Spatial Inequalities in Maternal Health: What Maps Tell Us That Tables Cannot

Numbers in a regression table tell you that an association exists. A map tells you where it exists — and where it is worst. In maternal health research across South Asia and Sub-Saharan Africa, this geographic dimension is not just interesting; it is essential for policy.

Why Spatial Analysis?

In Bangladesh, the national skilled birth attendance (SBA) rate is approximately 67%. But this aggregate number masks dramatic sub-national variation: Sylhet Division consistently records among the lowest rates (<50%), while Dhaka Division exceeds 80%. A national policy that ignores this heterogeneity will fail the women who need it most.

Our Approach: Combining ML with Geospatial Methods

In our multi-country studies, we follow a three-step spatial pipeline:

1. Cluster-Level Prediction

Using DHS GPS cluster coordinates, we generate ML-based predictions for each enumeration area and map them using sf and tmap in R (or geopandas in Python).

2. Spatial Autocorrelation Testing

We apply Moran’s I statistic to test whether low-SBA clusters are randomly distributed or spatially clustered. In all our studies, we find significant positive autocorrelation — meaning high-risk areas are geographically concentrated and mutually reinforcing.

3. Local Indicators of Spatial Association (LISA)

LISA maps identify specific “hot spots” (high-high clusters) and “cold spots” (low-low clusters) of poor maternal health outcomes. These become the direct targets for geographically prioritized interventions.

A Striking Finding: The Haor Region

In our study of pregnancy danger sign knowledge in the Haor (wetland) region of Bangladesh, women in these flood-prone, geographically isolated communities showed dramatically lower knowledge scores than the national average — even after controlling for education, wealth, and healthcare access. The spatial pattern revealed that geographic isolation itself is an independent determinant, above and beyond individual-level factors.

This finding has direct policy implications: mobile health units and community health worker programs must be geographically targeted to Haor districts, not just allocated proportionally by population.

Tools We Use

The combination of ML predictions + spatial visualization is now a standard part of our research pipeline. If you are working on similar problems and want to collaborate, reach out: saleksta@gmail.com