Health eHeart is a prospective longitudinal cohort study that has enrolled all interested adult volunteers (age≥18 years) who have an email address and have been residing within the United States since March 2013. Participation in the study is entirely web-based and recruitment was conducted on the internet through social media advertisements, email campaigns with advocacy and research organizations, and in person at clinics at the University of California San Francisco. We included participants who completed at least one IPAQ and opted before June 2019 to share data from commercially acquired Fitbits with Health eHeart. Participants were excluded only for missing or invalid data. Health eHeart was approved by the University of California San Francisco Committee on Human Research. Consent was obtained electronically on enrollment. We obtained deidentified data for this study.

Self-reported activity was assessed with the IPAQ short form, which assesses activity at discrete intensities (sedentary, walking, moderate activity, or vigorous activity) over 7 consecutive days [11,12]. Data were cleaned according to the IPAQ protocol with additional criteria developed to address erroneous values (Multimedia Appendix 1). Total active time (sum of time spent walking, moderately active, and vigorously active) and overall activity in weekly MET hours were computed [12].

Physical activity was measured objectively on the same 7 days with Fitbit-brand wearable activity trackers, which are triaxial accelerometers using proprietary algorithms to convert accelerometer counts to minute-by-minute step counts. The Fitbit devices that contributed study data were owned by study participants who obtained their devices outside of the Health eHeart study; the study did not request that participants acquire a Fitbit for this study and did not exclude specific Fitbit device types. Participants received no specific instructions on Fitbit use from Health eHeart. No specific devices were excluded and all were Fitbit brand (Tables S1 and S2 in Multimedia Appendix 1). We defined nonwear time as ≥120 contiguous minutes with 0 steps counted; all other time was considered wear time [27]. We recoded minutes with more than 200 steps as missing, as described previously in accelerometer studies [28,29]. Fitbit data were considered adequate for a given day when there were at least 10 hours of Fitbit wear time during that day. An activity observation was defined as a period of 7 days recalled by a participant on the IPAQ for which the same 7 days had adequate Fitbit data.

We then computed a series of Fitbit-derived physical activity estimates corresponding to standard estimates derived from the IPAQ [30]. Fitbit overall activity was defined as a within-participant average of steps/day during the week. We calculated time spent walking (10-100 steps/min, eg, slow walking), moderately active (101-130 steps/min, eg, brisk walking), vigorously active (131 to 200 steps/min, eg, vigorous walking, jogging, running), and sedentary time, defined as bouts shorter than 120 consecutive minutes with 0-9 steps, to allow for incidental movement [31] and typical sedentary bout lengths [32]. Minutes with at least 10 steps counted were considered active time.

This intensity-defined approach was designed to reflect energy expenditure classifications for the IPAQ [31,33,34]. In a review of literature correlating step rates and METs expended in adults, 64-96 steps/min ranged from 2.0 to 3.1 METs (5 studies), 102-129 steps/min ranged from 2.9 to 5.5 METs (8 studies), and 134 to 170 steps/min ranged from 6.8 to 13.0 METs (7 studies) [35], which are closely aligned with IPAQ definitions of walking (<3 METs), moderate activity (3-6 METs), and vigorous (>6 METs) activity [36,37].

Height, weight, and BMI were self-reported through Health eHeart. Values reported within 90 days of the activity observation were averaged for weight and BMI, and values within 365 days were averaged for height. When both were available, height and weight data were combined to generate a calculated BMI value associated with each activity observation. Calculated BMI values were merged with self-reported BMI values using the median when both were available for an activity observation (supplementary methods in Multimedia Appendix 1).

Demographic characteristics, health-related behaviors (ie, alcohol and smoking), and clinical characteristics (ie, history of coronary artery disease, hypertension, hyperlipidemia, and diabetes) were obtained through questionnaires. We used responses within 365 days of and closest to each activity observation.

Some participants contributed multiple observations over time (Figure S1 in Multimedia Appendix 1). Comparative analyses were restricted to the first activity observation (N=1498). We investigated agreement between IPAQ-measured and Fitbit-measured activity by using Wilcoxon matched-pairs signed-rank test and Spearman’s rank correlations on untransformed data. For the BMI analysis, we excluded participants lacking BMI or covariate data to generate a subsample of 697 observations from 586 participants (Multimedia Appendix 1). We compared normally distributed continuous variables by two-sided t test, nonnormal variables by Wilcoxon rank sum test and categorical variables by chi-squared test. Bivariate associations between activity and BMI were examined by Spearman’s rank correlation and nonparametric test-of-trend across heuristically defined physical activity categories at participant level restricted to the first activity observation. We next used multivariable regressions of BMI on Fitbit-measured and IPAQ-measured activity variables by using robust errors clustered by participants to account for activity outliers and nonindependence for participants with multiple observations. Activity distributions and BMI were not transformed. Twelve separate predictor models were fit—one for each of the 6 analogous activity variables measured by either Fitbit or IPAQ. Next, 6 head-to-head combined models were fit with both analogous measurements as predictors (eg, vigorous activity measured by IPAQ and Fitbit in the same model) to assess BMI association strength while controlling for the analogous measurement. All models were adjusted for potential confounding by Fitbit wear-time, device wear location (torso or wrist), and season of data collection and estimated in natural units and in standardized units to account for differences in self-reported and Fitbit-measured activity distributions. We then sequentially adjusted the models for demographic characteristics, health-related behaviors, and clinical characteristics.

