Helminth prevalence and fifth-grade reading performance in 390 Brazilian municipalities

Abstract

The association between intestinal parasitic infections and socioeconomic disadvantage is widely documented, as is the negative impact of material deprivation on learning. We investigated the extent to which municipal disparities in reading associated with the prevalence of four helminthiases — Ascaris lumbricoides, Trichuris trichiura, hookworms and Schistosoma mansoni — persist after adjustment for socioeconomic and school factors. In this ecological study, we linked the national parasitological survey (2010–2015), which examined 197,564 schoolchildren in 521 municipalities, to the 2019 results of the Brazilian national student assessment (Saeb), in the 390 municipalities with at least 100 schoolchildren examined. The crude difference between the extremes is large: in the highest A. lumbricoides fifth, 37.3% of students reached 200 reading points or more, against 59.8% in the lowest fifth — a gap of 22.4 percentage points. Adjustment for income, maternal schooling, sanitation and school structure attenuates that difference by 59% to 79%. What remains, 4 to 9 percentage points, is modest and unstable: it withstands the sensitivity checks, but diminishes once municipalities with smaller samples are included and ceases to be significant in the continuous specification under inverse-variance weighting, which the analysis plan defined as the primary scheme. Prevalence adds little to the full model, between 0.4 and 1.2 percentage points of R², and the distributions at the extremes overlap widely; on its own, it explains 17% of the variation. The four helminths showed continuous coefficients in the same direction. Because the data are municipal, nothing can be concluded about the learning of any individual infected child. Municipal A. lumbricoides prevalence functions, in these data, primarily as an indicator of territorial vulnerability, and adds little to the socioeconomic indicators already monitored. The analytical routines were programmed and executed by large language models from the authors' prompts — chiefly Claude Opus 5 and ChatGPT 5.6 —, using standard statistical libraries; the authors set the questions, approved the analytical decisions and checked the input data against the official sources. The central results were reproduced by independent recomputation and the spatial model re-estimated in the PySAL library; there was no review by an independent statistician and no re-extraction of the data from the primary sources, checks the authors are not trained to perform. Data, code, protocol and verification records accompany the work and are deposited in Zenodo (DOI 10.5281/zenodo.22664954).