🤖 machine learning

Age-Dependent Heterogeneity in the Association Between Physical Activity and Mental Distress: A Causal Machine Learning Analysis of 3.2 Million U.S. Adults

This study of over 3.2 million U.S. adults reveals that while physical activity significantly reduces mental distress in older populations, this protective effect is absent in young adults (18–24) and has eroded over the past decade, suggesting that the exercise-mental health link does not generalize to the youngest demographic facing a deepening mental health crisis.

Original authors: Yuan Shan (Department of Statistical Science, Duke University)

Published 2026-04-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuan Shan (Department of Statistical Science, Duke University)

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

The Big Idea: One Size Does Not Fit All

Imagine you have a magic shield that protects you from bad days, stress, and anxiety. For decades, scientists have told us that exercise is that shield. They say, "Just go for a run, and your mental health will get better."

This study, which looked at a massive group of 3.2 million Americans over ten years, asks a simple but crucial question: "Does this shield work the same way for a 20-year-old as it does for a 60-year-old?"

The answer is a resounding no.

The study found that the "mental health shield" of exercise is age-dependent. It works incredibly well for older adults, but for young adults, the shield is cracked, and in some recent years, it might not be working at all.

The "Magic Shield" Analogy: How It Works by Age

Think of physical activity as a vitamin for your mind.

For Adults Aged 55–64: Taking this vitamin is like a superpower. The study found that for people in this age group, exercising cuts their risk of frequent mental distress by 50%. It's like having a heavy-duty umbrella in a storm; it keeps you completely dry.
For Adults Aged 18–24: Taking this vitamin is like taking a single aspirin for a broken leg. The study found that for young adults, exercise only reduces the risk of mental distress by about 11%.
The Trend: The "protection" gets stronger the older you get. It starts weak at age 18 and gets stronger and stronger until it plateaus around age 55.

The Scary Trend: The Shield is Cracking for Young People

The most alarming part of the study is a time-travel story. The researchers looked at data from 2015 to 2024.

In 2015: Exercise still helped young people a little bit.
In 2018 & 2024: The data showed that for 18–24-year-olds, exercise had zero effect on their mental distress. The "magic shield" had completely vanished.

Why is this happening?
The authors suggest that the problems young people face today are different from the problems older people face.

Older adults might feel stressed about health, retirement, or loneliness. Exercise helps fix these physical and social issues.
Young adults are drowning in digital stress: social media comparison, economic anxiety, and the pressure of the "hustle culture."

The Analogy: Imagine a young person is being attacked by a swarm of angry bees (social media stress, debt, loneliness). Telling them to "go for a jog" is like telling them to "swat at the bees with a feather." It's a nice gesture, but it doesn't stop the swarm. The stressors are too big and too cognitive (in the head) for a simple jog to fix.

How Did They Figure This Out? (The Detective Work)

The researchers didn't just guess; they used two different detective tools to make sure they were right:

The Traditional Detective (Logistic Regression): They looked at the data like a spreadsheet, grouping people by age and checking the numbers. They saw the clear pattern: the older you are, the better exercise works.
The AI Detective (Causal Machine Learning): They used a sophisticated computer algorithm called a "Causal Forest." This AI didn't know the answer beforehand. It just looked at millions of data points and asked, "What factor changes the result the most?"
- The AI's Answer: "Age."
- The AI determined that age was the single most important reason why exercise helps some people and not others. It was 2.5 times more important than any other factor (like race, income, or gender).

The "Placebo" Check (Did We Mess Up?)

To make sure they hadn't made a mistake, they ran a "fake test." They asked: "Does exercise change your gender?" (Obviously, no).

If their method was broken, it might have found a weird link between exercise and gender that changed by age.
It didn't. The method only found the weird link between exercise and mental health for young people. This confirmed that the results were real and not a glitch.

The Takeaway for Everyone

1. Don't stop exercising!
Even for young people, exercise is still good for your heart, muscles, and sleep. It just might not be the "cure-all" for the specific type of anxiety and depression that is crushing young people today.

2. We need new tools for young people.
If a young person is struggling, telling them "just go run" might feel dismissive. Because the study shows that exercise alone isn't fixing the problem, we need to address the real causes: social media habits, economic pressure, and isolation.

3. The "One-Size-Fits-All" advice is outdated.
Public health advice needs to change. We can't tell a 20-year-old and a 60-year-old to do the exact same thing to fix their mental health. The older adult needs the "exercise shield," but the young adult needs a different kind of armor entirely.

Summary in One Sentence

While exercise is a powerful shield against mental distress for older adults, that shield has become ineffective for young adults in the last decade, suggesting that the unique, modern stressors facing youth require solutions beyond just "getting moving."

1. Problem Statement

While physical activity (PA) is widely established as a protective factor against mental distress, existing literature often assumes this benefit is uniform across demographic subgroups. A landmark study by Chekroud et al. (2011–2015) concluded that PA benefits were "consistent" across age groups. However, with the rising youth mental health crisis and the emergence of new stressors (e.g., social media, economic precarity), there is a critical gap in understanding whether the PA–mental health link holds for younger adults.

This study addresses two primary questions:

Does the protective association between leisure-time physical activity and frequent mental distress (FMD) vary systematically by age?
Has this association changed over the last decade (2015–2024), particularly for young adults?

2. Methodology

Data Source and Sample

Dataset: Pooled data from ten consecutive annual waves of the U.S. Behavioral Risk Factor Surveillance System (BRFSS) (2015–2024).
Sample Size: $N = 3,242,218$ U.S. adults (after complete-case analysis of an initial 4.4 million).
Design: Cross-sectional, nationally representative, using disproportionate stratified sampling (DSS) and survey weights (_LLCPWT).

