e.g., neural networks, random forests) and requires only mild smoothness (Hölder continuity) in a Banach space.
4. Intermediate Results: The paper derives new asymptotic results for the AIPW estimator with multivariate vector outcomes and the IPW estimator for functional outcomes, which are of independent interest.

4. Results

Simulation Study

The authors conducted extensive simulations comparing DR-FoS against Outcome Regression (OR) and Inverse Probability Weighting (IPW) estimators under various misspecification scenarios.

• Robustness: DR-FoS maintained high accuracy and low Mean Squared Error (MSE) even when one of the two models (propensity or outcome) was heavily corrupted by noise. In contrast, OR and IPW failed significantly when their respective models were misspecified.
• Coverage: The simultaneous confidence bands achieved nominal 95% coverage. The study also tested scenarios with discontinuities, showing that while bands widened to maintain validity, the method remained robust.
• Complex Models: Even when using complex non-linear models (like FunGCN) for the nuisance parameters, DR-FoS outperformed single-model approaches.

Real-World Application (SHARE Dataset)

The method was applied to the Survey of Health, Aging and Retirement in Europe (SHARE) to analyze the causal effect of chronic conditions (hypertension, high cholesterol) on functional quality-of-life indicators (Mobility Index, CASP scale) over 192 months.

• Findings: Both chronic conditions were found to have a statistically significant negative impact on quality of life and mobility.
• Temporal Dynamics: The adverse effects were observed to increase in magnitude over time, demonstrating the method's ability to uncover dynamic causal patterns that scalar summaries would miss.

5. Significance

• Bridging Fields: DR-FoS successfully bridges the gap between causal inference (specifically double robustness) and functional data analysis.
• Practical Utility: It provides researchers with a robust tool for analyzing complex, high-dimensional longitudinal data where model specification is uncertain—a common scenario in medicine and social sciences.
• Inference Rigor: By establishing convergence in the sup-norm, the paper solves the problem of constructing valid simultaneous confidence bands, allowing for rigorous inference over the entire functional domain rather than just pointwise or average effects.
• Future Directions: The framework sets the stage for extending causal inference to more complex structures, such as function-on-function treatments or non-i.i.d. data.

In summary, the paper presents a theoretically sound and practically robust framework for estimating causal effects in functional data settings, overcoming the limitations of traditional scalar methods and rigid parametric functional models.