Reproducibility of Stunting Determinant Rankings Across Age Cohorts and National Survey Waves

Abstract

Analysis code for a measurement-validity study of model-derived determinant rankings for childhood stunting, using two harmonized waves of the Indonesia Nutritional Status Survey (SSGI 2022 and SSGI 2024). The code does not develop or validate a deployable prediction model: a gradient-boosted tree is used solely as a measuring instrument, and the object of measurement is the determinant ranking itself. Reproducibility is summarized by a single named estimand, the rank-based reproducibility coefficient, defined as the Spearman rank correlation between two importance profiles, with permutation nulls and bootstrap confidence intervals. The pipeline partitions the data into four period-by-age-cohort cells (baduta 0-23 months and balita_tua 24-59 months, crossed with the two waves) and, under an anti-leakage protocol that excludes outcome-forming anthropometry, measures three layers of determinant structure: redundancy among determinants, additive main effects, and pairwise interactions. It then quantifies within-cohort reproducibility across waves, replication of between-cohort differences, sensitivity of the recovered structure to imputation, and a directly estimated measurement-noise floor against which the observed cross-wave instability is judged. This repository is one of three downstream studies on Applied Explainable AI for Health Risk Prediction, with childhood stunting in Riau Province, Indonesia, as the validation domain. It operates on the harmonized master dataset produced by the Stunting Harmonization Pipeline, which is archived separately. The microdata are governed by the Ministry of Health of the Republic of Indonesia and are not redistributed; the harmonized master Parquet is a derivative and is never committed. The synthetic test-data generator in the harmonization repository allows this pipeline to be run and verified without restricted data.