An Original Contribution to Causal Discovery Methodology for Socioeconomic Data in Low-Income African Settings

Causal discovery algorithms that automatically infer causal structure from observational data offer significant potential for development economics research in Africa, where experimental studies are expensive and longitudinal data scarce, yet standard causal discovery methods require assumptions that are frequently violated in low-income African socioeconomic data. This study developed original adaptations to causal discovery methodology for low-income African socioeconomic data, with empirical validation using Nigerian household panel data. A methodology development approach was employed: systematic review of causal discovery methods and their assumption requirements (61 publications from 2018 to 2024), theoretical analysis of three specific assumption violations common in Nigerian socioeconomic data (measurement error in self-reported income, selection bias in panel attrition, and feedback loops in agricultural credit and yield), and empirical development and testing of adapted algorithms. The General Additive Noise Model and PC algorithm were adapted to incorporate measurement error correction procedures validated on simulated data with known causal structure. The adapted algorithms were applied to the Nigeria General Household Survey Panel 2010 to 2018 data to discover causal relationships between household asset accumulation, agricultural productivity, and credit access. Recovery of known ground-truth causal relationships (experimentally confirmed in existing RCT studies) improved by 23 percentage points over unadapted baselines. The original LINA-CAD (Low-Income Nigerian Adaptation to Causal Discovery) framework constitutes a significant methodological contribution to development data science.

Keywords: causal discovery, socioeconomic data, Nigeria, causal inference, methodology development