An independent reconstruction of NHANES water records raises a substantive problem with Zhou et al. (2022)
The Null Institute | Independent research audit | 17 October 2026
Scope: NHANES 1999–2014; source-data verification and an exploratory mortality sensitivity analysis. This report has not been peer-reviewed.
Research publications: Updated scientific preprint (XSci, 26 October 2026; DOI: 10.66977/xsci.2610.002u) · Original research archive: DOI: 10.5281/zenodo.23439456 · Zenodo research report and archived materials.
The XSci version is a publicly available preprint undergoing review, not a peer-certified journal article. The Zenodo DOI identifies the original historical audit.
The result
A 2022 study of 35,463 American adults linked higher water consumption, including water obtained from food, to lower mortality. Our independent reconstruction of the original NHANES files now points to a likely accounting error in how those water sources were combined.
The critical change occurred in 2005–2006. From that cycle onward, NHANES recorded plain drinking water inside the 24-hour dietary recall, and that water was already present in its total-moisture variable. Yet the published methods of Zhou et al. describe food water as total dietary moisture minus non-plain beverages, and total water as food water plus other beverages plus plain water. Unless an unreported subtraction was made, this procedure counts post-2005 drinking water twice: once as supposed food water and again as plain water.
This concern is not based on wording alone. The reconstruction that retains the extra water reproduces all four reported water-source medians strikingly closely, while removing the repeated water changes the estimates substantially. The same error mechanism also creates a sharp discontinuity at precisely the cycle when NHANES changed its collection method.
Our assessment: double-counting is the leading explanation for these findings. We cannot establish that it occurred in the authors’ actual code without that code or an exact reconstruction of their analytic sample. But the published exposure measures should no longer be treated as secure without this check.
The arithmetic trap
NHANES explicitly documents that beginning in 2005–2006, drinking-water reports became items in the dietary recall. The same documentation specifies that the moisture variables include water consumed as a beverage. A separate variable, DR1_320Z (or DR2_320Z), still reports the mass of plain water drunk that day.
Imagine a day with 500 g of moisture from food, 1,000 g from other beverages and 800 g from plain water. The dietary total is already 2,300 g. Adding plain water again gives 3,100 g. Worse, calculating “food water” as 2,300 minus the other beverages gives 1,300 g, although food contributed only 500 g. The example is illustrative; the underlying problem is an accounting identity, not a statistical assumption.
Before 2005, the separately recorded plain-water intake was not inside the food diary moisture total in the same way. A formula that was previously appropriate could therefore become wrong when the data structure changed. Zhou’s own methods mention this change but do not describe the subtraction needed to harmonise the two eras.
The published numbers match the vulnerable calculation
The original audit processed the eight NHANES cycles, 1999–2000 through 2013–2014, from 96 entries in supplied source archives. The reproduction workflow was subsequently rerun independently using 52 directly downloaded official CDC/NCHS public-use files. We then compared two definitions on exactly the same 35,163 participants, holding the definition of other beverages fixed. Six additional records were excluded from both calculations because the provisional beverage grouping yielded a clearly impossible negative residual after subtraction. All our displayed medians are unweighted sample medians, not population estimates.
| Median daily intake | Zhou et al., published | Our literal reconstruction | Removing the repeated plain water |
|---|---|---|---|
| Total water | 2,970 g | 2,965 g | 2,523 g |
| Water assigned to food | 924 g | 949 g | 556 g |
| Plain drinking water | 753 g | 748 g | 748 g |
| Other beverages | 1,032 g | 1,002 g | 1,002 g |
Zhou’s values are from their published Table 1 (N = 35,463). Our paired columns use the identical participants, the same rules for classifying non-plain beverages and a common upper-tail screen based on the literal reconstruction (N = 35,163). The authors’ own exact selection and classification remain unknown. Differences in total moisture and plain-water mass also involve nutrient-versus-weight rounding.
These are not four perfect matches, and several choices remain different from Zhou’s. Nevertheless, the correspondence is unusually specific. The suspect formula produces both the paper’s food-water level and its much larger total-water level. Correcting the same arithmetic makes the median apparent food water about 393 g/day lower and median total water about 442 g/day lower in our matched sample.
