# Spatial and Machine Learning Analysis of District-Level Health Insurance Inequities in Ghana

## TL;DR

This study maps health insurance gaps in Ghana's districts, finding that uninsurance is clustered in northern regions and strongly linked to low literacy.

## Contribution

The paper introduces a novel combination of geospatial and machine learning methods to identify and predict health insurance inequities at the district level in Ghana.

## Key findings

- 42 'High-High' uninsurance clusters were identified, primarily in the northern savanna region.
- Illiteracy rate was the strongest predictor of uninsurance risk with 84.7% feature importance.
- A 30.3% illiteracy threshold significantly increases uninsurance risk in districts.

## Abstract

Background

Ghana's National Health Insurance Scheme (NHIS) remains central to achieving Universal Health Coverage (UHC). Nevertheless, UHC “resource commitment” is rarely paired with detailed access analyses, and inequities at the district level continue to persist. National aggregate data often overlook inequities at subnational levels, while the systematic and uneven resource distribution within the country’s most underdeveloped regions, sub-regions, or localities requires empirical identification.

Objective

The study aimed to examine the geographical distribution of uninsurance in Ghana and to perform predictive machine learning for the socioeconomic risk stratification of all 261 districts.

Methods

The study employed a quantitative, cross-sectional, and ecological design utilizing the 2021 Population and Housing Census data. The research integrates geospatial autocorrelation analysis with Decision Tree Classification (DTC). Spatial dependence and geographic clusters were assessed using Global Moran’s I and Bivariate Local Indicators of Spatial Association (LISA). For the machine learning model, which utilized Gini impurity as the splitting criterion, the features included variables of multidimensional poverty, illiteracy, and unemployment, while a 70/30 train-test split was used for all 261 Metropolitan, Municipal, and District Assemblies (MMDAs).

Results

There was significant positive spatial autocorrelation (Global Moran's I = 0.422, p < 0.001). This result identified 42 “High-High” uninsurance clusters, the majority of which were in the northern savanna region. The DTC model's overall accuracy was 82.5%, with the illiteracy rate being the primary predictor (84.7% feature importance) of uninsurance risk. Partial dependence analysis found a significant illiteracy threshold of 30.3%. The risk of a district being classified as "high" increased significantly as each successive unit above the threshold was recorded.

Conclusion

Health insurance unavailability in Ghana is geographically clustered and is strongly associated with literacy levels. While the design is ecological in nature and, as such, limits direct causal inference, it indicates that institutional complexity acts as a navigational barrier to enrollment. The 42 hotspots, in addition to the 10 “Red List” priority districts, serve as targets for interventions to reduce health inequities. Meeting these identified priorities is essential for the transition toward Universal Health Coverage and requires the integration of operational and system-level reforms with education-sensitive, geographically targeted resources.