📊 statistics

Scalar-on-distribution regression via generalized odds with applications to accelerometry-assessed disability in multiple sclerosis

This paper proposes a unified generalized odds regression framework that models subject-specific distributions via probability ratios to improve the prediction of disability in multiple sclerosis using accelerometry data, outperforming traditional scalar and survival-based approaches.

Original authors: Pratim Guha Niyogi, Muraleetharan Sanjayan, Kathryn C. Fitzgerald, Ellen M. Mowry, Vadim Zipunnikov

Published 2026-01-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pratim Guha Niyogi, Muraleetharan Sanjayan, Kathryn C. Fitzgerald, Ellen M. Mowry, Vadim Zipunnikov

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 a person's health by looking at their daily activity. Traditionally, doctors and researchers have used a "summary score," like the total number of steps taken in a day. The paper argues that this is like trying to understand a movie by only looking at the average brightness of the screen. You miss the dramatic explosions, the quiet whispers, and the intense action scenes that actually tell the story.

Here is a simple breakdown of what the paper does, using everyday analogies:

1. The Problem: The "Average" Lie

The researchers studied people with Multiple Sclerosis (MS) who wore wristwatches (accelerometers) that tracked their movement every minute.

The Old Way: They used to calculate the "average" activity. If someone sat all day but had one 10-minute burst of running, the average might look "okay."
The Issue: The paper suggests that for MS, the extreme moments matter most. It's not the average activity that predicts disability; it's the ability (or inability) to handle high-intensity bursts of movement. The "tails" of the data (the very high or very low activity levels) hold the secret clues.

2. The Solution: The "Generalized Odds" Framework

The authors created a new mathematical tool called Generalized Odds (GO).

The Analogy: Imagine you are comparing two people's activity. Instead of asking, "How much did they move?" you ask, "How much more likely is it that this person is running at full speed compared to just sitting on the couch?"
How it works: The GO framework calculates a ratio. It compares the probability of "extreme" events (like high-intensity activity) against "typical" events (like moderate activity).
Why it's special: This tool is flexible. It can act like a "Survival Function" (how long they keep moving), a "Hazard Function" (the risk of stopping), or a "Residual Life" (how much more activity they have left), but it does it all under one roof. It handles messy, real-world data (like when a watch falls off or data is missing) better than older methods.

3. The Method: Smoothing the Rough Edges

Because the data comes in thousands of tiny 1-minute chunks, it's too messy to analyze directly.

The Analogy: Think of the data as a jagged, rocky mountain range. The researchers used "splines" (a mathematical smoothing tool) to turn that jagged mountain into a smooth, flowing hill. This allows them to see the overall shape of the activity pattern without getting stuck on every single rock.
The "Penalty": To make sure they didn't overcomplicate the model (like trying to explain a movie with 10,000 different plot twists), they used a "penalty" system. This acts like a strict editor, cutting out unnecessary details and keeping only the patterns that truly matter.

4. The Results: A Clearer Picture

The team tested this new method on data from the HEAL-MS study (a group of MS patients).

The Comparison: They compared their new "Odds" method against the old "Average" method and other standard methods.
The Outcome: The new method was a clear winner.
- The old "Average" method explained about 9.5% of the variation in disability scores.
- The new "Generalized Odds" method explained about 21% of the variation.
What this means: By focusing on the ratio of high-intensity activity to low-intensity activity, the new model predicted the patients' disability levels (measured by the EDSS score) nearly twice as well as the previous best methods.

5. The Bottom Line

The paper claims that to understand complex health conditions like MS, we need to stop just counting the "total steps" and start looking at the balance between extreme effort and rest.

By using this "Generalized Odds" framework, researchers can see the "tail behavior" of a patient's activity—specifically, how they handle high-intensity moments. This provides a much sharper, more accurate picture of their disability than simply averaging their day. The paper concludes that this approach is a powerful new way to turn raw digital health data into meaningful medical insights.

Technical Summary: Scalar-on-Distribution Regression via Generalized Odds

Problem Statement
Digital health technologies (DHT) generate continuous, high-resolution behavioral data (e.g., minute-level accelerometry) that form rich subject-specific distributions. While these distributions contain more information than scalar summaries (e.g., total daily activity counts), identifying which distributional features are most predictive of clinical outcomes remains an open challenge. Previous work by Niyogi et al. [2024] demonstrated that hazard-based representations, which emphasize tail behavior (extreme activity levels), outperform other representations in predicting disability in multiple sclerosis (MS). However, existing methods often rely on specific distributional objects (densities, quantiles, survival, or hazard functions) that may not fully capture the contrast between extreme and typical values, particularly in sparse or discrete data contexts.

