# Physical activity and diabetes among US adults
## Research brief: NHANES 2011–2018
Analysis date: 28 September 2026.

**Adults reporting more recreational physical activity had lower diabetes prevalence. Survey weighting changed the descriptive picture, and adjustment for measured participant characteristics substantially reduced—but did not eliminate—the association. These cross-sectional data do not establish a preventive effect of activity.**

### What was studied
We combined fresh official public data from four NHANES cycles, 2011–2012 through 2017–2018. The eligible population comprised examined adults aged 20 or older, excluding known current pregnancy. NHANES represents the civilian noninstitutionalized population; the estimates describe this historical period, not present-day US adults.

Recreational activity was calculated as moderate minutes plus twice vigorous minutes per week, using reported days and minutes in a typical week. Categories were none, 1–149, and at least 150 moderate-equivalent minutes/week. Questions count bouts lasting at least 10 minutes: the highest category is not a complete measure of meeting current activity guidelines.

Diabetes was defined by a reported clinician diagnosis other than during pregnancy, current reported insulin or diabetic pills, or HbA1c at least 6.5%. A negative classification required a nonpositive diagnosis response and an observed HbA1c below 6.5%, with no contradictory medication evidence. Unknown values were not treated as negative. This identifies prevalent diabetes of any type, including biochemical evidence without a reported diagnosis; one HbA1c measurement is not a confirmed clinical diagnosis.

### Participants and missing data
Of 39,156 interviewed participants, 37,399 had positive examination weights and valid survey design information. There were 21,646 examined adults aged at least 20; excluding 247 known pregnancies left 21,399 eligible adults. Activity and diabetes were observed for 20,484. All five primary models used the same **18,241 adults**, including 3,310 classified with diabetes.

The common cohort retained 85.2% of eligible participants and 88.0% of their examination-weighted population. Income was missing for 2,120 eligible adults, diabetes classification for 884, BMI for 306, activity for 31, education for 24 and smoking for 19; these categories overlap. No imputation was used.

Excluded eligible adults had higher weighted diabetes prevalence among those with observed diabetes data (15.9% versus 13.5%) and less reported high recreational activity among those with observed activity (34.4% versus 38.8%). Their weighted mean age was 48.6 versus 47.9 years. Consequently, complete-case weighting does not by itself restore full national representativeness. Unknown pregnancy status was retained, including 172 participants explicitly coded as uncertain.

### How the picture changes
The table compares at least 150 minutes/week with no reported recreational activity. Each adjusted difference comes from predicting diabetes prevalence at each activity category over the same cohort's covariate distribution. Negative differences indicate lower prevalence in the more active category.

| Analysis | None: prevalence | >=150: prevalence | Difference, percentage points (95% CI) | Odds ratio (95% CI) |
|---|---:|---:|---:|---:|
| Unweighted crude | 22.71% | 11.68% | -11.03 (-12.19, -9.86) | 0.45 (0.41, 0.49) |
| Survey-weighted crude | 18.49% | 8.11% | -10.38 (-11.76, -8.99) | 0.39 (0.34, 0.45) |
| + Demographics and cycle | 16.30% | 9.71% | -6.59 (-8.01, -5.18) | 0.52 (0.45, 0.61) |
| + Education, income and smoking | 15.63% | 10.30% | -5.33 (-6.78, -3.88) | 0.59 (0.51, 0.69) |
| + BMI | 14.52% | 11.79% | -2.73 (-4.04, -1.41) | 0.75 (0.65, 0.86) |

Demographic adjustment includes age, sex, race/ethnicity and survey cycle. Age and BMI use prespecified nonlinear splines. Race/ethnicity is a social/demographic covariate, not an assumed biological cause. The last two rows add variables cumulatively. Adjusted prevalences are model-standardized summaries, not observed category-specific prevalences or predictions of an intervention.

**Survey design matters.** NHANES oversamples some population groups. Weighting lowered the estimated diabetes prevalence in each activity category. Strata and clusters also change uncertainty; weights alone do not make ordinary independent-observation standard errors valid. The log-odds-ratio SE for the high-versus-none contrast was 0.0466 unweighted and 0.0725 in the design-aware crude model. This comparison combines weighting and design effects, so it does not isolate clustering alone.

**Participant characteristics matter.** The weighted mean ages were 51.0 years in the no-activity category and 44.2 years in the highest category; mean BMI was 30.4 versus 28.0. Accounting for measured differences reduced the standardized prevalence gap substantially.

**BMI needs separate interpretation.** BMI may confound the association, may be influenced by physical activity, and may itself change following diabetes or its treatment. The BMI-adjusted row answers a different conditional comparison. Its attenuation cannot quantify the fraction of a causal effect mediated by BMI.

