The Null Institute | Technical supplements | 22 October 2026
This page reconstructs the supplementary material previously linked from the independent audit of Zhou et al. (2022). The original GitHub-hosted documents became unavailable; the calculations below were independently rerun on the NI Research Runner. This is not a direct copy of the lost GitHub files. The original research record and DOI remain unchanged.
2009–2010 plain-water coding correction
We independently downloaded the official 2009–2010 NHANES dietary item and total files for both recall days. The initial v1.1 diagnostic selected item codes 94000100, 94100100, and 94300100, and omitted flagged combination records. It therefore reported 47 mismatches above 1 g on each day (94 total).
The corrected reconstruction uses 94000100 and 94100100, together with the applicable flagged combination record on day 1 (237 g). This reconciles the item-level sum with the official plain-water variable.
| Diary day | Valid person-days | Old mismatches >1 g | Corrected mismatches >1 g |
|---|---|---|---|
| Day 1 | 9,623 | 47 | 0 |
| Day 2 | 8,288 | 47 | 0 |
| Total | 17,911 | 94 | 0 |
The corrected maximum absolute difference is below 10-9 g on both days. The four input file SHA-256 hashes are: DR1IFF_F d2e2cf8195695f8c8a3890b62231fa71b344824fd531a86a483da8e1ccd30034; DR2IFF_F f88569d8c7a147690b5ba36300308db3b041e846d55a37499108aa725252ed80; DR1TOT_F 7ff258b1cb2020509d0d0ad3afb49797ca5196d085bac3e45cc6df149d93091c; DR2TOT_F 214759831d18d9bc0f0b2975fed277a7836e4d35774c5dc655643b582a28adbe.
This is a validation of NHANES accounting, not a direct inspection of the original authors’ unpublished code.
Mortality sensitivity excluding 2009–2010
Using the archived v1.1 derived analytic dataset, we independently reproduced the historical two paired Cox models (31,046 participants, 5,898 deaths). Changing only the study-cycle filter to exclude 2009–2010, and refitting the models and quartile boundaries, produced:
| Exposure | Participants | Deaths | Q4/Q1 HR | 95% CI | p |
|---|---|---|---|---|---|
| Published literal accounting | 26,323 | 5,270 | 0.917595 | 0.805225–1.045646 | 0.196943 |
| Plain-water-isolated accounting | 26,323 | 5,270 | 0.969315 | 0.865392–1.085718 | 0.590140 |
Neither confidence interval excludes 1. The analysis is exploratory, uses mortality follow-up through 2019, and does not reproduce the Zhou authors’ full model, original follow-up period, or all survey-design features. The results do not establish a causal effect of water intake.
Research archive and version history
The complete historical v1.1 archive (42 unchanged entries, source SHA-256 6c5717bcf50eb71272deec5910f58db79e42bb41858a0cbfc80b3dda935ccb50) has been recovered. A new 52-file recovery bundle adds the dated errata, mortality sensitivity documentation, executable verification scripts, derived aggregate checks and file checksums. It is preserved in the Institute’s DigitalOcean private research storage.
The historical archive retains the superseded 94-discrepancy diagnostic as an audit trail; the correction above takes precedence. Raw CDC XPT files are not redistributed. The full 52-file recovery bundle remains private because it contains participant-level derived data. A curated public-safe supplement with analysis code, aggregate results, and figures is now available: download the verified code and aggregate-results ZIP (137,604 bytes). Its SHA-256 is 4908064d223b2784d4ef91b3ec586732652f68434568f755498b5331ea9cf2da. This public archive excludes the participant-level input tables and is not a turnkey reproduction of the authors’ original models. The Zenodo research record remains the citable historical version.
Independent research audit. Not peer reviewed. Supplementary checks rerun 22 October 2026.