The published result reproduces
The original coefficients can be recovered. The disagreement is about interpretation, not arithmetic.
Open discussions OPEN DISCUSSION DRAFT · VERSION 1.1 · 22 JULY 2026
DRINKING WATER · ARSENIC · RANDOMIZED EVIDENCE
When a “successful” health intervention
stops looking successful.
An independent reanalysis of a drinking-water arsenic information trial in India—and a practical example of how reasonable modelling decisions can change the scientific story.
同一批数据，采用不同但都合理的分析方法，为什么会得出完全不同的健康结论？
WHAT THIS ISA paper opened early,
A paper opened early,
while its conclusions can still improve.
This is not a journal publication or a final verdict on the original trial. It is a documented working argument. We are publishing it now so that readers can challenge the data decisions, statistical models, and wording before the paper is treated as finished.
这不是正式发表的论文，也不是对原研究的最终裁决。我们选择在结论仍可修改时公开，希望读者直接指出数据处理、模型选择和表述中的问题。
01 — WHAT CHANGEDThe numbers reproduce.
The numbers reproduce.
The interpretation does not.
The health finding is model-dependent
The apparent reduction in stomach problems emerges mainly after controlling for knowledge gained after the intervention.
Exposure reduction remains uncertain
The individual arm may have improved, but the adjusted concentration difference is imprecise and the group arm shows no clear reduction.
The dramatic skin curve is an artifact
One observation at 1.0 mg/L drives the apparently extreme nonlinear result. Removing it changes the joint p-value to 0.14.
02 — THE CENTRAL RESULTOne added variable
One added variable
changes the story.
The group intervention appears to reduce stomach problems by 9.34 percentage points in the published model. In the randomized total-effect model, the estimate is 4.36 points and the confidence interval includes no effect.
The main shift occurs when the model controls for knowledge gained after the intervention. Because knowledge was one of the intended effects of the programme, that adjustment changes the question being answered.
Reproducible code tells us that the calculation can be repeated. It does not, by itself, tell us which interpretation is justified.
03 — READ THE FULL DRAFTTwo languages.
Two languages.
One evidence record.
一项“有效”的健康干预，为什么换个合理模型就不再有效？
自然中文表达，包含完整方法、结果、讨论、图表和参考文献。
When a “Successful” Health Intervention Stops Looking Successful
The complete English paper, including methods, numerical results, figures, limitations, and references.
04 — OPEN REVIEWTell us where
Tell us where
the argument fails.
Useful criticism is more valuable than general approval. We especially welcome comments on:
- Whether the three conflicted villages were handled fairly;
- Whether post-treatment knowledge belongs in the main health model;
- Whether the health outcomes can support an arsenic-specific claim;
- Whether any conclusion is stronger than the evidence allows.
Comments appear with this article so that the evidence, criticism, and later revisions remain connected.
Loading comments…