We conducted 2 sensitivity analyses to investigate whether data sources affected the main findings of the study. The first examined whether the main findings were sensitive to the inclusion of participants whose data were supplied by MobileTrack (81/1392, 5.8% of comparative analysis cohort; 32/586, 5.5% of BMI analysis subset), which was derived from smartphone measurements and not a wearable activity tracker. The second examined whether the main findings were sensitive to the inclusion of participants whose activity trackers were torso-worn (397/1392, 28.5% of the comparative analysis cohort; 166/586, 28.3% of the BMI analysis subset). For both sensitivity analyses, we excluded the relevant cohort and repeated the study analysis.

Data preparation and analyses were performed using STATA 14 (StataCorp). Two-tailed P values less than .05 were considered significant. We report this study according to STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statement guidelines [38].

The age of the participants (N=1498) ranged from 19 to 92 years and they were predominantly Whites (1299/1498, 86.7%), females (895/1498, 59.7%), with at least a bachelor’s degree (1067/1415, 75.4%), and an annual income exceeding US $100,000 (695/1282, 54.2%) (Table 1). All characteristics in Table 1 are summarized at participant level at the first activity observation by using mean (SD) for normal continuous data, median (IQR) for nonnormal continuous, and not available for frequency data.

Physical activity observations were not evenly distributed among seasons (Wald test, P<.001) and 995 (71.5%) of 1392 Fitbit devices were wrist-worn (Table 2). In Table 2, all characteristics are summarized at participant level at the first activity observation by using mean (SD) for normal continuous data, median (IQR) for nonnormal continuous, and not available for frequency data. There were no missing data for the BMI analysis. The median overall activity at the first activity observation was 35.5 MET-h/wk (IQR 18.1-62.2) by IPAQ. Fitbits counted a median of 8622 steps/d (IQR 6191-11061) over the same week. Compared to Fitbit measurements, participants self-reported more moderate activity (median of 90 min/wk vs 75 min/wk, Wilcoxon matched-pairs sign-rank test P<.001) and vigorous activity (80 min/wk vs 0 min/wk, P<.001), but self-reported less time spent sedentary (42.0 h/wk vs 87.2 h/wk; P<.001) and walking (3.5 h/wk vs 20.2 h/wk, P<.001).

Overall activity measurements by Fitbit (steps/d) and IPAQ (MET-h/wk) were moderately correlated (Spearman’s rank r_s=0.48), as was total active time (r_s=0.40) and vigorous activity (r_s=0.29) (Figure 1, Figure S2 in Multimedia Appendix 1). By contrast, time spent sedentary, walking, or moderately active were weakly correlated (walking r_s=0.17, sedentary r_s=0.16, moderate activity r_s=0.09). Findings were very similar when we excluded participants whose data source was the MobileTrack app, and when we excluded participants whose data were obtained from an activity tracker worn on the torso.

The subset of participants with a BMI measurement was included in the BMI analysis sample (n=586). In bivariate analyses using categorical physical activity, Fitbit measurements of activity were significantly associated with BMI for all intensity-defined (Table 3) and overall activity measurements (Table 4) (P values range from <.001 to .01), whereas overall activity, total active time, sedentary time, and walking time measured by IPAQ were significantly associated with BMI (P values range from .001 to .01).

All Fitbit measurements, including those of lower intensity activities (Figure 2), higher intensity activities (Figure 3), and overall activity (Figure 4) were correlated with BMI (Spearman’s rank test, 6 comparisons, all P<.001). By contrast, IPAQ measurements of lower intensity activity (Figure 2) and overall activity (Figure 4) were correlated with BMI (P values range from .001 to .005) but not the higher intensity moderate activity (P=.26) or vigorous activity (P=.06) (Figure 3).

After adjustment for demographics, health-related behaviors, and clinical characteristics, an increase of 1 h/wk in Fitbit-measured vigorous activity was associated with a BMI that was 0.84 kg/m² lower (95% CI –1.35 to –0.32, P<.001) (Table 5). By contrast, an increase of 1 h/wk in self-reported vigorous activity on the IPAQ was associated with a BMI that was 0.17 kg/m² lower (95% CI –0.34 to –0.00, P=.045), which was significantly smaller than the coefficient for Fitbit-measured vigorous activity (P=.01).

In fully adjusted combined models with both analogous Fitbit and IPAQ measurements in the same model (eg, vigorous activity measured by Fitbit and IPAQ included as predictors in the same model), coefficients displayed a similar pattern of magnitude to the fully-adjusted single predictor models (Table 5). However, only Fitbit measurements were significantly associated with BMI in combined head-to-head models for moderate and vigorous activity, whereas the reverse was the case for sedentary time.

When using standardized units to account for differences in the typical ranges between self-reported and Fitbit-measured activity using fully-adjusted single predictor models, magnitudes of association with BMI were larger for Fitbit versus IPAQ-measured time spent walking, moderately active, or vigorously active, but these did not reach statistical significance (Table S5 in Multimedia Appendix 1).

In separate regressions, an increase in Fitbit-measured overall activity of 1 SD (3663 steps/d) was associated with a BMI that was 1.37 kg/m² lower after adjusting for demographic factors, health-related behaviors, and clinical characteristics (β=–1.37, 95% CI –1.84 to –0.90; P<.001), whereas 1 SD increase in IPAQ-measured overall activity (37.2 MET-h/wk) was associated with a BMI that was 0.72 kg/m² lower (β=–.72, 95% CI –1.14 to –0.29; P=.002) (Table 6, complete regression results in Table S3 in Multimedia Appendix 1). In Table 6, overall activity units were standardized due to their different scales (natural unit analyses of overall activity are presented in Table S4 in Multimedia Appendix 1). Findings were very similar when we excluded participants whose data source was the MobileTrack app and when we excluded participants whose data were obtained from an activity tracker worn on the torso.