Abstract

Alternative data sources such as satellite imagery, administrative records and citizen-generated data can give National Statistical Offices (NSOs) timelier and more granular statistics. They were not designed for statistical purposes, though, so they must be validated before use. In my presentation, I will discuss how machine learning (ML) can support that validation, drawing on African experiences, and what safeguards keep it trustworthy.
I will highlight
• Where ML helps- drawing from use cases in Nigeria, Tanzania, and across Africa). In each case the official benchmark is what makes the ML output credible.
• Where ML can mislead- addressing issues of prediction errors and testing for systematic and geographic bias, not just overall accuracy.
• What NSOs can do, using the UNECE Quality Framework for Statistical Algorithms covers accuracy, explainability, reproducibility, timeliness and cost-effectiveness, and complements existing quality frameworks rather than replacing them.
My presentation will show that transparent, well-validated ML is essential to public trust in official statistics in the era of AI.