A computational investigation examining public health records across Indonesia has demonstrated how combining multiple analytical architectures can untangle complex regional drivers of childhood stunting. Conducting a nationwide analysis of administrative records spanning 514 districts nested within 38 provinces, researchers evaluated an integrated multi-method analytical framework uniting Bayesian multilevel structural equation modeling, explainable machine learning with Shapley additive explanations, structural association discovery algorithms, and latent profile analysis. The study addressed longstanding methodological bottlenecks in hierarchical health surveillance, where lower-level administrative units exhibit severe geographic disparities while operating within higher-level provincial boundaries. In Indonesia, where childhood stunting affects roughly 21.6 percent of children under five years of age, district-level stunting prevalence ranged dramatically from 1.60 percent to 52.10 percent, with a national district mean of 28.47 percent. The investigative team sought to simultaneously measure unobserved latent constructs, discover structural relationships, capture potential non-linear associations, and identify distinct district typologies across the country.
Ecological public health evaluations frequently suffer from analytical trade-offs when relying upon any single algorithmic technique. Conventional structural equation models estimate relationships among unobserved theoretical constructs but depend on linear assumptions and require pre-specified structural pathways. Multilevel regression designs account for geographic clustering but struggle with latent variables and measurement noise. Conversely, data-driven causal discovery algorithms explore graph architectures automatically but often deteriorate when applied to modest sample sizes, aggregated units, or derived factor scores. Supervised machine learning algorithms like random forests excel at identifying complex interactions and non-linear associations but lack inherent structural transparency. Rather than merging these disparate statistical paradigms into a singular composite equation, the investigators arranged four established methodologies into a parallel triangulation framework. By comparing where diverse algorithms converged or diverged, the team extracted diagnostic insights regarding structural determinants, algorithmic limits, and policy priorities without obscuring underlying data nuances.
To establish a consistent empirical foundation, the authors compiled complete district-level and provincial-level administrative records for the fiscal year 2022 from the official Siperindu database maintained by Indonesia’s National Population and Family Planning Board. The dataset encompassed all 514 administrative districts grouped across 38 provinces without sampling exclusions, creating a complete national registry containing no missing values. The primary measured outcome was district-level stunting prevalence, defined as the proportion of children under five whose height-for-age fell below minus two standard deviations from the World Health Organization Child Growth Standards median. At the district level, nine observed indicators were grouped into three theoretical latent constructs informed by international nutritional frameworks: demographic pressure, operationalized through crude birth rate, total fertility rate, and dependency ratio; socioeconomic vulnerability, measured via district poverty rate, open unemployment rate, and the human development index; and access to basic services, tracked through clean water coverage, improved sanitation access, and contraceptive prevalence rate. Contextual provincial predictors included provincial poverty percentage and provincial population density.
The first analytical pillar implemented a Bayesian multilevel structural equation model using Hamiltonian Monte Carlo sampling to isolate within-province district variations from broader provincial disparities. District continuous predictors were group-mean centered around provincial averages and standardized to zero mean and unit variance. Estimating parameters across four parallel Markov chains with 4,000 warm-up iterations and 8,000 sampling iterations yielded 32,000 posterior draws. All structural and measurement parameters achieved complete convergence with a potential scale reduction factor of exactly 1.000 and effective sample sizes exceeding 20,000. In the measurement model, factor loadings confirmed the integrity of the latent constructs: total fertility rate loaded most strongly on demographic pressure with an unstandardized factor loading of 1.411, while the human development index showed a strong inverse loading of minus 1.365 on socioeconomic vulnerability. Improved sanitation access anchored the basic services construct, whereas contraceptive prevalence exhibited a notably weak loading of minus 0.220.
The structural equations within the Bayesian multilevel framework delineated prominent direct and indirect pathways governing regional child growth disparities. Demographic pressure exerted a substantial positive effect on socioeconomic vulnerability, registering an unstandardized path coefficient of 0.542 and a standardized coefficient of 0.505, with a 95 percent credible interval ranging from 0.425 to 0.670. Elevated socioeconomic vulnerability in turn severely suppressed access to basic infrastructure and services, producing a standardized coefficient of minus 0.636. When examining direct effects on district stunting prevalence, both demographic pressure and socioeconomic vulnerability proved statistically meaningful, yielding standardized direct paths of 0.218 and 0.171, respectively. In contrast, the direct path from basic service access to stunting was not statistically significant, recording a standardized coefficient of minus 0.057. This suggests that the protective influence of clean water and sanitation operates primarily through broad socioeconomic conditions rather than functioning as an isolated, independent countermeasure.
At the macro level, provincial context exerted a powerful influence on local nutritional outcomes. Provincial poverty emerged as the dominant structural determinant across the entire multilevel specification, exhibiting an unstandardized coefficient of 0.356 and a standardized coefficient of 0.629, with a credible interval between 0.203 and 0.509. Provincial population density produced a non-significant linear estimate of minus 0.085. In total, the Bayesian multilevel model explained 42.3 percent of the between-province variance in stunting prevalence, while capturing 14.0 percent of within-province district variation. Despite these contributions, substantial residual variance remained unaccounted for, with 86.0 percent of within-province variance and 57.7 percent of between-province variance unexplained by the specified indicators. These residual components highlight the operational impact of unmeasured clinical, dietary, maternal, and environmental factors across Indonesia’s diverse administrative landscape.
