A Strong Short-Term Mortality Association That Attenuated in a Later Mortality Linkage

A prespecified replication and reanalysis of a previously reported NHANES mortality association.

Publication type: Research Report
Version: 1.0
Date: 29 September 2026
Institution: The Null Institute
Author: Ilya Nikitin, Ph.D., Founder and Director, The Null Institute
Status: Independent reanalysis; not peer reviewed

Abstract

Background. A 2020 analysis of NHANES 2009–2012 reported a large association between a composite classification of “underhydration” and mortality during approximately 3–6 years of follow-up. Underhydration was associated with an adjusted all-cause mortality hazard ratio (HR) of 2.33 and an adjusted chronic-disease-associated mortality HR of 4.21. The analysis contained 52 all-cause and 33 chronic-disease-associated deaths.

Objective. To reproduce the original association using the historical mortality release and then test whether its magnitude persisted in the later 2019-vintage NCHS Linked Mortality Files, using an analysis protocol frozen before inspection of the later outcome.

Methods. We reconstructed the original NHANES cohort before examining the 2019-vintage mortality result. The reconstruction exactly reproduced the published sample size and hydration categories: 1,200 participants, including 319 classified as euhydrated, 22 as hyponatremic, and 859 as underhydrated. Historical 2015-vintage mortality data reproduced all 52 all-cause and 33 chronic-disease-associated deaths. Unadjusted Cox estimates reproduced to published rounding and fully adjusted point estimates differed by less than approximately 2%.

Before inspection of the target 2019-vintage result, we froze the cohort, exposure, primary outcome, covariates and interpretation criteria. Prespecified analyses included an age-and-sex model, a mortality-vintage transition analysis among participants classified alive in 2015, component decomposition, renal and diuretic sensitivities, flexible urine-osmolality modelling, cycle-specific analyses and proportional-hazards diagnostics.

Results. In the 2019-vintage analysis, the minimally adjusted all-cause estimate was approximately HR 1.03. The fully adjusted Model 2 estimate was HR 1.095 (95% CI 0.61–1.96), compared with approximately 2.3 in the historical short-follow-up analysis. The chronic-disease-associated estimate similarly attenuated from approximately 4.2 to approximately 1.64 (95% CI 0.68–3.95).

Among 1,128 participants in the primary euhydrated-versus-underhydrated contrast who were classified alive in the 2015 vintage, 65 were classified deceased in the 2019 vintage. The fully adjusted mortality-vintage transition estimate was approximately OR 0.75 (95% CI approximately 0.32–1.77).

The two NHANES cycles did not show a common positive direction: age-and-sex estimates were approximately HR 1.30 (0.65–2.59) for 2009–2010 and 0.66 (0.24–1.82) for 2011–2012.

Sensitivity analyses did not reveal a hidden large positive association. Serum sodium ≥145 mmol/L occurred in only two participants. Excluding eGFR <60 mL/min/1.73 m² moved the age-and-sex HR from 1.032 to 0.923. Excluding 241 diuretic users yielded HR 1.035, and exclusion of 21 observations with missing urine osmolality yielded HR 1.001. Secondary flexible urine-osmolality and time-interaction diagnostics likewise did not reveal a clear hidden positive signal.

Conclusions. The large short-follow-up mortality association was reproducible in the historical public-use data, but its magnitude did not persist in the prespecified analysis using the later 2019 mortality linkage. Because the 2019 release both extended follow-up and introduced an enhanced linkage algorithm, the attenuation cannot be attributed uniquely to the passage of additional follow-up time. The results specifically weaken the proposition that this one-time NHANES composite classification provides a large, stable long-term mortality signal.

Introduction

In 2020, an analysis of NHANES data reported an unusually large association between a laboratory-based classification of hydration status and subsequent mortality among U.S. adults aged 51–70 years.

The classification combined three measurements: serum sodium, the volume of a first spot urine specimen and spot urine osmolality. Participants with serum sodium from 135 to <145 mmol/L, urine volume ≥50 mL and urine osmolality <500 mOsm/kg were classified as euhydrated. Participants failing one or more criteria, except those classified separately as hyponatremic, were considered underhydrated.