Variables

Outcome: Frequent Mental Distress (FMD), defined as $\ge$ 14 days of poor mental health in the past 30 days.
Exposure: Leisure-time Physical Activity (PA), binary indicator (active vs. inactive in the past 30 days).
Covariates: Sex, Age Group (6 categories), Race/Ethnicity, Education, Income, BMI, and Survey Year.

Analytical Framework

The study employs a triangulation approach combining traditional epidemiology with causal machine learning:

Survey-Weighted Logistic Regression:
- Models the log-odds of FMD with PA and covariates.
- Tests for effect modification via interaction terms ( $PA \times Age$ ) and stratified models.
- Uses sandwich estimators for heteroskedasticity-consistent standard errors.
Causal Forest via Double Machine Learning (DML):
- Goal: To estimate the Conditional Average Treatment Effect (CATE) in a non-parametric, data-driven manner without pre-specifying subgroups.
- Implementation: Used the CausalForestDML class from the EconML library.
- Mechanism: Residualizes outcome and treatment against confounders using Gradient Boosting (HistGradientBoosting), then estimates CATE via honest causal forests.
- Feature Importance: Quantifies which covariates drive heterogeneity in treatment effects.
- Post-hoc Analysis: Applied a CART tree to estimated CATEs to derive interpretable subgroup rules.
Robustness and Sensitivity Checks:
- Temporal Validation: Year-specific stratified regressions and a three-way interaction ( $PA \times Age \times Year$ ).
- E-value Analysis: Quantifies the strength of unmeasured confounding required to nullify observed associations.
- Propensity Score Overlap: Ensures adequate common support between active/inactive groups.
- Placebo Test: Estimated PA effects on sex (a pre-determined covariate) to rule out systematic artifacts.
- Imputation Sensitivity: Compared complete-case results with conditional random imputation for missing income data.

3. Key Contributions

Discovery of a Monotonic Age Gradient: The study provides the first large-scale evidence that the protective effect of PA on mental distress is not uniform but increases monotonically with age.
Temporal Erosion in Young Adults: It documents a novel trend where the PA benefit for adults aged 18–24 has eroded over the past decade, reaching statistical nullity in 2018 and 2024.
Causal ML Validation: It independently confirms age as the dominant driver of heterogeneity using Causal Forests (feature importance = 0.39), validating parametric findings without pre-specified subgroup structures.

4. Key Results

Age-Stratified Associations (Pooled 2015–2024)

The adjusted Odds Ratios (OR) for PA reducing FMD show a dramatic gradient:

18–24 years: OR = 0.89 (11% reduction; weak protection).
25–34 years: OR = 0.80.
35–44 years: OR = 0.67.
45–54 years: OR = 0.53.
55–64 years: OR = 0.50 (50% reduction; strongest protection).
65+ years: OR = 0.53.

Temporal Trends (18–24 Age Group)

The protective effect for young adults has weakened significantly over time:

2015: OR = 0.86.
2018: OR = 1.01 (Null effect).
2022: OR = 0.995 (Near null).
2024: OR = 1.01 (Null effect).
The three-way interaction ( $PA \times Age \times Year$ ) was highly significant ( $p < 0.001$ ), confirming the effect is evolving.

Causal Machine Learning Findings

Feature Importance: Age was the single most important predictor of treatment effect heterogeneity (Importance Score = 0.39), outperforming the next best predictor (Year, 0.156) by 2.5x.
CATE Distribution: The Causal Forest estimated an overall Average Treatment Effect (ATE) of -0.061. The CATE distribution showed substantial heterogeneity, with the 55–64 age group showing the most significant negative effect (CATE = -0.093), while the youngest group hovered near zero.

Sensitivity Analysis

E-values: For older groups (OR $\approx$ 0.50), the E-value exceeded 3.2, indicating robustness to unmeasured confounding. For the 18–24 group (OR = 0.89), the E-value was only 1.49, consistent with a genuinely weak effect easily nullified by minor confounding.
Placebo Test: The age-heterogeneous pattern was not observed in the placebo test (PA effect on sex), ruling out systematic methodological artifacts.

5. Significance and Implications

Re-evaluating Public Health Guidelines: The findings challenge the assumption that "exercise is good for everyone's mental health" in a uniform way. For young adults (18–24), PA alone may be insufficient to mitigate distress driven by modern cognitive-social stressors (e.g., social media, economic instability).
Mechanistic Insights: The authors suggest that while PA benefits older adults via neurobiological pathways (inflammation reduction, HPA axis regulation), young adult distress may be more cognitive-social in nature, requiring different interventions.
Methodological Advancement: The study demonstrates the power of combining traditional survey-weighted regression with Double Machine Learning (Causal Forests) to uncover complex, non-linear heterogeneity in observational data that might be missed by standard stratification.
Policy Recommendation: Mental health strategies for young adults must move beyond generic PA recommendations and integrate interventions targeting specific stressors (e.g., digital well-being, economic support) alongside physical activity.

6. Limitations

Causality: Despite using Causal Forests, the cross-sectional design prevents definitive causal inference; unconfoundedness cannot be verified.
Measurement: PA is binary (any/none), preventing dose-response analysis.
Self-Report: All data are self-reported, subject to recall and social desirability bias.
Subsample Size: The Causal Forest analysis used a subsample ( $N=50,000$ ) for computational feasibility, resulting in wider confidence intervals compared to the full-sample regression.