Population differences in wearable device wear time: Rescuing data to address biases and advance health equity

This study analyzes Fitbit data from over 11,000 participants to reveal how demographic and health factors influence wearable device wear time, demonstrating that standard compliance thresholds disproportionately exclude data from disease populations and proposing a flexible methodological framework to mitigate these biases and advance health equity.

The Gist

Imagine you are trying to understand how much people walk by asking them to wear a pedometer (a step-counter) on their wrist. You want to know if people with depression walk less than people who are happy, or if older people walk more than younger people.

But here's the problem: Some people forget to wear the watch. Some take it off to shower, some lose it, and some just stop wearing it because they don't feel like it.

This paper is about what happens when researchers decide to throw away all the data from the days people didn't wear their watches. The authors argue that this is a huge mistake, especially for people who are sick or struggling.

Here is the breakdown of their findings using some simple analogies:

1. The "Perfect Day" Trap

Most researchers use a rule called the "10-Hour Rule." They say, "If the watch wasn't worn for at least 10 hours that day, we throw that day's data in the trash."

• The Analogy: Imagine you are a teacher grading a class. You decide that if a student misses more than 2 hours of class, you will throw away their entire test score for that day.
• The Problem: Who misses the most class? Usually, it's the students who are sick, stressed, or dealing with a crisis. By throwing away their "bad days," you aren't just removing bad data; you are erasing the very people you are trying to study. You end up with a class full of healthy, happy students and think, "Wow, everyone in this class is doing great!" while ignoring the ones who are struggling.

2. Who Actually Wears the Watch?

The authors looked at data from nearly 12,000 people and found that wearing the watch isn't random. It's like a popularity contest where certain groups are more likely to participate:

• The "Golden Ticket" holders: Men, older adults, people with more money, and people with higher education levels tend to wear their watches more consistently.
• The "Forgotten" groups: People with lower incomes, people of color, and people dealing with mental health struggles (like depression or anxiety) are much more likely to take the watch off or forget to wear it.

The Metaphor: It's like trying to study the weather by only looking out the windows of houses with big, expensive skylights. You'll get a great view of the sunny days, but you'll miss all the rain because the people in the smaller houses (or the ones with broken windows) aren't looking up.

3. The Depression Connection (The Big Discovery)

This is the most critical finding. The researchers looked at people with Major Depressive Disorder (MDD).

• The Result: People with depression wore their watches significantly less than healthy people.
• The "Trash" Stat: When the researchers applied the strict "10-Hour Rule," they threw away 74.4% of the data days for people with depression. For healthy people, they only threw away 20.9%.
• The Analogy: Imagine you are trying to measure how heavy a backpack is. But, every time someone with a heavy backpack (depression) tries to show you, you say, "Oh, you're holding it too loosely, I can't measure it," and you throw their backpack in the trash. Meanwhile, you keep measuring the light backpacks of healthy people. You end up thinking, "Everyone's backpack is light!" because you threw away the heavy ones.

4. The Solution: Don't Throw the Data Away!

The authors propose a new way to handle this. Instead of throwing away the "bad" days, we should keep them and adjust our math.

They suggest five new tools (like a mechanic's toolkit) to fix this:

1. The "Adjustment" Knob: Keep all the data, but tell the computer, "Hey, this person wore the watch less today, so let's adjust the math to account for that."
2. The "Speedometer" Method: Instead of counting total steps (which is low if you didn't wear the watch), count steps per hour of wear. This is like measuring how fast a car is going, regardless of how long the trip was.
3. The "Balanced Scale": Use a statistical trick to match sick people with healthy people who have similar watch-wearing habits, so the comparison is fair.
4. The "Flexible Rule": Instead of a hard 10-hour rule, maybe use an 8-hour rule if that fits the specific group better.

Why This Matters

If we keep using the old "10-Hour Rule," we are accidentally biasing our research. We are making it look like sick people are healthier than they really are, or that certain groups are more active than they are, simply because we threw away their data.

The Bottom Line:
Wearable devices are amazing tools for health, but they are only useful if we stop treating "forgetting to wear the watch" as a mistake to be deleted. Forgetting to wear the watch is actually a symptom. It tells us something important about a person's health, their stress levels, or their daily struggles. By keeping that data and analyzing it correctly, we can build a fairer, more accurate picture of health for everyone, not just the people who are lucky enough to remember to charge their watches.

Technical Summary

1. Problem Statement

Wearable devices (e.g., Fitbits) offer transformative potential for personalized healthcare by providing continuous, passive digital biomarkers. However, a critical methodological gap exists: individual variations in device wear time are often non-random and correlated with demographic factors, social determinants of health (SDoH), and disease states.

Current research practices typically apply a rigid compliance threshold (e.g., discarding days with <10 hours of wear time) to ensure data quality. The authors argue that this approach introduces selection bias, systematically excluding data from populations with lower adherence (e.g., individuals with depression, lower socioeconomic status, or specific chronic conditions). This exclusion not only reduces statistical power but also distorts clinical effect estimates by removing data that may itself be a clinically meaningful signal of disease burden.

2. Methodology

The study utilized data from the All of Us (AoU) Research Program (Controlled Tier v7), analyzing 11,901 participants who linked their personal Fitbit devices.

