District-level estimation of vaccination coverage: Discrete vs continuous spatial models

Area of Science:

• Epidemiology
• Biostatistics
• Geographic Information Systems (GIS)

Background:

• Health and development indicators (HDIs) like vaccination coverage are crucial for monitoring progress in low- and middle-income countries.
• Household surveys are often used due to unreliable routine data collection systems.
• There's a growing need for subnational (e.g., district-level) HDI estimates for targeted interventions and evaluations.

Purpose of the Study:

• To compare discrete spatial smoothing models with continuous Gaussian process (GP) models for estimating district-level vaccination coverage.
• To evaluate the predictive performance of these spatial modeling approaches using Bayesian inference.
• To assess the impact of accounting for between-cluster variation in GP models.

Main Methods:

• Utilized a fully Bayesian framework with Integrated Nested Laplace Approximation (INLA) and Stochastic Partial Differential Equations (SPDE).
• Compared discrete spatial smoothing models (district-level data) with continuous GP models (geolocated cluster-level data).
• Analyzed vaccination coverage data from the 2014 Kenya DHS and 2015-16 Malawi DHS.

Main Results:

• Continuous GP models demonstrated strong predictive performance, serving as a viable alternative to discrete spatial models.
• Accounting for between-cluster variation within continuous GP models did not significantly alter district-level estimates.
• Both modeling approaches provided reliable estimates for vaccination coverage.

Conclusions:

• Continuous Gaussian process models are a credible and effective tool for generating subnational vaccination coverage estimates.
• These spatial modeling techniques enhance the reliability of health and development indicator data in data-limited settings.
• The findings offer practical guidance for utilizing model-based approaches in public health surveillance and program evaluation.