There is another internal check. From 2003 through 2014, we reconciled daily dietary moisture with the sum of item-level moisture for 101,153 matched person-days, to numerical precision; seven other records did not match. Separately, reconstructing official plain-water totals from three item codes (94000100, 94100100, 94300100), gram weights and combination-item flags yielded zero discrepancies above 1 g in 2005–2006, 2007–2008, 2011–2012 and 2013–2014. The 2007–2008 cycle was checked explicitly: 16,833 valid person-days, zero discrepancies. The initial 2009–2010 check flagged 47 records on each diary day. A subsequent independent audit resolved these discrepancies with the appropriate 2009–2010 water codes (94000100 and 94100100) and the applicable combination-item convention; the revised check found zero differences above 1 g among 17,911 eligible person-days across both diary days. See the dated supplementary validation. This distinction matters because water used to prepare a drink is not necessarily water drunk plain.
The discontinuity begins in 2005
Under the vulnerable calculation, median food-water share jumps from about 20–21% before 2005 to 41–46% thereafter. With drinking water removed once, the share stays within about 20–25% across all eight cycles.
The problem alters who is compared with whom. 98.3% of the highest food-water quartile under the literal calculation belongs to 2005–2014 survey cycles. Under the corrected calculation, that proportion falls to 67.7%, close to the post-2005 cycles’ 66.4% share of the screened sample. Among post-2005 participants, about 61.0% change food-water quartile when the accounting is corrected and quartiles are recalculated for each definition.
This makes the supposed “food-water” exposure partly a marker for which NHANES collection era a person entered. Survey cycle is not listed among the adjustment variables in Zhou’s published Cox-model descriptions. Even with cycle adjustment, double counting would remain a measurement error; without it, the calendar imbalance creates an additional interpretative problem.
What happens to the mortality association?
Zhou reported a food-water hazard ratio of 0.84 (95% CI 0.71–0.98) for all-cause mortality in the highest versus lowest intake quartile in Model 4. For total water, the corresponding estimate was 0.77 (0.66–0.90) in Model 3. These are associations, not evidence that eating wetter foods or drinking more water prevents death.
We re-estimated an exploratory food-water contrast with the same participants, covariates, non-plain beverage classification and mortality follow-up under the two accounting choices. Our models contained 31,046 people and 5,898 deaths, linked through 2019.
| Our all-cause mortality sensitivity model, Q4 vs Q1 | Hazard ratio (95% CI) |
|---|---|
| Literal food-water residual | 0.910 (0.805–1.029) |
| Residual after removing repeated plain water | 0.979 (0.883–1.085) |
The estimated association weakens towards the null once the extra water is removed. This is evidence that the measurement choice matters for a mortality analysis, not a claim that Zhou’s published HR 0.84 has been replaced by 0.979. Neither interval in our exploratory analysis excludes 1; we have not formally tested the difference between the two coefficients. Our period of follow-up, covariates, selection and survey weighting are not identical to the original paper. New checks restricted to 2003–2014 produced HRs of 0.984 versus 1.014, and requiring valid second-day recalls produced 1.026 versus 1.046. All corresponding confidence intervals included 1. These exploratory results do not establish a corrected mortality effect.
Why did the corrected HR change from 0.990 in the previous manuscript to 0.979 here? The earlier 0.990 estimated exposure from a separate, item-derived working food classification. The current 0.979 uses the same other-beverage classification as the literal arm and changes only the accounting identity after 2005: food_isolated = (M − B) − P, versus food_literal = M − B. The literal HR stays 0.910. Both versions use the same 31,046-person complete-case sample, but the corrected estimates are not alternative numerical runs of an identical exposure definition. We now report both instead of replacing one without explanation.
The archived earlier common-sample cohort analyses yielded the following Q4 versus Q1 results (2019 mortality follow-up; same v1 model and screen within each row; cohort-specific quartiles):
| v1 cohort / period | People / deaths | Literal food residual HR (95% CI) | Earlier item-derived food proxy HR (95% CI) |
|---|---|---|---|
| Pooled | 31,046 / 5,898 | 0.910 (0.805–1.029) | 0.990 (0.893–1.096) |
| Before 2005 | 10,169 / 2,987 | 0.928 (0.812–1.060) | 0.928 (0.812–1.060) |
| 2005 onward | 20,877 / 2,911 | 0.996 (0.765–1.297) | 1.076 (0.935–1.240) |
These estimates make clear that the period pattern was not a single stable result. The pre-2005 v1 contrast was HR 0.928 (0.812–1.060) under either definition, whereas an earlier, materially different pooled-weight specification gave HR 0.730 (0.595–0.896) before 2005. In other words, a protective association before 2005 appears under some specifications but not all; the estimated magnitude is model-dependent. The pooled HR need not lie between stratum-specific HRs because Cox HRs are non-collapsible, the quartile cut points are recalculated within each cohort, and different pooled/stratified models have different risk sets and adjustments. The full ledger of 48 archived food-, total-water- and food-share sensitivity contrasts, including screening variants and preliminary survey-weighted fits, is supplied below. These are exploratory analyses, not independent replications or a preregistered model set.
