Blood-pressure reading and quality-control choices move India's raised-blood-pressure prevalence several times beyond its sampling error: an executable multiverse analysis of NFHS-4 and NFHS-5

Background. National estimates of raised blood pressure from the same survey microdata differ substantially across published analyses of India's National Family Health Survey (NFHS). NFHS measures blood pressure on a single occasion, so every estimate here concerns raised blood pressure on one measurement occasion, not clinically diagnosed hypertension. How much of the spread across published estimates is attributable to defensible analytical choices, rather than to sampling variation, has not been established. Methods. We encoded the analytical decisions of 20 published NFHS hypertension papers into an executable specification registry and ran 144 prospectively specified analyses (72 per round) on NFHS-4 and NFHS-5 person-recode microdata, crossing eligible population, blood-pressure reading-selection algorithm, quality-control scope, case definition and survey estimand. Estimates use a Taylor-linearized ratio estimator with sex-by-five-year-age post-stratification to household-roster targets, which corrects the differential blood-pressure coverage of the NFHS-4 male subsample. Two estimand families are reported separately rather than pooled. Results. Holding case definition, eligible population and estimand fixed, the five reading-selection and quality-control algorithms move prevalence by 1.58 to 3.07 percentage points, which is 4.5 to 19.6 times the design-based standard error of the estimate. Only 70 of 480 sibling estimates fall inside a given specification's own 95% confidence interval, and 66 of 120 specifications have confidence intervals containing no sibling at all. Within a fixed case definition, state rankings are reproducible: every within-definition bootstrap interval lies entirely above the across-definition interval in both rounds. Disparity directions are robust, with all 432 specification-by-gap combinations positive and 428 surviving Holm correction. On the specifications matching India's national NCD monitoring indicator, the reading-rule choice alone moves the monitored quantity by 46% of the reduction the 25% relative-reduction target demands. Conclusions. Reading-selection and quality-control choices, each individually defensible and each drawn from published practice, move an Indian national prevalence estimate several times beyond its reported precision. These are the fields least often reported: the blood-pressure validity thresholds are clearly reported by 3 to 4 of 20 papers, and no paper reports every field needed to reconstruct its own analysis. Levels should be reported as ranges over defensible specifications; rankings and disparity directions can be reported as point findings. Code, specification registry with per-field source quotes, and the full results archive are deposited.