The second pillar probed the empirical sensitivity boundaries of automated causal discovery when applied to real-world observational factor scores. Evaluating both the score-based Bayesian Network Structure Learning via hill climbing under the Bayesian Information Criterion and the constraint-based Peter-Clark algorithm using Gaussian conditional independence testing, the researchers attempted to reconstruct the six theoretically specified structural paths. To prevent independence assumption violations from repeated provincial measures, causal discovery inputs were strictly restricted to district-level indicators. Across non-parametric bootstrapping with 100 resamples at an edge stability threshold of 0.4, both discovery algorithms recovered identical fractions of the theoretical structure: only two of the six theoretical edges, or 33.3 percent, were successfully detected, specifically demographic pressure leading to stunting and socioeconomic vulnerability leading to stunting.
The modest 33.3 percent recovery rate illustrated critical practical boundaries for deploying artificial intelligence causal discovery tools within public health settings. The researchers documented that path coefficients below standardized thresholds of 0.20 to 0.30 fell beneath the empirical detection sensitivity of standard Gaussian independence tests under sample sizes of 514 units. Furthermore, utilizing latent factor scores rather than raw indicators introduced measurement error attenuation that dampened observable statistical associations. Based on these observed empirical constraints, the authors proposed cautious heuristics for public health analytics, suggesting that automated graph recovery requires standardized effect sizes exceeding 0.50, sample sizes exceeding 1,000 independent observations, raw indicator inputs, and explicit accommodations for potential non-linearity and hierarchical clustering before causal inferences can be reliably posited.
To capture complex non-linear associations inaccessible to linear structural equations, the third analytical component deployed a random forest ensemble trained on all eleven raw observed indicators without factor transformation. The ensemble was configured with 500 regression trees and tuned candidate split variables, evaluated through five-fold cross-validation grouped by province to prevent spatial data leakage. The random forest pipeline captured 38.95 percent of the variance in district-level stunting prevalence, with an out-of-bag root mean squared error of 0.7806 and mean absolute error of 0.6089. This predictive performance substantially exceeded both a baseline null model and a conventional linear regression model comprising all eleven predictors, which reached a coefficient of determination of only 27.51 percent, an error of 0.8506, and mean absolute error of 0.6860.
Permutation importance and Shapley additive explanations illuminated the exact nature of the predictive gains achieved by the machine learning ensemble. Provincial poverty maintained top ranking across metrics, causing a 43.697 percent increase in mean squared error when permuted, with positive Shapley values denoting increased stunting risk. Strikingly, population density ranked as the second most important determinant, generating a 29.825 percent increase in prediction error upon permutation, despite having appeared completely non-significant in the linear multilevel model. Shapley value decomposition and local regression smoothing revealed a distinct non-linear protective threshold: population density exerted minimal influence at lower levels but exhibited strong protective effects once urban concentration surpassed specific density margins. This divergence underscored the diagnostic utility of multi-method triangulation, showing that agglomeration benefits, including service proximity and economic diversification, act non-linearly to mitigate stunting burdens.
The fourth component employed latent profile analysis using an unrestricted Gaussian finite mixture formulation to identify hidden district typologies based on standardized factor scores and stunting rates. Model selection supported a three-profile solution, maximizing the Bayesian Information Criterion at minus 2789.27 with an entropy score of 0.502 and average class assignment probabilities near 0.80. Profile 1 encompassed 320 districts, representing 62.3 percent of the country, characterized by near-average demographic pressure, slight socioeconomic vulnerability, and average stunting prevalence of minus 0.006 standard deviations. Profile 2 captured 102 vulnerable districts, comprising 19.8 percent of the sample, characterized by concurrent deficits across demographic pressure, socioeconomic vulnerability, and deficient service access, coupled with the nation’s highest stunting burden at 0.356 standard deviations. Profile 3 comprised 92 districts, or 17.9 percent, displaying favorable circumstances with low vulnerability, superior infrastructure access, and low stunting prevalence at minus 0.372 standard deviations.
The authors contextualized their findings within several analytical boundaries inherent to cross-sectional administrative datasets. Because the Siperindu records capture an ecological cross-section from the 2022 fiscal year, observed directional pathways represent statistical associations and structural hypotheses rather than strict counterfactual causal determinations. Ecological aggregation across districts and provinces also precludes direct inferences regarding household behaviors or individual child growth trajectories. Additionally, key biological and dietary determinants, such as exclusive breastfeeding duration, complementary feeding diversity, micronutrient supplementation, and acute pediatric morbidity, were unavailable in administrative files. Nevertheless, by demonstrating how structural equation modeling, explainable machine learning ensembles, and latent classification mutually inform public health surveillance, the multi-method framework provides a transparent blueprint for diagnosing regional vulnerability and evaluating policy priorities across complex administrative hierarchies.
Subject of Research: Multi-method artificial intelligence framework for childhood stunting surveillance in Indonesia
Article Title: A multi-method AI framework integrating explainable machine learning, Bayesian multilevel modeling, and latent profile analysis for childhood stunting surveillance in Indonesia
Article References: Thamrin, S. A., Aidid, M. K., Arsyad, D. S., Amir, S. B. H., Lawi, A., & Nur, D. (2026). A multi-method AI framework integrating explainable machine learning, Bayesian multilevel modeling, and latent profile analysis for childhood stunting surveillance in Indonesia. Discover Artificial Intelligence, 6(1), Article 1862. https://doi.org/10.1007/s44163-026-02635-9
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02635-9
Keywords: childhood stunting, Indonesia public health, Bayesian multilevel SEM, machine learning, Random Forest, SHAP analysis, latent profile analysis, causal discovery, Bayesian network structure learning, PC algorithm, provincial poverty, population density, Siperindu database, demographic pressure, socioeconomic vulnerability, basic service access, hierarchical observational data, health surveillance
News Source: Denise Maddox. (November 13, 2026). AI Framework Uncovers Stunting Drivers and Clustering Patterns Across Indonesia. Scienmag.