Nor can we infer that all reported associations were artifacts. The original study examined cancer mortality, heart disease, coffee and tea; those results need their own checks. But its central estimates for both “food water” and “total water” depend on exposure measures whose construction now has a serious unresolved problem. Plain-water and other-beverage variables need not be arithmetically duplicated themselves, yet adjusted associations can still change when the other exposure covariates are corrected.
What the evidence warrants
The CDC documentation establishes the change in measurement. The item-level records confirm the arithmetic. The same-sample comparison shows a large and predictable distortion. The published Zhou water distributions resemble the vulnerable reconstruction far more than the reconstruction without double counting.
We therefore consider a double-counting error in Zhou et al. (2022) likely, with potentially material consequences for their water-source and total-water conclusions. That conclusion is stronger than a general concern about how the methods were worded. It is still an inference about unobserved author code, not direct observation of that code.
The decisive final test is straightforward: recover the precise 35,463-person selection, the authors’ complete food/beverage mapping and the 2015 mortality follow-up, then run the published models under both accounting rules. Release of the original analysis code would resolve the key implementation question much faster.
Until such a check is available, we think it is premature to interpret the reported food-water association as evidence about water obtained from food rather than a possible artifact of combining incompatible NHANES cycles. Nothing in this audit supports an allegation of misconduct.
Methods, validation and reproducibility
Data provenance and selection
The analysis used NHANES 1999–2014 dietary data and the public-use mortality linkage through 2019. The historical supplied CDC-origin archives contained 96 entries, with internal ZIP checks and SHA-256 verification against supplied manifests. A separate GitHub Actions run independently downloaded 52 official CDC/NCHS public-use files and reproduced the core analytical tables and aggregate outputs. These hashes do not independently authenticate the archives against official CDC signatures. Records were linked by SEQN.
We reproduced the beginning of Zhou’s participant flow: 82,091 surveyed, 78,518 with examinations, 41,659 aged at least 20, and 40,315 after the pregnancy exclusion. Our next first-day eligibility count was 37,658 versus 37,708 from Zhou’s stated exclusions. The residual 50-person difference and differences after trimming remain unresolved. The paired exposure comparison used the 35,169 participants retained by the literal four-variable 97.5th-percentile screen, minus six with residual food water below −1 g, yielding 35,163. Three remaining residual values in [−1, 0) g were set to zero for display only. No identical claim is made about Zhou’s selected 35,463 participants.
Accounting identities
Let M be daily aggregate dietary moisture, P the official plain-water mass and B moisture assigned to non-plain beverages under the working taxonomy. The two source definitions are:
- Before 2005: total =
M + P; food =M − B. - 2005 onward, vulnerable formula: total =
M + P; food =M − B. - 2005 onward, no-double-count formula: total =
M; food =M − B − P.
These definitions keep the same B in both arms. The last is an accounting diagnostic, not yet a fully validated physiological measure of water originating in non-beverage food. P is a mass variable whereas M is nutrient moisture; small differences and a few negative residuals can arise. The working classification of non-plain beverages is based on code groupings and is not an independently validated cycle-specific FNDDS beverage taxonomy. A more exact implementation should classify item-level moisture using verified FNDDS descriptors, including combination-food flags.
The historical v1.1 plain-water check used item codes 94000100, 94100100 and 94300100 across 2005–2014. Subsequent 2009–2010 verification instead requires 94000100 and 94100100 (without 94300100), item gram weights (DR1IGRMS / DR2IGRMS) and combination flags (DR1CCMTX / DR2CCMTX). The historical cross-cycle check excluded combination-item water when reconstructing DR1_320Z / DR2_320Z; the corrected 2009–2010 check includes the relevant flagged combination item, eliminating the remaining 237 g discrepancy. The complete cycle-by-cycle check is shown below; the cycle is named by both years, rather than shorthand start years. Discrepancies are absolute differences greater than 1 gram between the official summary and the reconstruction.