Methodology
The authors propose a unified Generalized Odds (GO) framework to represent subject-specific distributions and a corresponding scalar-on-odds regression model.

Generalized Odds Definition:
The framework defines a subject-specific generalized odds function, $h_{4,F}$ , as the ratio of probabilities of two arbitrary events within the sample space. For a random variable $X$ with distribution $F$ , and intervals $A=[u_1, u_2]$ and $B=[u_3, u_4]$ , the GO is defined as:
$h_{4,F}(u_1, u_2, u_3, u_4) = \frac{P(X \in [u_3, u_4])}{P(X \in [u_1, u_2])} = \frac{F(u_4) - F(u_3)}{F(u_2) - F(u_1)}$
This formulation generalizes existing distributional objects:
- Probability Mass Function (PMF): $h_{4,F}(0, D, k, k-1) = p(k)$ .
- Hazard Function: $h_{4,F}(k, D, k-1, k) = \lambda(k)$ .
- Residual Life Distribution: $h_{4,F}(t, D, t, t+u)$ .
- 1-index and 2-index objects: Special cases comparing survival to cumulative probability ( $S(u)/F(u)$ ) or survival to survival ( $S(u_2)/F(u_1)$ ).
Regression Model:
The authors formulate a scalar-on-functional regression model where the clinical outcome $y_i$ (e.g., EDSS score) is linked to the GO function $h_{F_i}$ via a link function $g$ :
$g(\mu_i) = \alpha + \int \beta(u) h_{F_i}(u) du$
where $\mu_i = E[y_i | X_i]$ . The coefficient function $\beta(u)$ is estimated using a spline-based functional representation. Specifically, $\beta$ is approximated via a tensor product of B-spline basis functions:
$\beta(u_1, u_2, u_3, u_4) = \sum_{k_1}\sum_{k_2}\sum_{k_3}\sum_{k_4} b_{k_1,k_2,k_3,k_4} B_{k_1}(u_1)B_{k_2}(u_2)B_{k_3}(u_3)B_{k_4}(u_4)$
Estimation and Regularization:
To handle the high dimensionality of the tensor coefficients and identify informative distributional regions, the authors employ penalized likelihood estimation. They evaluate four penalty functions to induce sparsity and variable selection:
- Lasso ( $L_1$ )
- Elastic Net
- Smoothly Clipped Absolute Deviation (SCAD)
- Minimax Concave Penalty (MCP)

Data and Application
The methodology is applied to data from the HEAL-MS study (Home-based Evaluation of Actigraphy to Predict Longitudinal Function in Multiple Sclerosis), involving 248 participants with MS.

Predictors: Wrist-worn accelerometry data (GT9X Actigraph) collected over 2 weeks, processed into 1-minute epoch activity counts (log-transformed) from 8 AM to 8 PM.
Outcome: Expanded Disability Status Scale (EDSS) scores (0–10).
Comparison: The GO models are compared against classical scalar summaries (average activity counts) and scalar-on-survival models.

Key Results
Using 5-fold cross-validation with 100 replications, the study reports the following predictive performance (measured by cross-validated $R^2$ ) for EDSS scores:

Scalar Summary (Average Activity): $R^2 \approx 0.095$ .
Scalar-on-Survival: $R^2 \approx 0.115$ .
Scalar-on-1-index Odds: $R^2 \approx 0.151$ (Lasso).
Scalar-on-2-index Odds: $R^2 \approx 0.146 - 0.151$ .
Scalar-on-4-index Odds (Generalized): $R^2 \approx 0.208 - 0.209$ .

The 4-index GO model consistently outperformed all other approaches, nearly doubling the predictive performance of the scalar summary and significantly improving upon the survival-based model. The Elastic Net and Lasso penalties yielded the highest $R^2$ values for the 4-index model.

Significance and Claims
The paper claims that the Generalized Odds framework offers a significant methodological advance by:

Directly Targeting Tail Behavior: Unlike single-index representations (e.g., densities) that summarize overall shape, GO functions explicitly contrast extreme values against typical values, capturing the "tail behavior" identified as critical in prior MS research.
Robustness: The ratio-based formulation remains well-defined and stable even for sparse, bounded, or discrete data where direct density estimation might fail.
Unified Framework: It subsumes hazard, survival, and residual life representations as special cases, providing a flexible, unified approach for distributional covariates.

The authors conclude that odds-based distributional covariates provide superior predictive value for DHT data in clinical settings, specifically demonstrating that tail behavior in physical activity is a critical driver of disability outcomes in multiple sclerosis. They note that while the application focused on MS, the framework is broadly applicable to other domains where contrasts between extreme and typical behaviors are scientifically meaningful.