During approximately 3–6 years of mortality follow-up, underhydration was associated with an adjusted all-cause mortality HR of 2.33 (95% CI 0.97–5.60). For chronic-disease-associated mortality, the adjusted HR was 4.21 (1.29–13.78).

Those are large point estimates. They were also based on only 52 all-cause and 33 chronic-disease-associated deaths. The original study explicitly identified short follow-up, limited event counts, observational design, residual confounding and the absence of serial hydration measurements as limitations.

A later release of the NCHS Linked Mortality Files created an opportunity for a stronger test. But the test is not as simple as “wait four more years.”

The 2019 public-use mortality release extends follow-up through December 31, 2019, but also uses an enhanced probabilistic linkage algorithm and supersedes the earlier mortality files. NCHS reports that vital status remained unchanged for at least 98% of participants when equivalent follow-up intervals were compared, but acknowledges that a small number of classifications changed between releases.

The later mortality vintage therefore contains two changes: additional mortality follow-up and revised mortality linkage.

Our question was consequently framed narrowly:

If the original cohort and exposure are first reproduced, and the analysis rules are frozen before inspecting the later mortality outcome, does the large original prognostic association remain visible in the 2019-vintage data?

Methods

Study population

We used NHANES 2009–2010 and 2011–2012.

The target cohort consisted of adults aged 51–70 years meeting the source study’s restrictions, including valid baseline dietary and examination data, serum creatinine from 0.5 to 1.5 mg/dL, urinary albumin-to-creatinine ratio <30 mg/g, at least eight hours of fasting, and sufficient information for the source chronic-condition definitions.

Cohort reconstruction was completed before inspection of the target 2019 mortality result.

The final cohort included 609 women and 591 men.

Hydration classification

Euhydrated: serum sodium ≥135 and <145 mmol/L, first spot urine volume ≥50 mL, and urine osmolality <500 mOsm/kg.

Hyponatremic: serum sodium <135 mmol/L.

Underhydrated: not euhydrated, not hyponatremic, and at least one of serum sodium ≥145 mmol/L, first spot urine volume <50 mL or inability to void under the source definition, or urine osmolality ≥500 mOsm/kg.

The reconstructed groups exactly matched the source study:

• euhydrated: 319;
• hyponatremic: 22;
• underhydrated: 859.

The primary contrast excluded the 22 hyponatremic participants.

Mortality vintages

The reproduction stage used historical NCHS public-use mortality files with follow-up through December 31, 2015.

The prespecified later analysis used the 2019-vintage NCHS Public-Use Linked Mortality Files.

The latter require two interpretive qualifications. First, for disclosure protection, synthetic values can replace follow-up time or underlying cause of death for selected records. Vital status itself is not altered by this disclosure-protection procedure.

Second, the 2019 files were produced using an enhanced linkage algorithm and supersede the earlier files. Consequently, small numbers of participants can differ in vital-status or cause-of-death classification across mortality vintages.

We therefore treat final all-cause vital status as more robust than fine-grained cause-specific or exact survival-time comparisons, and we do not interpret every difference between vintages as necessarily representing a newly occurring death.

Reproduction gate

The historical mortality data exactly reproduced:

The unadjusted point estimates reproduced to published rounding:

• all-cause: published 2.37, reproduced 2.373;
• chronic-disease-associated: published 4.44, reproduced 4.440.

The reconstructed full Model 2 point estimates were similarly close:

• all-cause: approximately 2.29 versus published 2.33;
• chronic-disease-associated: approximately 4.19 versus published 4.21.

The reproduction gate was therefore considered passed.

Frozen analysis

Before inspection of the target 2019-vintage outcome, the following were fixed:

• cohort definition;
• exposure definition;
• primary euhydrated-versus-underhydrated contrast;
• all-cause primary outcome;
• principal Model 2 covariate set;
• minimal age-and-sex model;
• mortality-vintage transition analysis;
• component decomposition;
• renal-function sensitivity;
• diuretic sensitivity;
• flexible urine-osmolality analysis;
• cycle-specific estimates;
• proportional-hazards diagnostics;
• interpretation rules.