| Cycle | Valid day 1 | Valid day 2 | Total person-days | Discrepancies >1 g |
|---|---|---|---|---|
| 2005–2006 | 9,169 | 8,264 | 17,433 | 0 |
| 2007–2008 | 9,118 | 7,715 | 16,833 | 0 |
| 2009–2010 | 9,623 | 8,288 | 17,911 | 94 |
| 2011–2012 | 8,389 | 7,496 | 15,885 | 0 |
| 2013–2014 | 8,531 | 7,453 | 15,984 | 0 |
The 2007–2008 cycle passed on both days. The 2009–2010 values shown above record the initial v1.1 diagnostic; a corrected, cycle-specific item-code audit subsequently reconciled both days with zero deviations above 1 g (methods update, 21 October 2026). Having no mismatch above 1 g is a consistency test against the CDC summary, not independent verification of every FNDDS code or beverage category. Separately, item moisture reconciled with total moisture in 101,153 matched person-days during 2003–2014.
Quartiles and follow-up
The paired exposure comparisons above use the same literal-screened sample and unweighted medians, with separately recomputed pooled quartiles. The 98.3% versus 67.7% late-cycle shares use Q4 membership within the pooled sample; the 61.0% reclassification figure uses quartiles recalculated within the post-2005 subgroup. The two percentages therefore have different denominators and must not be combined as one transition statistic.
Mortality sensitivity used a common analysis sample with non-missing model covariates (N = 31,046; 5,898 deaths), with Efron-tied Cox regression and sandwich covariance clustered by NHANES cycle, stratum and PSU, without full survey probability weighting. Adjustment included age, age squared, sex, BMI, ethnicity, education, marital status, income-to-poverty ratio, survey cycle, energy, protein, carbohydrate, fat, fibre, plain water and other-beverage water. The pair varies only the construction of the food-water exposure and its quartile boundaries. It is not the original Model 4: in particular, smoking, alcohol, physical activity and multiple medical-history variables were unavailable in this specification, and follow-up extends to 2019 rather than 2015. These are conditional associations, not causal estimates.
Complete archived model-specification ledger
The table enumerates 48 recoverable Q4/Q1 or explicitly labelled alternative-estimand contrasts from the saved results and execution logs. It is not a claim that no other ad-hoc runs took place. Model families are not commensurable; in particular, continuous food-share HRs measure an increment of 25 percentage points, not a quartile contrast. The source_v1 corrected measure and the matched_isolation_v2 corrected measure are constructed differently. For exact confidence intervals, p-values, all three quartile coefficients and machine-readable provenance, consult exploratory_model_ledger.csv and its source files in the research package.
View all 48 saved exploratory contrasts
| Run | Cohort | Exposure | Contrast | Adjustment | N | HR (95% CI) |
|---|---|---|---|---|---|---|
| early_unweighted_31633 | all | food_correct | food Q4/Q1 | age_sex_cycle | 31,633 | 0.686 (0.636–0.740) |
| early_unweighted_31633 | all | food_correct | food Q4/Q1 | plus_income_diet | 31,633 | 0.921 (0.826–1.027) |
| early_unweighted_31633 | all | food_correct | food Q4/Q1 | plus_other_water | 31,633 | 0.945 (0.845–1.057) |
| early_unweighted_31633 | all | food_literal | food Q4/Q1 | age_sex_cycle | 31,633 | 0.679 (0.620–0.743) |
| early_unweighted_31633 | all | food_literal | food Q4/Q1 | plus_income_diet | 31,633 | 0.903 (0.812–1.005) |
| early_unweighted_31633 | all | food_literal | food Q4/Q1 | plus_other_water | 31,633 | 0.878 (0.783–0.985) |
| early_unweighted_31633 | pre_2005 | food_correct | food Q4/Q1 | age_sex_cycle | 11,259 | 0.672 (0.608–0.741) |
| early_unweighted_31633 | pre_2005 | food_correct | food Q4/Q1 | plus_income_diet | 11,259 | 0.800 (0.699–0.915) |
| early_unweighted_31633 | pre_2005 | food_correct | food Q4/Q1 | plus_other_water | 11,259 | 0.818 (0.713–0.939) |
| early_unweighted_31633 | pre_2005 | food_literal | food Q4/Q1 | age_sex_cycle | 11,259 | 0.672 (0.608–0.741) |
| early_unweighted_31633 | pre_2005 | food_literal | food Q4/Q1 | plus_income_diet | 11,259 | 0.800 (0.699–0.915) |
| early_unweighted_31633 | pre_2005 | food_literal | food Q4/Q1 | plus_other_water | 11,259 | 0.818 (0.713–0.939) |
| early_weighted_31633 | all | food_correct | food Q4/Q1 | pool-weighted+Breslow; age/sex/year/diet/PIR/other-water | 31,633 | 0.901 (0.768–1.056) |
