Methods & reproducibility
Every number, and how to reproduce it.
This is the technical record behind Counted Health for reviewers, statisticians, and anyone who wants to check our work. Exact datasets and variables, how the burden score is built, the validation models with confidence intervals, what the evidence does and does not support, and the scripts that regenerate every figure. Everything is U.S. federal open data.
1. Data sources
No private, proprietary, or personally identifying data is used. Every input is a free, public federal file.
| Federal source | Role |
|---|---|
| CDC NHANES 2013–2014 | Source cycle for every respondent in this atlas. Variable files combined: DEMO_H (demographics), CBC_H (complete blood count), BIOPRO_H (biochemistry profile), GHB_H (glycohemoglobin), PFQ_H (physical functioning), DPQ_H (depression screener), HSQ_H (current health status), SLQ_H (sleep disorders), PAXDAY_H (physical activity monitor), and MCQ_H (medical conditions). |
| CDC — National Health and Nutrition Examination Survey (NHANES) | Program landing page: survey design, purpose, and how NHANES data collection works. |
| NCHS — NHANES Analytic Guidelines | Official guidance on survey weighting and variance estimation; the basis for the survey-weighted (chronicWt) rates reported in this brief. |
| NCHS — Public-Use Linked Mortality Files (through 2019) | Independent death-certificate linkage (National Death Index) used only to VALIDATE the burden strata — never to build the score. Basis for the Cox mortality analysis (NHANES 2013–2014 respondents followed through December 2019). |
2. Cohort & survey design
Adults from the NHANES 2013–2014 cycle, complete-case on the modeled variables (no imputation — median-imputing high-missing features produced artifact groups, so we excluded them). Analytic n differs slightly by analysis because each uses complete cases on its own variables:
- ~3,919 — the burden Atlas (physiology + activity + symptom features).
- 3,898— tri-modal association & sensitivity set (self-reported chronic condition (MCQ160-series + MCQ010)).
- 3,915 — mortality-eligible subset (public-use linkage), 285 deaths, median follow-up 71 months.
NHANES is a complex, multistage probability sample. Rates are survey-weighted (MEC exam weight WTMEC2YR) with variance estimated using the design strata and PSUs (SDMVSTRA × SDMVPSU), per the NCHS Analytic Guidelines.
3. The three modalities
Three kinds of measurement, taken on the same person, that normally live in separate files:
| Modality | NHANES files | Signals |
|---|---|---|
| Physiology | CBC_H, BIOPRO_H, GHB_H | Complete blood count, biochemistry, HbA1c, and a neutrophil-to-lymphocyte ratio (NLR) inflammation proxy. |
| Behavior | PAXDAY_H | ~1 week of wrist-worn accelerometry, summarized to day-level activity intensity and day-to-day irregularity. |
| Experience | DPQ_H, PFQ_H, HSQ_H, SLQ_H | Validated instruments: PHQ-9 depression, physical functioning / fatigue, self-rated health, and sleep. |
4. Burden score & the five regions
Each feature is standardized (z-scored) across the cohort, then combined into a single continuous impairment score per person. People are ranked by that score and split into five equal groups (quintiles) — the five burden regions, Region 1 (lowest) to Region 5 (highest). Two deliberate constraints: the scoring rule is fixed a priori (it cannot be tuned to a flattering answer), and the regions are strata, not clusters — HDBSCAN labelled 77–100% of respondents as noise and did not track burden, because the burdened sit in a diffuse tail rather than a dense blob. Binary items were excluded (they dominate z-scored Euclidean distance).
5. Primary association
Diagnosed chronic-condition prevalence rises monotonically across the regions: 33% → 72% survey-weighted (0.326, 0.408, 0.487, 0.575, 0.719). Top-vs-bottom, the adjusted odds ratio is 2.45 (95% CI 1.92–3.12) via a survey-weighted logistic, cluster-robust SE by PSU (SDMVSTRA×SDMVPSU), adjusting for age, sex, race, and BMI. Adjustment shrinks the association from the unadjusted 5.23 — it never crosses 1 — so the gradient is not an artifact of confounding by those factors.
6. Mortality validation (non-circular)
The hardest test: link the same respondents to NCHS death-certificate records — an endpoint the burden score never used. Model: Cox proportional hazards, adjusted for age, sex, and BMI.
| Region | n | Deaths | Crude mortality | Mean age |
|---|---|---|---|---|
| 1 · lowest | 782 | 12 | 1.5% | 42 |
| 2 | 783 | 27 | 3.4% | 47 |
| 3 | 783 | 34 | 4.3% | 49 |
| 4 | 784 | 73 | 9.3% | 54 |
| 5 · highest | 783 | 139 | 17.8% | 57 |
- Region 5 vs Region 1: HR 5.01 (95% CI 2.72–9.23, p < 0.001); Region 4 vs 1 HR 2.82.
- Clean dose-response: trend HR 1.55 per region (p < 0.001); continuous severity HR 1.59 per SD.
- An objective-only score (labs + activity, no self-report) predicts death at HR 1.53 per SD (p < 0.001) — so the signal is not merely self-report, and not merely age (age contributes HR ≈ 3.4 per SD alongside it).
7. Sensitivity analyses
| Adjustment set | Top-vs-bottom OR (95% CI) |
|---|---|
| Fully adjusted (age+sex+race+BMI) | 2.45 (1.92–3.12) |
| Age dropped | 4.00 (3.24–4.93) |
| BMI dropped | 3.18 (2.50–4.04) |
| Age & BMI dropped | 4.87 (3.90–6.09) |
| Unadjusted | 5.23 (4.11–6.65) |
- Split-sample: the trend holds out of sample — train trend OR 1.23 vs test 1.23 (fixed seed 20260705).
- E-value: 4.33 (CI limit 3.25) — an unmeasured confounder would need associations this strong with both burden and outcome to explain away the result.
- Incremental value: the three modalities fused reach AUC 0.693 — higher than the best single modality (labs only, 0.653) — evidence that combining the signals adds discrimination, not just noise.
8. What this validates — and what it does not
It does not validate, and we do not claim:
- Any diagnosis for an individual — the regions are population strata, never disease labels.
- Causation — high burden goes with earlier death; it does not prove cause.
- The Phase-2 cost model — mortality validates the severity strata, not the dollar figures attached to them.
- Long COVID or Lyme specifically — the cycle predates COVID, so this is a fatigue / functional-impairment phenotype.
One candor point we surface rather than hide: after adjusting for age, sex, race, and BMI, the objective signals do not independently predict the self-reported chronic-condition label (that label is noisy) — yet the same objective score strongly predicts death. We report both directions and never claim objective-signal independence for the self-report outcome.
9. Limitations
- Mortality is an objective endpoint never used to build the burden score, so this is not circular.
- Cox models are adjusted for age, sex, and BMI; survey-design weighting is a Phase 2 refinement.
- Association, not causation: high burden predicts higher mortality risk; it does not prove cause.
- Age is a strong independent co-driver (adjOR ≈2.19/SD); the claim is 'independently associated,' not 'independent of age.'
- Outcome is self-reported.
- Pre-COVID cycle (2013–2014): a fatigue/functional-impairment phenotype, not Long COVID.
- Design-aware approximation (var_weights + cluster-robust SE), not a full survey package.
10. Reproduce every number
The pipeline is deterministic. Public NHANES and NCHS files go in; committed scripts emit the exact JSON this page and the whole site read — no hand-entered statistics.