The primary model used pooled fasting subsample weights, 0.5 × WTSAF2YR, together with the NHANES survey strata and PSU variables.

Most importantly, statistical significance was not specified as the sole decision rule. A large persistent association required the later HR to remain materially above 1 and to be reasonably coherent across supporting analyses. Movement toward 1, lack of a later positive signal, or failure of the constituent components to reproduce the pattern were prespecified evidence against a stable large prognostic association.

Results

Historical association reproduced

The historical finding was reproducible. The cohort size, hydration groups and mortality-event totals were reconstructed exactly. The unadjusted Cox coefficients reproduced to rounding, and the multivariable coefficients were very close to those originally published.

The subsequent attenuation therefore cannot reasonably be attributed simply to failure to reconstruct the source analysis.

2019-vintage primary analysis

In the later mortality vintage, the large association was no longer present.

The age-and-sex-adjusted point estimate was approximately HR 1.032.

The fully adjusted Model 2 estimate was HR 1.095, 95% CI 0.61–1.96.

The key change is therefore in effect magnitude, not merely in whether a conventional p-value lies above or below 0.05. The historical all-cause estimate was approximately 2.3. The later estimate was approximately 1.1.

The confidence interval remains wide enough that the data do not establish an exact null effect. Both potentially meaningful reductions and increases in risk remain compatible with the estimate. What is no longer supported is a stable approximately twofold mortality association.

Chronic-disease-associated mortality

The same attenuation occurred for the secondary chronic-disease outcome.

The historical adjusted estimate was approximately 4.2. In the 2019-vintage analysis it was approximately HR 1.64, 95% CI 0.68–3.95.

Because public-use cause-of-death information can contain synthetic substitutions and because cause classification can differ across linkage vintages, this result receives less inferential weight than the all-cause endpoint. Nevertheless, it does not provide independent support for persistence of the original large association.

Mortality-vintage transition analysis

The prespecified “late-death” analysis is more precisely described as a mortality-vintage transition analysis.

Within the primary euhydrated-versus-underhydrated contrast, 1,128 participants were classified alive in the 2015 mortality vintage. Of these, 65 were classified deceased in the 2019 vintage.

The approximate estimates were:

• unadjusted OR: 0.85;
• age-and-sex OR: 0.73;
• Model 2 OR: 0.75, with 95% CI approximately 0.32–1.77.

These transitions therefore did not preserve the positive direction of the original short-follow-up association.

They should not be interpreted as a pure cohort of deaths occurring after January 1, 2016, because the mortality linkage itself changed between releases. Their role is narrower: they provide a prespecified falsification test showing that the participants newly classified as deceased in the later mortality vintage were not disproportionately drawn from the baseline underhydrated group.

Cycle consistency

The two survey cycles also did not provide a common positive signal.

• 2009–2010: HR 1.30 (95% CI 0.65–2.59);
• 2011–2012: HR 0.66 (95% CI 0.24–1.82).

The individual estimates are imprecise, but their opposite directions argue against interpreting the pooled estimate as a consistently replicated positive association across both NHANES cohorts.

Component decomposition

The composite classification combined three different physiological measurements. The components did not reveal a concealed large positive association.

Serum sodium ≥145 mmol/L occurred in only two participants out of 1,200, with one death. It is therefore essentially non-informative as an independent exposure in this sample.

The low-urine-volume and urine-osmolality ≥500 mOsm/kg components likewise produced estimates near or below 1 rather than recreating the large historical composite effect.

Renal sensitivity

Fifty participants had eGFR <60 mL/min/1.73 m². After excluding them, the age-and-sex estimate changed from 1.032 → 0.923. Thus impaired renal function was not masking a large positive association.

A separate albuminuria exclusion would add little information because the original reconstructed cohort already excluded UACR ≥30 mg/g.

Diuretics

There were 241 diuretic users. After excluding them, the estimated HR was 1.035.

Missing urine osmolality

Excluding 21 observations with missing urine-osmolality information in the relevant sensitivity analysis produced HR 1.001.

Continuous urine osmolality

A binary 500 mOsm/kg threshold could theoretically obscure a nonlinear association. A prespecified flexible continuous analysis showed no clear evidence of such a hidden pattern.

