Target trial emulation of physical activity and cardiovascular disease risk: What is impact of the exposure assessment method?

This study demonstrates that in target trial emulation of physical activity and cardiovascular disease risk, relying on self-reported exposure measurements obscures intervention effects due to measurement error and increases Type II error risk, whereas objective wearable-based assessments reveal a significant protective effect of meeting physical activity guidelines.

The Gist

Imagine you are trying to figure out if switching from a sedentary lifestyle to an active one actually saves your heart. To do this perfectly, you would ideally run a massive experiment: take 10,000 people, lock them in a room, force half to sit on the couch for a year and force the other half to run marathons, and then see who gets sick.

But obviously, we can't do that. It's unethical and impossible. So, scientists use a clever trick called "Target Trial Emulation." Think of this like a high-tech flight simulator. You can't test a new plane in a real storm, so you build a perfect digital simulation of the storm and the plane to see what happens. In this study, the researchers built a "flight simulator" using real-world data from the UK Biobank to see if getting active protects the heart.

However, this study asks a very specific, critical question: What happens if the sensors in our flight simulator are broken?

The Two Sensors: The "Guessing Game" vs. The "Truth Machine"

The researchers wanted to see if the way they measured physical activity changed the results. They compared two very different tools:

1. The "Guessing Game" (Self-Reports): This is like asking someone, "How much did you run last week?" and them trying to remember while sitting at a kitchen table.
   • The Problem: Humans are terrible at this. We forget the short walks, we exaggerate the gym sessions, and we only remember the "big" activities. It's like trying to measure the ocean with a teaspoon; you miss almost everything.
2. The "Truth Machine" (Wearables): This is a smartwatch or fitness tracker that sits on your wrist 24/7.
   • The Benefit: It doesn't care if you remember or if you want to look good. It counts every step, every shuffle, and every sprint with laser precision. It sees the whole ocean, not just the teaspoon.

The Experiment: The Great Heart Race

The researchers took two groups of people who were currently inactive (sitting on the couch). They watched them over time to see who started moving enough to meet health guidelines (150 minutes of activity a week) and who stayed inactive.

They then ran their "flight simulator" twice:

• Run A: They used the Guessing Game data to decide who was active.
• Run B: They used the Truth Machine data to decide who was active.

The Results: A Tale of Two Stories

Here is where it gets fascinating. The two sensors told completely different stories about the same people.

Story 1: The "Guessing Game" (Self-Report)
When the researchers used the self-reported data, the results were boring and confusing.

• The "Active" group and the "Inactive" group had almost the exact same risk of heart disease.
• It looked like moving your body didn't help at all.
• The Metaphor: It's like trying to hear a whisper in a noisy room. The signal (the benefit of exercise) was so drowned out by the static (bad memory and lying) that you couldn't hear it. The scientists concluded, "Exercise doesn't seem to matter," but they were actually just listening to the wrong microphone.

Story 2: The "Truth Machine" (Wearables)
When they switched to the smartwatch data, the results were dramatic and clear.

• The group that actually moved (according to the watch) had a massive drop in heart disease risk compared to the couch potatoes.
• The risk was nearly cut in half!
• The Metaphor: This is like turning on the noise-canceling headphones. Suddenly, the whisper becomes a clear voice. The "Truth Machine" saw the real movement, and the data screamed, "Exercise saves lives!"

Why Did This Happen?

The study found that the "Guessing Game" was so inaccurate that it created a Type II Error. In science-speak, this means the study failed to find a real effect because the measuring tool was too fuzzy.

Because people misremembered their activity, the "Active" group in the self-report study was actually full of people who thought they were active but weren't really. They were mixed in with the "Inactive" group. This muddied the water so much that the scientists couldn't see the difference between the two groups.

The Big Takeaway

This paper is a wake-up call for public health.

For years, many studies have used the "Guessing Game" (questionnaires) and found weak or non-existent links between exercise and health. This study suggests those studies might have been wrong not because exercise doesn't work, but because we were measuring it with a broken ruler.

The Conclusion:
If you want to know if a new medicine works, you don't ask patients to guess how much they took; you check the pill bottle. Similarly, if we want to know if exercise saves lives, we can't rely on people's memories. We need the smartwatches.

The authors are saying: "Stop guessing. Start measuring." If we want to make the best health policies for the world, we need to use the "Truth Machines" (wearables) to see the real picture, or we might miss the fact that moving our bodies is one of the most powerful medicines we have.

Technical Summary

1. Problem Statement

Target Trial Emulation (TTE) is a rigorous observational framework designed to strengthen causal inference by mimicking the protocol of a hypothetical Randomized Controlled Trial (RCT). While TTE effectively addresses confounding and Type I errors (false positives), its vulnerability to Type II errors (false negatives) when the exposure variable is measured imprecisely remains under-explored.

In physical activity epidemiology, the "gold standard" for exposure assessment has historically been self-reported questionnaires (e.g., IPAQ). These tools suffer from:

• Non-differential misclassification: Random errors in recall that bias results toward the null.
• Partial behavioral capture: They often miss short bouts of activity and rely on long-term recall.
• Low precision: They fail to capture the full intensity spectrum of movement.