For 1–149 minutes/week, the socioeconomic/smoking-adjusted prevalence difference versus none was -2.90 points (95% CI -5.03 to -0.77). With BMI added it was -1.60 points (-3.70 to 0.49), so that interval included no difference. The data do not establish a sharp threshold at 150 minutes.

### Planned robustness checks
- **Broader crude sample:** among 20,484 adults with observed activity and diabetes, weighted prevalence was 18.57%, 12.67% and 8.27% across the three activity categories. The corresponding common-cohort values were 18.49%, 12.39% and 8.11%. Crude patterns were similar, although this does not rule out selection bias in adjusted associations.
- **Diagnosis-only diabetes:** on 18,240 matched participants, the socioeconomic/smoking-adjusted high-versus-none gap was -4.24 points (-5.47 to -3.01), versus -5.34 (-6.79 to -3.89) using the primary definition on those same participants.
- **Total reported activity:** including work and transport activity, on 18,179 matched participants, the corresponding gap was -6.35 points (-7.77 to -4.93), versus -5.30 (-6.75 to -3.85) for recreational activity. Total and recreational activity define different exposure groups; this does not rank causal benefits by activity domain.

Both alternative definitions retained an inverse association. These were prespecified descriptive robustness checks, not independent confirmatory studies.

### What the data can and cannot tell us
The analysis supports a reproducible population-survey association and demonstrates that its size depends on weighting, participant composition, adjustment and measurement choices.

It cannot determine whether activity preceded diabetes. Diabetes symptoms, disability, advice after diagnosis and treatment can change activity, creating reverse causation. Residual confounding may remain from diet, health status, medication use and other unmeasured or imperfectly measured factors. Income and older age are top-coded; adjustment models are approximations. Activity is self-reported, excludes short bouts and may be misclassified. HbA1c alone misses some diabetes detected by other clinical tests, and current medications may have other indications. Diabetes types are not separated.

Survey weights address sampling and certain interview/examination nonresponse, not every missing questionnaire or laboratory value. Complete-case selection may bias associations. Excluding known pregnancy does not eliminate uncertainty about all pregnancies. Laboratory documentation reports a testing-site change in 2013–2014 and a minor secondary-method change in 2017–2018; original released HbA1c values were used, with cycle adjustment.

Odds ratios are not prevalence ratios, especially with a common outcome. They are also noncollapsible: changes after covariate adjustment cannot all be assigned to confounding. Standardized prevalence differences help describe absolute association size but are still not causal effects. Confidence intervals quantify sampling/model uncertainty under these assumptions, not the full uncertainty from measurement and selection bias.

### Methods and verification
We used examination weights WTMEC2YR/4, 59 cycle-specific strata and 121 masked PSUs, retaining the full examination design before adult-domain subsetting. There were 62 design degrees of freedom. Fifty-six strata had two PSUs and three had three; this corrected an overly restrictive two-PSU expectation before model estimation and is documented separately.

Survey logistic models used Taylor coefficient covariance. Standardized prevalence and differences used 121 stratified delete-one-PSU jackknife replicates, refitting and restandardizing in each replicate. Survey intervals use model residual design degrees of freedom; prevalence intervals are transformed from the logit scale. Unweighted intervals use ordinary model covariance and conditional delta-method standardization. No multiplicity-adjusted confirmatory testing was intended.

All models converged with full rank. No activity-by-outcome cells crossed with the categorical adjustment variables had fewer than 10 participants. Weight rescaling left estimates and SEs unchanged. Crude standardized prevalence matched direct survey estimates, and crude jackknife SEs agreed with Taylor SEs within 0.1%. The delivered script was rerun through native Notebook on retained inputs; cohort flow, primary odds ratios, primary standardized results and sensitivity standardized results matched within 1e-10. This is an analysis rerun, not a claim of using a separate native Reproduction feature.

### Sources and files
Data and codebooks: CDC NHANES, [2011–2012](https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?BeginYear=2011), [2013–2014](https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?BeginYear=2013), [2015–2016](https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?BeginYear=2015), [2017–2018](https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?BeginYear=2017); DEMO, PAQ, DIQ, GHB, BMX and SMQ components.
Methods: [CDC NHANES weighting tutorial](https://wwwn.cdc.gov/nchs/nhanes/tutorials/weighting.aspx), [physical activity codebook](https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2017/DataFiles/PAQ_J.htm), [diabetes codebook](https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2017/DataFiles/DIQ_J.htm), [HbA1c codebook](https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/2017/DataFiles/GHB_J.htm). All accessed 28 September 2026.

The reproducibility archive contains the 24 original XPTs, retained codebooks, source hashes, full numeric tables, frozen specification and clarification, R script, figure, runtime information and rerun checks.

Figure description: the left panel shows lower diabetes prevalence after weighting in all three activity categories; the right panel shows the high-versus-none prevalence difference moving toward zero as adjustment is added. All displayed intervals are 95% confidence intervals. The CSV tables provide a text-accessible alternative.
