📊 statistics

Efficient Longitudinal Function-on-Function Regression

This paper proposes a computationally efficient, three-step marginal approach for longitudinal function-on-function regression that enables accurate estimation and valid inference on high-dimensional wearable data, as demonstrated in a physical activity intervention trial and implemented in an R package.

Original authors: Leif Verace, Siobhan McMahon, Erjia Cui

Published 2026-05-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Leif Verace, Siobhan McMahon, Erjia Cui

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 trying to understand how a group of people's daily routines change over time, but instead of just asking "Did they exercise today?", you have a super-detailed video recording of their every move, minute-by-minute, for months. That is the kind of data scientists are dealing with in this paper: wearable device data (like Fitbits) that tracks physical activity not just as a single number, but as a continuous, flowing curve throughout the day.

The researchers wanted to answer a specific question: Do different types of intervention strategies (like doing exercises alone vs. in a group) change when during the day older adults are most active?

Here is the breakdown of their work using simple analogies:

The Problem: The "Too Big to Fit" Puzzle

Imagine trying to solve a giant jigsaw puzzle where every piece is a tiny, moving video frame.

The Data: They had nearly 300 people, each visited 4 times over a year. At every visit, they had minute-by-minute activity data.
The Challenge: Traditional statistical methods are like trying to solve this puzzle by looking at every single piece at once, in 3D, while the pieces are moving. It's so computationally heavy that the computer gets stuck. In fact, for a dataset this size, the old methods would take days to finish, or sometimes fail to finish at all.

The Solution: The "Three-Step Sandwich"

The authors, Leif Verace, Siobhan McMahon, and Erjia Cui, invented a new, faster way to solve this puzzle called ELFFR (Efficient Longitudinal Function-on-Function Regression). Instead of trying to solve the whole giant puzzle at once, they broke it down into three manageable steps:

Step 1: The "Snapshot" Approach
Instead of looking at the whole day at once, they looked at the data one minute at a time. They asked, "What is happening at 8:00 AM?" and solved that tiny puzzle. Then they asked, "What about 8:01 AM?" and solved that one.
- Analogy: Imagine watching a movie by freezing it on a single frame, analyzing that frame, then moving to the next. Because each frame is small, the computer can solve it instantly.
Step 2: The "Smoothing" Glue
Once they solved the puzzle for every single minute, they had a million tiny, slightly jagged answers. They used a mathematical "smoother" to connect the dots, turning those jagged lines into a beautiful, flowing curve that makes sense across the whole day.
- Analogy: It's like taking a rough, pixelated sketch and running a smoothing brush over it to reveal the clear, continuous picture underneath.
Step 3: The "Confidence Check"
They needed to make sure their answers weren't just lucky guesses. They used two methods to draw "confidence bands" (like a safety net) around their results. One method is a quick math formula (for data that looks like a bell curve), and the other is a "resampling" trick where they pretend to run the study over and over again on a computer to see if the results hold up.

The Result: Speed and Clarity

Speed: The old method was like a snail; the new method is a race car. For a large dataset, the old method took over 13 hours, while their new method finished in about 35 minutes.
Accuracy: Despite being faster, their results were just as accurate as the slow, heavy methods.

What They Found in the Real World (The RS3 Study)

They applied this new method to a real study called Ready Steady 3.0, which tested ways to get older adults to move more. They compared "Intrapersonal" strategies (doing it alone) vs. "Interpersonal" strategies (doing it in a group).

The Discovery: The old way of looking at the data (just averaging the whole day) missed the nuance. The new method revealed a specific "magic window."
The Finding: People in the group (interpersonal) intervention became significantly more active specifically between 9:00 AM and 2:00 PM.
The Contrast: The "alone" (intrapersonal) groups did not show this same boost.
The Takeaway: It turns out that doing activities with others doesn't just make people move more in general; it specifically supercharges their activity during the morning and early afternoon hours.

Why This Matters

This paper isn't just about math; it's about giving researchers a tool to see the "hidden details" in massive amounts of data. Before this, analyzing this kind of minute-by-minute, year-long data was too slow and difficult. Now, scientists can quickly see exactly when and how interventions work, allowing for better, more targeted health strategies in the future.

In short: They built a fast, efficient engine to drive through a mountain of complex data, revealing that group activities specifically boost older adults' energy levels in the late morning and early afternoon.

1. Problem Statement

The paper addresses the statistical challenge of analyzing longitudinal functional data where both the outcome and the predictors are functions observed over time.

Context: Modern studies (e.g., the Ready Steady 3.0 study) collect high-dimensional wearable data (minute-level physical activity) across multiple visits and subjects.
The Gap: Existing methods for longitudinal functional regression typically handle either scalar-on-function or function-on-scalar scenarios. While Functional Additive Mixed Models (FAMM) can theoretically handle function-on-function regression with longitudinal structures, they suffer from severe computational limitations. As the number of subjects ( $N$ ) and the density of functional observations increase, FAMM becomes computationally intractable (e.g., failing to finish in a day for datasets with ~300 subjects).
Goal: Develop a method that provides accurate estimation and valid inference for longitudinal function-on-function regression while scaling efficiently to large datasets.

