A cross-cycle water-accounting vulnerability in NHANES: a paired audit of a published mortality exposure
Ilya Nikitin
Abstract
Background. Zhou et al. (2022) reported inverse associations of dietary water-source intake with mortality in 35,463 NHANES participants from 1999–2014. Beginning in 2005–2006, NHANES recorded plain drinking water as itemised intake and included it in total dietary moisture. We examined whether a literal implementation of the published exposure definitions could add this water twice.
Methods. We compared CDC dietary-total and item-level records across eight NHANES cycles and constructed two source-water definitions in the same screened sample (N = 35,163). One followed a potentially duplicative literal residual; the other removed post-2005 plain water from the residual exactly once. We examined cycle discontinuities, distributional proximity to Zhou et al.'s published summary, and quartile reassignment. Exploratory mortality analyses used a distinct, non-equivalent Cox model with NCHS public-use follow-up through 2019. New robustness analyses refitted the archived model after restricting survey cycles to 2003–2014, with and without requiring a valid second-day recall.
Results. The original article's exposure definitions imply the potentially duplicative M + P construction unless an unreported harmonisation step was used. In 101,153 matched person-days, item-moisture sums reconciled with the reported aggregate to numerical precision. The literal versus corrected definitions yielded paired-sample median food-water residuals of 949 versus 556 g/day; the published Zhou median was 924 g/day. The vulnerable reconstruction created a post-2005 jump in median apparent food-water share from approximately 21% to 41–46%; the non-duplicative calculation remained approximately 20–25%. In the pooled exploratory 2019 model (31,046 participants; 5,898 deaths), the Q4/Q1 hazard ratios were 0.910 (95% CI 0.805–1.029) and 0.979 (0.883–1.085). In the 2003–2014 restricted cohort they were 0.984 and 1.014; with only valid second-day recalls, 1.026 and 1.046. All corresponding intervals included 1. A complete 48-row archive of exploratory model fits also includes materially different specifications, and cannot support a stable mortality-effect claim.
Conclusions. The literal post-2005 accounting vulnerability and large exposure-classification differences are demonstrable. Similarity of a vulnerable reconstruction to published marginal distributions raises a specific implementation concern, but Zhou et al.'s actual code is unavailable; we cannot establish that the original authors made that error. The mortality-model difference in our audit is exploratory and sensitive to survey-era and recall-completeness choices; it is not a corrected estimate of the original published association.