Short-term PM10 exposure and hospital admissions for respiratory illness in children under five in Rio de Janeiro: a distributed lag study with negative-control evaluation of fifteen estimators

Daily time-series studies of air pollution and health rest on modeling choices that fit statistics cannot validate, since a bias shared by every candidate model is invisible to criteria computed on the same data. This study linked daily hospital admissions for respiratory illness among children under five in Rio de Janeiro to daily PM10 from eight municipal monitoring stations over 2011 to 2019, fitting a negative binomial distributed lag model across a 14-day window. Every estimator was refitted to exposure series shifted by one and two whole years, which preserves seasonality while breaking day-to-day correspondence, so any surviving association is not an exposure effect. Across 37,137 admissions, a 10 µg/m³ increase in PM10 was associated with a 3.75% rise in admissions (95% CI 0.63 to 6.97), of which same-day exposure recovered only 1.67%, and an estimated 449 admissions (80 to 795) were attributable to concentrations above the World Health Organization 2021 daily guideline. Eleven of fifteen estimators passed the control and four neural estimators failed. Seasonal adjustment at four degrees of freedom per year returned a near-null estimate whose shifted controls reached 6.53 percentage points, revealing broken identification rather than absence of an effect.