Multiscale spatial modelling to support targeting of childhood stunting in Rwanda : From kernel-weighted local models to neural network-learned locality
(2026)- Abstract
- Childhood stunting (height‑for‑age z‑score < −2) remains a global public health challenge, and national averages often mask subnational clustering and spatial heterogeneity of underlying determinants. Across four papers, this thesis integrates geocoded household data with gridded geospatial surfaces from nationally representative survey data to map stunting hotspots, quantify spatially varying and non‑linear associations between stunting and its determinants, and produce decision‑support surfaces to guide targeted intervention prioritisation. In Papers I, II and III, we use a population‑based household survey conducted in November–December 2021 across 137 villages in the Northern Province of Rwanda. This survey links anthropometry for... (More)
- Childhood stunting (height‑for‑age z‑score < −2) remains a global public health challenge, and national averages often mask subnational clustering and spatial heterogeneity of underlying determinants. Across four papers, this thesis integrates geocoded household data with gridded geospatial surfaces from nationally representative survey data to map stunting hotspots, quantify spatially varying and non‑linear associations between stunting and its determinants, and produce decision‑support surfaces to guide targeted intervention prioritisation. In Papers I, II and III, we use a population‑based household survey conducted in November–December 2021 across 137 villages in the Northern Province of Rwanda. This survey links anthropometry for children aged between 1 and 36 months with a wide range of child, maternal, household and environmental characteristics, and household geolocations. Paper IV scales up nationally using stunting and determinant surfaces aggregated to lower‑level administrative units to generate predictions aligned with implementation scales. In Paper I, global logistic regression is complemented by geographically weighted logistic regression (GWLR) to identify where associations are strongest. GWLR improves model fit and resolves residual spatial autocorrelation, uncovering substantial local variation in the influence of child, maternal and household factors. Paper II extends local modelling using multiscale GWR (MGWR) and generalized additive models (GAMs), identifying predictor specific spatial scales and nonlinear effects of elevation, slope, rainfall, vegetation index and temperature. The resulting thresholds provide insights to support geographically tailored interventions. Paper III further extends geographically weighted modelling using Geographically Neural Network Weighted Regression (GNNWR), which learns spatial weights and captures non‑linear local effects, producing coherent coefficient surfaces that demonstrate the value of hybrid spatial deep learning for geographically nuanced inference. At the national scale, the fourth study compares spatial econometric models with XGBoost, interpreted using SHapley Additive exPlanations (SHAP), and a spatial regression graph convolutional neural network (SR-GCNN). SR-GCNN achieves higher predictive accuracy and the lowest residual spatial autocorrelation. SHAP and spatial Durbin model analyses consistently identify sanitation, literacy, maternal health care and malaria prevention as key determinants of HAZ. Scenario simulations show that combined, equity‑oriented improvements across these domains could produce the largest predicted reductions in high‑burden areas. These counterfactuals inform prioritisation but are not causal estimates. Overall, stunting is strongly clustered and its determinants vary in magnitude and functional form across space. This thesis contributes a transferable modelling ladder, from geographically weighted regression to spatial deep learning, integrated with a scenario engine that converts multi‑source geospatial data into shareable risk, attribution and “what‑if” surfaces, offering practical support for geographically targeted public‑health decision‑making. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/193c1339-c1ad-495f-8c20-e17842e84a6d
- author
- Kagoyire, Clarisse LU
- supervisor
- opponent
-
- Professor C. Kyriakidis, Phaedon, Cyprus University of Technology, Cypern.
- organization
- publishing date
- 2026-03
- type
- Thesis
- publication status
- published
- subject
- keywords
- child stunting, spatial heterogeneity, geographically weighted regression, spatial deep learning, scenario mapping, targeted interventions, Rwanda
- pages
- 75 pages
- publisher
- Lund University
- defense location
- Geocentrum II, Pangea hall (229). Join via zoom: https://lu-se.zoom.us/j/63545620054
- defense date
- 2026-04-10 09:00:00
- ISBN
- 978-91-8104-888-9
- 978-91-8104-889-6
- language
- English
- LU publication?
- yes
- id
- 193c1339-c1ad-495f-8c20-e17842e84a6d
- date added to LUP
- 2026-03-06 13:51:28
- date last changed
- 2026-03-16 14:27:18
@phdthesis{193c1339-c1ad-495f-8c20-e17842e84a6d,
abstract = {{Childhood stunting (height‑for‑age z‑score < −2) remains a global public health challenge, and national averages often mask subnational clustering and spatial heterogeneity of underlying determinants. Across four papers, this thesis integrates geocoded household data with gridded geospatial surfaces from nationally representative survey data to map stunting hotspots, quantify spatially varying and non‑linear associations between stunting and its determinants, and produce decision‑support surfaces to guide targeted intervention prioritisation. In Papers I, II and III, we use a population‑based household survey conducted in November–December 2021 across 137 villages in the Northern Province of Rwanda. This survey links anthropometry for children aged between 1 and 36 months with a wide range of child, maternal, household and environmental characteristics, and household geolocations. Paper IV scales up nationally using stunting and determinant surfaces aggregated to lower‑level administrative units to generate predictions aligned with implementation scales. In Paper I, global logistic regression is complemented by geographically weighted logistic regression (GWLR) to identify where associations are strongest. GWLR improves model fit and resolves residual spatial autocorrelation, uncovering substantial local variation in the influence of child, maternal and household factors. Paper II extends local modelling using multiscale GWR (MGWR) and generalized additive models (GAMs), identifying predictor specific spatial scales and nonlinear effects of elevation, slope, rainfall, vegetation index and temperature. The resulting thresholds provide insights to support geographically tailored interventions. Paper III further extends geographically weighted modelling using Geographically Neural Network Weighted Regression (GNNWR), which learns spatial weights and captures non‑linear local effects, producing coherent coefficient surfaces that demonstrate the value of hybrid spatial deep learning for geographically nuanced inference. At the national scale, the fourth study compares spatial econometric models with XGBoost, interpreted using SHapley Additive exPlanations (SHAP), and a spatial regression graph convolutional neural network (SR-GCNN). SR-GCNN achieves higher predictive accuracy and the lowest residual spatial autocorrelation. SHAP and spatial Durbin model analyses consistently identify sanitation, literacy, maternal health care and malaria prevention as key determinants of HAZ. Scenario simulations show that combined, equity‑oriented improvements across these domains could produce the largest predicted reductions in high‑burden areas. These counterfactuals inform prioritisation but are not causal estimates. Overall, stunting is strongly clustered and its determinants vary in magnitude and functional form across space. This thesis contributes a transferable modelling ladder, from geographically weighted regression to spatial deep learning, integrated with a scenario engine that converts multi‑source geospatial data into shareable risk, attribution and “what‑if” surfaces, offering practical support for geographically targeted public‑health decision‑making.}},
author = {{Kagoyire, Clarisse}},
isbn = {{978-91-8104-888-9}},
keywords = {{child stunting; spatial heterogeneity; geographically weighted regression; spatial deep learning; scenario mapping; targeted interventions; Rwanda}},
language = {{eng}},
publisher = {{Lund University}},
school = {{Lund University}},
title = {{Multiscale spatial modelling to support targeting of childhood stunting in Rwanda : From kernel-weighted local models to neural network-learned locality}},
url = {{https://lup.lub.lu.se/search/files/244160385/Thesis_Clarisse_Kagoyire_LUCRIS.pdf}},
year = {{2026}},
}