2. Methodology: Efficient Longitudinal Function-on-Function Regression (ELFFR)

The authors propose a novel marginal three-step approach that decomposes the complex joint modeling problem into simpler, parallelizable tasks.

Step 1: Pointwise Longitudinal Scalar-on-Function Regression

Instead of fitting a joint model, the method fits a separate Longitudinal Penalized Functional Regression (LPFR) model at each discrete time point $s_l$ along the functional domain.

Model: $Y_{ij}(s_l) = X_{ij}^T \beta(s_l) + Z_{ij}^T b_i(s_l) + \sum \int W_{ijk}(u)\gamma_k(s_l, u)du + \epsilon_{ij}(s_l)$ .
Mechanism: This treats the functional coefficient $\gamma(s, u)$ as a series of pointwise coefficients. The random effects $b_i(s_l)$ capture within-subject correlation at each specific time point.
Advantage: These models are independent across time points, allowing for massive parallelization.

Step 2: Smoothing

The pointwise estimates obtained in Step 1 are noisy. The method applies smoothing to reconstruct the continuous functional surfaces.

Scalar Coefficients: $\hat{\beta}(s_l)$ are smoothed along the functional domain $S$ using penalized splines (P-splines).
Functional Coefficients: The estimated surfaces $\hat{\gamma}_k(s, u)$ are smoothed along the bivariate domain ( $S \times U$ ) using a sandwich smoother.
Guideline: The authors recommend using a small number of knots (e.g., <10) on the predictor domain to ensure stability and prevent overfitting.

Step 3: Inference (Confidence Bands)

Two approaches are provided for constructing confidence bands:

Analytic Approach (Gaussian Data): Derives a closed-form solution for the covariance of the estimators. It accounts for the correlation between time points via a method-of-moments estimator for the random effect covariance structure $G(s_1, s_2)$ . The final covariance is computed using tensor product properties of the smoother matrices.
Cluster Bootstrap (General): A nonparametric bootstrap of study participants (resampling subjects, not time points) is used for both Gaussian and non-Gaussian data to ensure robust inference.

3. Key Contributions

Computational Efficiency: The primary contribution is a method that scales linearly (or near-linearly) with the number of subjects, unlike the exponential scaling of joint mixed models. It leverages parallel computing to fit thousands of pointwise models simultaneously.
Novel Extension: Extends the marginal approach (previously used for function-on-scalar) to the more complex function-on-function setting. This is non-trivial because it requires handling bivariate smoothing and complex covariance structures for functional coefficients.
Valid Inference: Provides a rigorous inferential framework with analytic confidence bands for Gaussian data and a flexible bootstrap procedure for general cases, ensuring valid coverage probabilities.
Software Implementation: The method is implemented in an R package (available on GitHub), making it accessible to researchers.

4. Simulation Results

The authors conducted extensive simulations comparing ELFFR against the standard FAMM approach.

Accuracy: ELFFR achieves estimation accuracy (measured by Mean Integrated Squared Error) comparable to FAMM, with slight improvements at larger sample sizes.
Inference: The method achieves near-nominal (95%) coverage for confidence bands, particularly with larger sample sizes and the analytic approach.
Speed: The computational gains are dramatic.
- At $N=400$ : ELFFR took ~5.4 minutes vs. FAMM's >68 minutes.
- At $N=800$ : ELFFR took ~35 minutes vs. FAMM's ~13.4 hours.
Scalability: The method remains feasible for large datasets where FAMM fails to converge within a reasonable timeframe.

5. Application: Ready Steady 3.0 (RS3) Study

The method was applied to a physical activity intervention study involving 294 older adults with minute-level data collected over 4 visits.

Data Structure: Functional outcome (PA throughout the day) and functional predictor (baseline PA throughout the day), adjusted for scalar covariates (age, gender, area deprivation index).
Key Findings:
- Interpersonal Interventions: Significantly increased physical activity specifically during the morning hours (9 a.m. – 2 p.m.). This effect was not observed for intrapersonal strategies.
- Temporal Specificity: The analysis revealed that the intervention effect is concentrated in peak activity hours, a nuance lost when data is aggregated into daily totals.
- Covariates: Age showed a significant negative effect on PA in the late afternoon (4 p.m. – 7 p.m.), and Area Deprivation Index (ADI) showed a negative effect in the late morning.
Significance: The study demonstrates that functional regression can uncover time-specific treatment effects that traditional aggregated analyses miss, offering actionable insights for designing behavioral interventions.

6. Significance and Future Directions

Impact: This work bridges the gap between high-dimensional wearable data and rigorous statistical modeling, enabling researchers to analyze complex longitudinal functional data without prohibitive computational costs.
Limitations & Future Work:
- Sensitivity to smoothing parameters (knot selection) requires further investigation for optimal automated selection.
- Analytic inference becomes computationally heavy as the density of the functional domain increases (though subsampling can mitigate this).
- Future research could explore alternative pointwise fitting methods, such as Longitudinal Dynamic Functional Regression (LDFR).

In summary, ELFFR is a scalable, accurate, and practical solution for modern longitudinal functional data analysis, successfully applied to reveal time-specific effects in a large-scale physical activity trial.