| early_weighted_31633 | all | food_literal | food Q4/Q1 | pool-weighted+Breslow; age/sex/year/diet/PIR/other-water | 31,633 | 0.850 (0.711–1.018) |
| early_weighted_31633 | post_2005 | food_correct | food Q4/Q1 | pool-weighted+Breslow; age/sex/year/diet/PIR/other-water | 20,374 | 1.311 (1.027–1.675) |
| early_weighted_31633 | post_2005 | food_literal | food Q4/Q1 | pool-weighted+Breslow; age/sex/year/diet/PIR/other-water | 20,374 | 1.196 (0.845–1.691) |
| early_weighted_31633 | pre_2005 | food_correct | food Q4/Q1 | pool-weighted+Breslow; age/sex/year/diet/PIR/other-water | 11,259 | 0.730 (0.595–0.896) |
| early_weighted_31633 | pre_2005 | food_literal | food Q4/Q1 | pool-weighted+Breslow; age/sex/year/diet/PIR/other-water | 11,259 | 0.730 (0.595–0.896) |
| source_v1_2019_common_trim | all | correct | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 31,046 | 0.990 (0.893–1.096) |
| source_v1_2019_common_trim | all | literal | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 31,046 | 0.910 (0.805–1.029) |
| source_v1_2019_common_trim | pre2005 | correct | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 10,169 | 0.928 (0.812–1.060) |
| source_v1_2019_common_trim | pre2005 | literal | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 10,169 | 0.928 (0.812–1.060) |
| source_v1_2019_common_trim | post2005 | correct | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 20,877 | 1.076 (0.935–1.240) |
| source_v1_2019_common_trim | post2005 | literal | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 20,877 | 0.996 (0.765–1.297) |
| source_v1_2019_retain_literal | all | correct | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 31,769 | 0.942 (0.847–1.047) |
| source_v1_2019_retain_literal | all | literal | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 31,769 | 0.901 (0.799–1.016) |
| source_v1_2019_retain_literal | pre2005 | correct | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 10,479 | 0.877 (0.761–1.009) |
| source_v1_2019_retain_literal | pre2005 | literal | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 10,479 | 0.876 (0.761–1.009) |
| source_v1_2019_retain_literal | post2005 | correct | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 21,290 | 1.055 (0.912–1.221) |
| source_v1_2019_retain_literal | post2005 | literal | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 21,290 | 0.998 (0.777–1.283) |
| source_v1_2019_retain_correct | all | correct | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 31,398 | 0.992 (0.895–1.100) |
| source_v1_2019_retain_correct | all | literal | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 31,398 | 0.896 (0.791–1.015) |
| source_v1_2019_retain_correct | pre2005 | correct | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 10,169 | 0.928 (0.812–1.060) |
| source_v1_2019_retain_correct | pre2005 | literal | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 10,169 | 0.928 (0.812–1.060) |
| source_v1_2019_retain_correct | post2005 | correct | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 21,229 | 1.075 (0.935–1.236) |
| source_v1_2019_retain_correct | post2005 | literal | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 21,229 | 0.943 (0.746–1.193) |
| matched_isolation_v2 | all | food_literal | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 31,046 | 0.910 (0.805–1.029) |
| matched_isolation_v2 | all | food_isolated | food Q4/Q1 | age,sex,BMI,race,edu,marital,PIR,diet,plain,B,cycle | 31,046 | 0.979 (0.883–1.085) |
| early_weighted_total_31633 | all | total_correct | total-water Q4/Q1 | weighted available-covariate model | 31,633 | 1.132 (1.000–1.281) |
| early_weighted_total_31633 | all | total_literal | total-water Q4/Q1 | weighted available-covariate model | 31,633 | 1.128 (0.995–1.279) |
| early_weighted_total_31633 | pre_2005 | total_correct | total-water Q4/Q1 | weighted available-covariate model | 11,259 | 1.110 (0.930–1.326) |
| early_weighted_total_31633 | pre_2005 | total_literal | total-water Q4/Q1 | weighted available-covariate model | 11,259 | 1.110 (0.930–1.326) |
| early_weighted_total_31633 | post_2005 | total_correct | total-water Q4/Q1 | weighted available-covariate model | 20,374 | 1.144 (0.958–1.366) |
| early_weighted_total_31633 | post_2005 | total_literal | total-water Q4/Q1 | weighted available-covariate model | 20,374 | 1.139 (0.963–1.347) |
