Equitable AI: Exploring the role of gender in poverty estimation models using geospatial data.

Area of Science:

• Socioeconomic data analysis
• Geospatial statistics
• Machine learning applications

Background:

• Household surveys are crucial for poverty measurement but have spatial and temporal limitations.
• Machine learning (ML) methods using geospatial data can bridge these gaps for poverty mapping.
• Gender-specific performance differences in ML poverty prediction models remain understudied.

Purpose of the Study:

• To investigate gender-related differences in the performance of ML models for poverty prediction.
• To assess ML model accuracy using geospatial data for male- versus female-headed households in Ghana.
• To identify factors contributing to performance disparities in poverty mapping models.

Main Methods:

• Utilized random forest ML models with accessible geospatial data.
• Trained and validated models using Ghana's Demographic & Health Survey asset holdings data.
• Differentiated model performance by aggregating asset holdings of female- and male-headed households.

Main Results:

• ML models trained on male-headed household data achieved high accuracy (R² = 0.85).
• Models trained on female-headed household data showed lower but reasonable accuracy (R² = 0.75).
• The accuracy gap is partially attributed to the smaller sample size of female-headed households in survey data.

Conclusions:

• ML models effectively extend the spatial and temporal reach of survey data for poverty analysis.
• Performance differences in ML poverty prediction are influenced by survey sampling design, particularly for female-headed households.
• Future survey designs should aim for larger samples of female-headed households to enhance ML model accuracy.