• Data Sources:
  • Wearable Data: Fitbit intra-day step data used to estimate wear time.
  • Demographics/SDoH: Age, sex, race/ethnicity, income, education, and insurance status.
  • Clinical Data: Electronic Health Records (EHR) for diagnoses (Major Depressive Disorder [MDD], Anxiety Disorder [AD], and various chronic conditions) and survey data for symptoms (depression, anxiety, anhedonia).
• Wear Time Metrics:
  • Percentage of Days Worn: Calculated as days with step data / total days of potential wear.
  • Daily Hours Worn: Calculated as hours with non-zero step counts.
• Statistical Analysis:
  • Linear regression and linear mixed-effects models (accounting for within-participant correlation) were used to compare wear time across groups.
  • Models were adjusted for age, sex, race/ethnicity, income, education, and insurance status.
  • Methodological Comparison: The authors evaluated five distinct analytical approaches for handling wear time variability using MDD as a case study:
    1. Standard Compliance Filtering: Excluding days <10 hours.
    2. Covariate Adjustment: Including daily wear hours as a covariate in the model.
    3. Metric Normalization: Converting total steps to "steps per hour."
    4. Propensity Score Matching: Matching participants on wear time distributions.
    5. Adaptive Thresholds: Systematically testing lower thresholds (1–10 hours).

3. Key Results

A. Demographic and SDoH Disparities

Wear time was significantly non-random across populations:

• Demographics: Wear time increased with age (peaking at 60–70 years) and sex (males wore devices more than females).
• SDoH: Higher income and education levels correlated with increased wear time. Individuals with health insurance wore devices significantly more than those without.
• Race/Ethnicity: Hispanic/Latino and Black non-Hispanic individuals wore devices fewer days and hours compared to White non-Hispanic counterparts.
• Lifestyle: No significant differences were found based on smoking or alcohol use.

B. Mental Health and Disease Associations

• Symptoms vs. Diagnoses: Individuals reporting depressive, anxiety, and anhedonia symptoms wore devices fewer days. This effect was amplified in diagnosed populations (MDD/AD).
  • Example: Individuals with MDD wore devices 59.7% fewer days and 9.59 fewer hours per day compared to controls.
• Chronic Conditions: Wear time patterns varied by disease class.
  • Reduced Wear: Observed in genitourinary, injury/poisoning, musculoskeletal, and respiratory disorders.
  • Increased Wear: Observed in blood disorders, infectious diseases, and perinatal conditions (potentially due to increased monitoring needs).
• Pre/Post Diagnosis: For many conditions, wear time declined significantly after diagnosis, suggesting the condition itself (or treatment side effects) impacts device engagement.

C. Impact of Analytical Approaches (The MDD Case Study)

Comparing daily step counts between MDD and non-MDD groups revealed how handling wear time affects results:

• Data Loss: The standard ≥10-hour threshold excluded 74.4% of data days for the MDD group, compared to only 20.9% for the control group.
• Bias in Estimates:
  • Standard Filtering: Showed the largest difference in steps (999 steps/day), likely inflated by the exclusion of low-activity days in the MDD group.
  • Covariate Adjustment: Retained 100% of data and showed a smaller, yet significant, difference (796 steps/day), suggesting part of the "gap" in filtered data was due to wear time variation rather than activity alone.
  • Propensity Matching: Resulted in massive data loss (93% excluded), rendering the result statistically insignificant due to low power.
  • Metric Normalization: Confirmed reduced activity (55 fewer steps/hour) while retaining 78.8% of data.

4. Key Contributions

1. Systematic Characterization: This is the first large-scale study to map wearable wear time across a diverse spectrum of demographics, SDoH, mental health symptoms, and chronic diseases.
2. Evidence of Selection Bias: Demonstrated that rigid compliance thresholds disproportionately exclude data from vulnerable and diseased populations, threatening health equity in digital health research.
3. Methodological Framework: Proposed and empirically validated a flexible framework for handling wear time, moving beyond a "one-size-fits-all" threshold. The authors advocate for:
   • Covariate Adjustment to retain maximum data and control for wear time statistically.
   • Metric Normalization for rate-based comparisons.
   • Adaptive Thresholds tailored to specific study populations.
4. Reframing Wear Time: Posited that wear time itself is a digital biomarker reflecting disease severity or engagement, rather than just a nuisance variable to be filtered out.

5. Significance

• Advancing Health Equity: By adopting flexible analytical methods, researchers can prevent the systematic exclusion of data from underrepresented and clinically complex populations, ensuring that digital health insights are generalizable and equitable.
• Improved Statistical Rigor: Moving away from strict filtering prevents the distortion of clinical effect sizes and increases statistical power by utilizing the full dataset.
• Clinical Utility: Recognizing that reduced wear time is often a symptom of the disease (e.g., fatigue in depression) allows researchers to use wear time as an additional feature in predictive models, rather than discarding it.
• Guidance for Future Research: The paper provides a practical toolkit for researchers to choose the appropriate method (covariate adjustment, normalization, etc.) based on their specific research question and population, ensuring robust and inclusive digital health analytics.