# Underhydration–mortality replication: reproduction gate audit

**Project:** Does NHANES-Defined Underhydration Predict Long-Term Mortality?
**Audit version:** 2.0
**Date:** 2026-09-29
**Status:** Reproduction gate passed; 2019 target outcome had not been inspected when this audit was frozen.

## Purpose

Before inspecting the prespecified 2019 mortality extension, reconstruct the published NHANES 2009–2012 cohort, hydration classification, mortality event counts and weighted Cox results reported by Stookey et al. (2020).

## Cohort reproduction

| Gate | Published benchmark | Reproduced | Status |
|---|---:|---:|---|
| Initial age 51–70 source population | 3,077 | 3,077 | Exact |
| After creatinine + albuminuria restrictions | 2,578 | 2,578 | Exact |
| After ≥8 h fasting + positive fasting weight | 1,243 | 1,243 | Exact |
| After complete chronic-condition information | 1,202 | 1,202 | Exact |
| Mortality-eligible final cohort | 1,200 | 1,200 | Exact |
| Female | 609 | 609 | Exact |
| Male | 591 | 591 | Exact |
| Euhydrated | 319 | 319 | Exact |
| Hyponatremic | 22 | 22 | Exact |
| Underhydrated | 859 | 859 | Exact |

Cycle-level final cohort:
- NHANES 2009–2010: 660 participants.
- NHANES 2011–2012: 540 participants.
- Total: 1,200 participants.

## Mortality-event reproduction

Using the recovered 2015-vintage public-use Linked Mortality Files:

| Group | N | All-cause deaths | Chronic-disease-associated deaths |
|---|---:|---:|---:|
| Euhydrated | 319 | 11 | 5 |
| Hyponatremic | 22 | 2 | 0 |
| Underhydrated | 859 | 39 | 28 |
| **Total** | **1,200** | **52** | **33** |

These totals exactly reproduce the source publication.

The chronic-disease mortality definition used the public-use mortality fields:
- underlying heart disease or malignant neoplasm, or
- diabetes-associated death flag, or
- hypertension-associated death flag.

## Weighted Cox point-estimate reproduction

The primary two-category contrast excludes the 22 hyponatremic participants and compares underhydrated with euhydrated adults.

Pooled fasting subsample weights were `0.5 × WTSAF2YR`.

| Outcome/model | Published HR | Reproduced HR | Agreement |
|---|---:|---:|---|
| All-cause, unadjusted | 2.37 | 2.373 | Exact to published rounding |
| Chronic-disease mortality, unadjusted | 4.44 | 4.440 | Exact to published rounding |
| All-cause, full multivariable Model 2 | about 2.33 | 2.290 | Close; ~1.7% lower |
| Chronic-disease mortality, full multivariable Model 2 | 4.21 | 4.186 | Close; <1% lower |

The full Model-2 reconstruction used the published adjustment domains: age, sex, height, race/ethnicity, education, poverty, physical activity, total energy, carbohydrate share of energy, alcohol, dietary solute load, smoking, prescription medication, examination season, and fasting duration.

The small residual difference in the fully adjusted point estimates is consistent with minor unreleased coding details or categorical handling. It is not large enough to alter the reproduction conclusion because:
1. the complete cohort flow is exact;
2. the hydration-group counts are exact;
3. the mortality-event counts are exact;
4. the unadjusted weighted Cox estimates reproduce to published rounding;
5. the fully adjusted exposure coefficients remain within approximately 0.6–1.7% of the published values.

No attempt was made to tune covariate coding after inspecting 2019 outcomes.

## Survey-variance limitation

The reproduction above validates the Cox coefficient point estimates. Full NHANES strata/PSU design-based standard errors and confidence intervals have not yet been recreated in this audit implementation. They will be used for the final manuscript-quality analysis.

This does not affect whether the exposure coefficient itself reproduces.

## Minor published-table inconsistencies noticed during reconstruction

Several descriptive counts in the published Table 1 appear internally inconsistent with the total N or with the raw NHANES variables. For example, the reported poverty-category counts sum to more than 1,200, and the reported two fasting-duration counts (319 and 881) exactly duplicate the hydration-group split rather than the distribution obtained from `PHAFSTHR`.

These observations are treated as possible table/coding discrepancies, not as evidence of scientific error. The analysis uses the raw NHANES variables and the published Methods definitions.

## Gate decision

**PASSED.**

The 2015 result has been independently reconstructed closely enough to proceed to the prespecified 2019 follow-up extension without changing:
- the cohort definition,
- the hydration exposure,
- the primary outcome,
- or the interpretation rules frozen in protocol v1.

At the time this audit was frozen, the target-cohort 2019 mortality result had not been inspected.