📊 statistics

Estimating causal effects of functional treatments with modified functional treatment policies

This paper proposes a novel framework for causal inference with functional treatments by introducing the Modified Functional Treatment Policy (MFTP) estimand, which uses a functional principal component analysis (FPCA) decomposition to overcome the challenges of averaging over infinite-dimensional trajectories, and provides a suite of identifiable estimators validated through simulations and real-world NHANES data.

Original authors: Ziren Jiang, Erjia Cui, Jared D. Huling

Published 2026-02-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ziren Jiang, Erjia Cui, Jared D. Huling

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor trying to figure out the best way to help a patient live longer. You notice that people who move around more during the day tend to stay healthy. But "moving around" isn't just one number—it’s a complex, wavy line of activity that changes every minute of the day.

This paper introduces a new way to study these "wavy lines" (which scientists call functional data) to understand cause and effect.

The Problem: The "One-Size-Fits-All" Trap

Before this paper, if scientists wanted to study the effect of activity, they usually used a method called ADRF.

Think of ADRF like a "Universal Remote Control." It tries to ask: "What would happen if we forced every single person on Earth to follow the exact same activity schedule—say, walking for 30 minutes at exactly 10:00 AM every day?"

This is a problem for two reasons:

It’s unrealistic: You can’t force a 90-year-old grandmother and a 20-year-old athlete to follow the exact same high-intensity workout. It’s physically impossible for some, which breaks the math.
It’s hard to visualize: Trying to describe a "universal" wavy line for everyone is like trying to draw a single "average" face that represents every human being—it ends up looking like a blurry, meaningless blob.

The Solution: The "Personalized Adjustment" (MFTP)

The authors propose a new idea called MFTP (Modified Functional Treatment Policy).

Instead of a Universal Remote, think of this as a "Smart Thermostat." A smart thermostat doesn't try to set every house in the world to exactly 72 degrees. Instead, it looks at how you currently live and says: "What if we just nudged your current habits slightly?"

If you are someone who stays very still at night, the MFTP asks: "What if we just reduced your nighttime movement by 20%?" If you are someone who sits too much during the day, it asks: "What if we boosted your daytime activity by 10%?"

Why this is better:

It’s realistic: It only suggests small, "plausible" changes to a person's actual life.
It’s meaningful: It answers questions doctors actually care about, like "Does reducing nighttime restlessness help people live longer?"

The Math Magic: The "Accordion" Trick

The biggest headache in the math is that a "wavy line" has infinite points. You can't calculate an average for something that never ends.

To solve this, the researchers use a trick called FPCA. Imagine a complex, beautiful piece of music. Instead of trying to analyze every single microscopic vibration of the air, you identify the main melody, the harmony, and the rhythm. Once you have those three main "components," you can recreate the whole song.

The researchers do this with the activity lines: they find the "main melodies" of the movement patterns, which makes the math manageable without losing the essence of the data.

The "Safety Net" (Double Robustness)

The researchers also built a "safety net" into their math called Double Robustness.

Imagine you are trying to predict if a cake will taste good. You use two methods: one looks at the ingredients (the outcome model) and one looks at the baker's reputation (the weighting model).

If you get the ingredients wrong, but the baker is a genius, you'll still get the right answer.
If the baker is terrible, but you perfectly measured the ingredients, you'll still get the right answer.

You only fail if both you and the baker are wrong. This makes their method very reliable even when the data is messy.

Real-World Proof: The NHANES Study

To prove it works, they applied this to real human data from the NHANES study (which uses wearable sensors). They found two powerful things:

Nighttime Restlessness: Reducing "disruptive" movement at night (like tossing and turning) is linked to lower mortality.
The Sedentary Slump: Increasing activity during "low-activity" periods during the day also helps people live longer.

In short: This paper gives scientists a way to move away from "one-size-fits-all" medicine and toward "nudge-based" medicine, using math that respects the complexity of human life.