The secondary spline diagnostic yielded an LR statistic of approximately 2.27 on 3 df, corresponding to approximately p=0.52.

This diagnostic is supportive rather than decisive. Its complex-survey inferential implementation has not yet undergone the same independent technical audit as the primary Cox estimate and should not be treated as a primary hypothesis test.

Proportional hazards

Descriptively, the exposure coefficient appeared to decline with increasing follow-up time, with approximate HRs near 1.41 around 3 years, 1.06 around 5 years and 0.71 around 10 years.

However, the secondary time-interaction diagnostic was not statistically persuasive, with approximately p=0.25. We therefore do not claim a demonstrated violation of proportional hazards or a proven biological time-varying effect.

The simpler observation is sufficient: the large historical point estimate was not reproduced in the later mortality vintage.

Discussion

The study produced two results that initially appear uncomfortable together. The original association was reproducible. The later association was much smaller.

There is no contradiction between them. The historical result describes what was present in the earlier linked dataset during relatively short follow-up. The later result asks whether that magnitude remains when mortality information is updated and extended under a prespecified analysis. It did not.

What explains the attenuation?

The available data do not identify a unique mechanism. At least three possibilities remain.

1. Early-risk concentration

The baseline classification may have captured transient frailty, existing disease, medication effects, renal physiology or another short-horizon correlate of mortality rather than a stable long-term exposure. Under this explanation, early deaths carry much of the original association and later mortality dilutes it. The data are compatible with this mechanism but do not prove it.

2. Sampling variability

Only 52 deaths were present in the original analysis. The original confidence intervals already indicated considerable uncertainty. An adjusted all-cause HR of 2.33 had a 95% CI from 0.97 to 5.60, while the chronic-disease HR of 4.21 had a CI from 1.29 to 13.78.

The large early estimates may therefore partly reflect ordinary sampling variability in a relatively sparse-event dataset.

3. Mortality-linkage revision

The later public-use file is not merely the earlier file plus additional years. NCHS introduced an enhanced linkage algorithm and states that a small number of vital-status classifications changed relative to previous releases.

The present design therefore cannot decompose the attenuation precisely into additional follow-up versus revised mortality linkage.

This is why the primary scientific statement concerns lack of persistence across mortality vintages rather than claiming that longer biological follow-up alone caused the HR to collapse.

Why the composite matters

The label “underhydration” sounds physiologically unitary. The operational definition is not.

It combines serum sodium ≥145 mmol/L, spot urine volume <50 mL and spot urine osmolality ≥500 mOsm/kg. These measurements reflect related but non-identical physiology.

In the reconstructed cohort, one component, high serum sodium, was almost absent. The other components did not recreate a large positive mortality association when examined separately.

The data therefore caution against treating the composite category as equivalent to a directly measured chronic state of inadequate water intake.

Spot urine is not habitual water intake

A spot urine specimen reflects short-term physiological regulation and is affected by recent fluid and food intake, solute load, renal concentrating capacity, medication, environment and timing.

The analysis therefore evaluates a baseline prognostic classification. It does not directly estimate the causal effect of drinking additional water. It cannot establish that deliberate water supplementation would reduce mortality. Nor can it establish that chronic clinically significant dehydration is harmless.

What This Study Weakens

The results weaken the specific proposition:

A one-time NHANES composite classification of underhydration in adults aged 51–70 years identifies a large and stable long-term mortality hazard of approximately twofold or greater.

The historical data supported a large short-horizon association. The later mortality vintage did not.

The results do not establish the stronger proposition that hydration physiology has no influence on human health or mortality. That conclusion is outside the identification capacity of this analysis.

Nor does HR 1.095 (0.61–1.96) demonstrate that the true effect is exactly zero. The interval remains broad.

The useful conclusion lies between these extremes:

A large observational mortality association could be reproduced in the historical dataset, but its magnitude did not persist under a prespecified test using the later mortality linkage.

Strengths

The main methodological strength is separation between reproduction and testing.