| early_share_31633 | all | sharefood | food share +25 percentage points | age_sex_total | 31,633 | 0.791 (0.745–0.841) |
| early_share_31633 | all | sharefood | food share +25 percentage points | plus_diet_socio | 31,633 | 0.999 (0.927–1.077) |
| early_share_31633 | all | sharefood_literal | food share +25 percentage points | age_sex_total | 31,633 | 0.807 (0.762–0.854) |
| early_share_31633 | all | sharefood_literal | food share +25 percentage points | plus_diet_socio | 31,633 | 0.936 (0.881–0.996) |
Audit trail: Early N=31,633 analyses used limited demographic/dietary covariates with cluster-robust or experimental pool-weighted fits. The v1 XPT-derived N=31,046 models added BMI, ethnicity, education and marital status. The revised matched-isolation models fix non-plain beverages by construction and adjust only the post-2005 plain-water subtraction. An earlier quartile_cox_2019.csv contains only pooled and pre-2005 rows: its post-2005 runs cannot be recovered from that CSV and are not fabricated here. Early model1_results_2019.csv was overwritten during execution; the 18 archived runs were recovered from model1_execution.log and saved to model1_results_restored_from_execution_log.csv. No claim of complete analytic pre-registration is made.
Research report, DOI and reproducibility
Permanent research record: 10.5281/zenodo.23439456 (Zenodo record and historical research report). For citations, use the bibliographic details on the Zenodo record. The audit has not been peer reviewed. The 22 October 2026 supplements separately document reconstructed checks without modifying the historical Zenodo deposit.
Reproduce the results
Reproducibility materials, restored 22 October 2026. The Institute recovered the historical v1.1 archive with its analysis scripts, derived working data and 48-row exploratory-model ledger. That historical package is preserved as an unmodified audit record in private DigitalOcean research storage. The independently reconstructed 2009–2010 water-coding correction and mortality robustness calculation are publicly documented, with input SHA-256 hashes and methodological details. A reproducibility package containing derived public-use analytic tables, scripts, model outputs, the 48-row exploratory-model ledger and checksums was publicly released on 26 October 2026: download the verified research dataset and code (ZIP, 2.96 MB). SHA-256: 929f1706f8f6a88ba2aa86056c1e61635a51055d6455dd171611a2a842758d50. The package supports replay of the eight reported exploratory Cox estimates from derived inputs; it is not a complete reconstruction from the original authors’ code or all original CDC XPT files. The research-archive status and version notes explain the distinction. The existing Zenodo DOI continues to refer to the original historical version. The 48 historical contrasts have not all been regenerated, raw CDC XPT files are not redistributed, and the original authors’ unobserved code, follow-up differences and provisional beverage taxonomy preclude an exact replication of Zhou’s original mortality models.
# Core same-sample accounting comparison (2005 onwards)
food_literal = M - B
food_no_doublecount = M - B - P
total_literal = M + P
total_no_doublecount = M
# M: diary nutrient moisture; B: non-plain beverages; P: plain-water mass
References
- Zhou H-L, Wei M-H, Cui Y, et al. Association Between Water Intake and Mortality Risk: Evidence From a National Prospective Study. Frontiers in Nutrition. 2022;9:822119.
- US Centers for Disease Control and Prevention, NHANES 2005–2006. Dietary Interview: Total Nutrient Intakes, First Day (DR1TOT_D), including the water-collection change and moisture variable definitions.
- US Centers for Disease Control and Prevention, NHANES 2011–2012. Dietary Interview: Total Nutrient Intakes (DR1TOT_G), including
DR1_320Zand combination coding. - National Center for Health Statistics. Public-use linked mortality files.
Research transparency: This is an independent computational audit based on public NHANES data; the interpretation does not represent the original authors. It has not been peer-reviewed. AI tools assisted code review and manuscript preparation under human direction. The work identifies a probable data-processing issue, not evidence of intentional wrongdoing and not a recommendation about how much water to consume.