Technical Summary: Estimating Causal Effects of Functional Treatments with Modified Functional Treatment Policies (MFTP)

1. Problem Statement

Traditional causal inference often deals with scalar treatments, but modern biomedical research increasingly utilizes functional data (e.g., continuous physiological trajectories or minute-level physical activity). While existing literature has attempted to estimate the Causal Average Dose-Response Functional (ADRF), this approach suffers from two major flaws:

Strong Positivity Violations: ADRF requires that a specific treatment trajectory $a(\cdot)$ be feasible for every individual in the population. In infinite-dimensional spaces, it is highly improbable that any single trajectory is applicable to all subjects, making the estimand scientifically irrelevant or unidentifiable.
Poor Interpretability: The ADRF is itself an infinite-dimensional functional, making it difficult to summarize or communicate to clinicians.

The authors seek to address these issues by shifting the focus from "what happens if everyone receives the same trajectory" to "what happens if everyone slightly modifies their own observed trajectory."

2. Methodology

The paper introduces the Modified Functional Treatment Policy (MFTP).

A. The Estimand:
Instead of a fixed $a(\cdot)$ , the MFTP defines a counterfactual scenario via a treatment modification policy $q(x, a(\cdot))$ . This function takes an individual's covariates $x$ and their observed trajectory $a(\cdot)$ and produces a modified trajectory $a_q(\cdot)$ . The target estimand is:
$\mu_q = E[Y(q(X, A(\cdot)))]$
This allows for "plausible" interventions, such as reducing nighttime activity by 50% for each individual, rather than forcing everyone into a single, potentially impossible, activity profile.

B. Overcoming Infinite Dimensionality:
A core mathematical challenge is that functional variables do not have a density function, making the standard iterative expectation used in scalar MTPs ill-defined. The authors propose a novel definition of the population average over functional variables using Functional Principal Component Analysis (FPCA). They define the average as the limit of expectations over the first $K$ principal components as $K \to \infty$ . They prove (via Lemma 2) that this definition is consistent with the standard expectation of the observed outcome.

C. Estimation Framework:
The authors derive three estimators based on the identified MFTP:

Outcome Regression (OR) Estimator: Uses a model $\hat{m}(x, a(\cdot))$ to predict outcomes under the modified treatment.
Inverse Probability Weighting (IPW) Estimator: Uses the density ratio of the principal component scores of the modified treatment versus the observed treatment.
Doubly Robust (AIPW) Augmented Estimator: Combines OR and IPW. This estimator is constructed using sample splitting to ensure asymptotic normality and avoid overfitting.

3. Key Contributions

New Causal Estimand: The MFTP provides a more flexible, interpretable, and mathematically feasible alternative to the ADRF.
Mathematical Foundation: Established the causal identifiability of MFTPs by bridging FPCA with causal inference theory.
Strong Double Robustness: Proved that the augmented estimator achieves root- $n$ consistency even if the nuisance models (outcome regression and density ratio) are estimated at slower rates, provided their combined convergence rate meets a specific threshold ( $\alpha_m + \alpha_e \geq 1/2$ ).
Dimension Reduction: Demonstrated that the infinite-dimensional problem can be effectively handled by weighting the finite-dimensional principal component scores.

4. Results

Simulation Studies: The authors validated that the AIPW estimator maintains root- $n$ convergence and high coverage rates across various scenarios, including model misspecification. They also showed that increasing the number of principal components ( $K$ ) generally improves MSE, particularly for the weighting estimator.
NHANES Application: Using accelerometer data, the framework was applied to estimate the causal effect of physical activity on all-cause mortality.
- Nighttime Activity: They showed that reducing "disruptive" nighttime activity (intensity $< 10$ units) causally reduces mortality.
- Low-Activity (Sedentary) Time: They demonstrated that increasing activity intensity during daytime "low-activity" periods (to reduce sedentary duration) also has a beneficial causal effect on mortality.

5. Significance

This work represents a significant advancement in Functional Causal Inference. By moving away from the rigid requirements of the ADRF, the authors provide a framework that is both mathematically rigorous and practically applicable to real-world longitudinal and high-frequency monitoring data. It allows researchers to ask "what-if" questions that are clinically relevant (e.g., "What if we reduce sedentary time?") while bypassing the impossibility of enforcing a single universal treatment trajectory on a diverse population.