Before inspecting the target 2019-vintage result, we reconstructed the cohort, hydration categories, historical mortality events and the original mortality association. The subsequent protocol froze the target exposure, outcome, covariates and interpretation rules.

Several potential rescue explanations were also specified before outcome inspection. This sharply limits the ability to redefine the study after seeing an inconvenient result.

A second strength is that the later result is not merely “non-significant.” The primary point estimate itself moved from approximately 2.3 to approximately 1.1.

Limitations

• This remains an observational analysis. It does not estimate the causal effect of increasing water intake.
• Hydration status was inferred from one baseline examination rather than repeated long-term measurement.
• The later all-cause confidence interval remains broad.
• Public-use follow-up times and causes of death can be synthetically substituted for disclosure protection in selected NCHS records, although vital status is not changed by that procedure.
• The 2019 mortality files use an enhanced linkage algorithm and therefore cannot be treated as a perfectly unchanged mortality ascertainment system with only additional calendar years appended.
• The exact consolidated 2019 event-count table by hydration group was not retained as a separate machine-readable artifact in the current publication archive.
• The primary survey-weighted estimates are more important than the secondary spline and proportional-hazards diagnostics. Those diagnostics should receive a dedicated survey-design technical check before journal submission.

Conclusion

The historical association between NHANES-defined underhydration and short-follow-up mortality was reproducible.

Using the later 2019 mortality linkage and an analysis protocol frozen before inspection of the target result, the association was much smaller.

The primary fully adjusted estimate was HR 1.095 (95% CI 0.61–1.96).

The minimally adjusted estimate was already close to 1. The mortality-vintage transition analysis did not retain a positive association, the two NHANES cycles did not show a common positive direction, the individual components did not reveal a hidden large signal, and renal, diuretic and missing-data sensitivities remained near null.

Because the mortality linkage methodology changed at the same time that follow-up was extended, the analysis cannot determine exactly how much of the attenuation reflects later deaths versus revised linkage.

The narrower conclusion is stronger because it requires fewer assumptions:

The large mortality association observed in the historical short-follow-up data did not behave as a stable prognostic association across the later mortality linkage.

Reproducibility and Audit Status

The project contains a frozen protocol dated September 29, 2026, created before inspection of the target 2019-vintage outcome. The historical reproduction gate was completed before the later result was examined.

The following analyses were prespecified:

• primary 2019-vintage analysis;
• minimal age-and-sex model;
• mortality-vintage transition analysis;
• component decomposition;
• renal-function sensitivity;
• continuous urine-osmolality analysis;
• cycle consistency;
• proportional-hazards diagnostics.

The archived analysis trail preserves the prespecified models and reported effect estimates, but a consolidated machine-readable table of 2019 event counts by exposure group was not retained.

Before submission to a peer-reviewed journal, the technical archive should additionally contain archived coefficients, design-based standard errors and confidence intervals for every reported Cox model; the exact strata/PSU/weight implementation; an independently rerun survey-design check of spline and time-interaction diagnostics; and executable analysis code with a versioned source manifest.

No additional outcome-driven hypothesis search is required.

References

1. Stookey JD, Kavouras SA, Suh H-G, Lang F. Underhydration Is Associated with Obesity, Chronic Diseases, and Death Within 3 to 6 Years in the U.S. Population Aged 51–70 Years. Nutrients. 2020;12(4):905. doi:10.3390/nu12040905.
2. National Center for Health Statistics. Public-Use Linked Mortality Files: File Description. Updated May 2022. Mortality follow-up through December 31, 2019.
3. National Center for Health Statistics. National Death Index Data Linkage: Public-Use Linked Mortality Files.
4. The Null Institute. Frozen Analysis Protocol: Does NHANES-Defined Underhydration Predict Long-Term Mortality? A Replication and Extended Follow-up of U.S. Adults Aged 51–70 Years. Version 1.0. September 29, 2026.
5. The Null Institute. Underhydration–Mortality Replication: Reproduction Gate Audit. Version 2.0. September 29, 2026.

Research provenance

The analysis was conducted as part of The Null Institute’s independent research program. AI tools assisted source recovery, data handling, checking and drafting under human direction. Responsibility for the published interpretation remains with the named author.