work · 2022 lancet global health geospatial modeling

Mapping solid-fuel cooking and attributable under-five respiratory mortality across 98 countries

Density plot of world population and under-five pneumonia deaths distributed across PM2.5 air-pollution exposure, 2000 vs 2018, split by country

Where the world’s people and children’s pneumonia deaths fall across the air-pollution exposure scale, 2000 vs 2018, by country. Read the paper →

This analysis estimated solid-fuel cooking prevalence at 5 km resolution across 98 low- and middle-income countries from 2000–2018. Those estimates fed household-air-pollution exposure and attributable under-five respiratory mortality estimates, with forecasts to the 2030 clean-cooking target.

5 km

resolution

countries

2.1M

survey observations

663

surveys

National averages

Progress toward the clean-cooking target, SDG 7.1, is commonly reported at national scale. Those averages cannot show where solid-fuel use is concentrated within a country or how exposure differs between districts. Estimating prevalence at 5 km resolution made those within-country gradients available for exposure, burden, and 2030 trajectory estimates.

Model pipeline

A model-based-geostatistics (MBG) pipeline that turns millions of scattered survey responses into continuous 5 km surfaces:

Assemble: 2.1 million household observations from 663 surveys, geo-located and collapsed to cluster-level solid-fuel-use rates, joined to dozens of satellite/geospatial covariates (nightlights, travel time, climate, land cover, population, ambient PM2.5).
Stack: a constrained ensemble of machine-learning submodels (GAM, gradient-boosted trees, penalized regression) fit with cross-validation. Their predictions become the mean surface.
Geostatistics: a Bayesian hierarchical model adds a spatial Gaussian process (SPDE) and country and sub-national random effects, fit with R-INLA. It borrows strength across space to estimate, with uncertainty, even where surveys are sparse.
Calibrate & aggregate: raked to national GBD estimates, aggregated to admin 0/1/2.
Forecast & burden: annualized-rate-of-change projections to 2030 vs. SDG 7.1, combined with ambient PM2.5 and relative-risk curves to estimate attributable under-five respiratory mortality.

Mapped results

In 2018, 593 million people lived in districts where solid-fuel use exceeded 95%. Only five of the 98 countries had a greater than 95% probability of reaching universal clean-cooking access in every district by 2030. The burden model attributed 205,000 under-five deaths from lower respiratory infections (95% UI 147,000–257,000) to household air pollution in 2018.

Map of the modeled probability of attaining SDG 7.1 clean-cooking target by 2030, red where unlikely, blue where likely

Probability of attaining the 2030 clean-cooking target (SDG 7.1). Red indicates where it’s unlikely on current trends.

My role

First author, responsible for the full analysis from survey extraction through the modeled surfaces, forecasts, and burden estimates.
The stacked-ensemble + Bayesian geostatistical model and its covariate/validation design.
The 2030 SDG forecasts and the exposure-to-mortality burden chain.
The paper, published in Lancet Global Health.

Published outputs

The published outputs include annual 5 km surfaces for 2000–2018, summaries at three administrative levels, 2030 clean-cooking projections, and attributable under-five respiratory mortality estimates.

Paper and code

the paper in lancet global health analysis code on github

Attribution. Published as Frostad et al., Lancet Global Health (2022); funded by the Bill & Melinda Gates Foundation; built on IHME’s model-based geostatistics framework. Figures are from the paper.