The central research question is: Does the reliance on self-reported physical activity in TTE designs obscure true causal effects on Cardiovascular Disease (CVD) risk, leading to erroneous null findings, compared to high-precision wearable device data?

2. Methodology

The study utilized data from the UK Biobank cohort to emulate a randomized trial comparing two groups of physically inactive adults:

• Intervention Group: Participants who adopted guideline-concordant physical activity ($\ge$150 minutes of Moderate-to-Vigorous Physical Activity [MVPA] per week) at follow-up.
• Control Group: Participants who remained physically inactive (<150 MVPA mins/week) at follow-up.

Study Design & Samples:
The authors conducted parallel TTE analyses on two distinct sub-samples to isolate the effect of the measurement method:

1. Wearables Sample ($N=490$): Participants with repeat accelerometer assessments (Axivity AX3). Exposure was quantified using a 2-stage Random Forest machine learning classifier on raw acceleration data (high temporal resolution, 100Hz sampling).
2. Self-Report Sample ($N=11,302$): Participants with repeat International Physical Activity Questionnaire (IPAQ) assessments. Exposure was calculated using standard MET-minutes scoring.

Analytical Approach:

• Eligibility: Adults aged 40–69 who were inactive at baseline.
• Propensity Score Matching (PSM): A 1:1 nearest-neighbor matching (caliper 0.05 SD) was applied to balance covariates including age, sex, education, ethnicity, diet, smoking, sleep, screen time, medication use, and prevalent comorbidities.
• Outcome: Incident CVD (ischemic heart disease, heart failure, stroke, hypertensive heart disease) identified via linked NHS hospital records and mortality registries.
• Statistical Models:
  • Fine-Gray subdistribution hazard models: To estimate cumulative incidence accounting for competing risks (death from other causes).
  • Modified Poisson regression: To estimate Risk Ratios (RR) and Risk Differences.
  • Sensitivity Analyses: Included negative control outcomes (accidental deaths), exclusion of early events, and restricting to older adults ($\ge$60 years).

3. Key Contributions

• Direct Comparison of Measurement Methods: This is the first study to directly compare TTE causal estimates for the same research question using self-report vs. objective wearable data within the same cohort.
• Demonstration of Measurement-Induced Bias: The study empirically proves that exposure measurement error is a dominant source of bias in TTE, capable of masking true causal effects even when confounding is rigorously controlled.
• Refutation of "Null Effect" Hypotheses: The findings challenge previous TTE studies in physical activity that reported weak or null associations with health outcomes, suggesting these null results may be artifacts of poor measurement rather than a lack of biological efficacy.

4. Results

Wearables-Based Analysis (Objective Measurement):

• Sample: 245 intervention vs. 245 control participants.
• Follow-up: Mean 4.4 years.
• Cumulative Risk: The intervention group had a significantly lower cumulative CVD risk (8.0%) compared to the control group (17.0%).
• Effect Estimates:
  • Hazard Ratio (HR): 0.59 [95% CI: 0.36, 0.98]
  • Relative Risk (RR): 0.45 [95% CI: 0.28, 0.72]
  • Risk Difference: -2.59% (statistically significant).

Self-Report-Based Analysis (Subjective Measurement):

• Sample: 5,651 intervention vs. 5,651 control participants (much larger sample size).
• Follow-up: Mean 6.3 years.
• Cumulative Risk: The risk trajectories for both groups were nearly identical. Intervention: 21.2% vs. Control: 21.6%.
• Effect Estimates:
  • Hazard Ratio (HR): 0.98 [95% CI: 0.89, 1.08]
  • Relative Risk (RR): 0.92 [95% CI: 0.84, 1.00]
  • Risk Difference: -1.24% (statistically non-significant).

Supplemental Findings:

• Matching the self-report sample to the wearables sample on demographics and health factors did not alter the null findings in the self-report group.
• Sensitivity analyses (excluding early events, using higher activity thresholds, negative controls) confirmed the robustness of the wearables findings and the persistence of null findings in self-reports.

5. Significance and Implications

• Validity of Causal Inference: The study concludes that objective measurement is a prerequisite, not just a refinement, for valid causal inference in physical activity TTE studies. Relying on self-reports introduces non-differential misclassification that systematically attenuates effect sizes toward the null (Type II error).
• Policy Impact: Previous guidelines or policy decisions based on TTE studies using self-reported data may have underestimated the protective benefits of physical activity against CVD.
• Methodological Shift: Future observational studies aiming to emulate trials for physical activity must prioritize device-based, high-resolution measurements. Increasing sample size (as seen in the self-report group) cannot compensate for poor exposure measurement precision.
• Reinterpretation of Literature: The authors suggest that many existing TTE studies reporting null or weak links between physical activity and chronic diseases (COPD, cancer, mental health) may need re-evaluation, as their null findings could be driven by measurement error rather than a lack of biological effect.

In summary, the paper provides compelling evidence that measurement precision is the critical bottleneck in emulating physical activity interventions, and without objective data, TTE designs risk failing to detect life-saving interventions.