Abstract
In rural populations where traditional economic metrics like income or expenditure are difficult to collect, researchers often rely on asset-based surveys to quantify socio-economic status (SES). However, these surveys typically yield a complex mix of binary, ordinal, and nominal data that standard statistical methods struggle to integrate. This paper proposes a novel Bayesian framework, the Mixture of Factor Analyzers for Mixed Data (MFA-MD), to cluster households into homogeneous socio-economic groups. By unifying Item Response Theory (IRT) for binary and ordinal items with factor analysis for nominal items, the model identifies distinct social strata within the Agincourt Health and Demographic Surveillance System (HDSS) in South Africa. The resulting clusters provide a more nuanced understanding of rural poverty and modernization than traditional linear wealth indices.
Introduction
Quantifying socio-economic status is essential for public health research and policy planning, yet in many low-income settings, the lack of formal employment and high reliance on subsistence activities make income a poor metric. The Agincourt HDSS has conducted biannual household asset surveys since 2001 to bridge this gap. While asset ownership is a reliable proxy for wealth, the resulting data is "messy," containing binary indicators (e.g., owning a car), ordinal rankings (e.g., sanitation levels), and nominal categories (e.g., primary cooking fuel).
Historically, researchers have used Principal Component Analysis (PCA) to collapse these variables into a single wealth index. However, PCA often fails to account for the categorical nature of the data and assumes that poverty exists on a simple linear gradient. This paper argues for a model-based clustering approach that treats SES as a set of distinct social clusters rather than a single continuous line.
Methodology: The MFA-MD Framework
The authors introduce the Mixture of Factor Analyzers for Mixed Data (MFA-MD), which treats observed survey responses as manifestations of underlying, unobserved (latent) continuous variables. The model is built on three pillars:
- Hybrid Latent Modeling: To handle mixed data types, the framework employs Item Response Theory (IRT) for binary and ordinal items, assuming a latent trait determines the probability of a specific response. For nominal items, it uses a factor analysis approach similar to a multinomial probit model, which utilizes a multi-dimensional latent variable to handle unordered choices.
- Unification: These structures are combined into a single, high-dimensional latent vector for each household, allowing the model to process disparate data types simultaneously.
- Clustering via Mixtures: The model assumes the population is a mixture of $G$ different Gaussian densities within the latent space. Each cluster is defined by its own mean and "loadings," which describe how strongly each asset relates to that specific group's status.
Bayesian Estimation
The model is estimated using a Markov chain Monte Carlo (MCMC) algorithm, specifically a Metropolis-within-Gibbs sampler. This Bayesian approach allows for the simultaneous estimation of cluster memberships and model parameters while providing a robust framework for accounting for uncertainty.
Data and Application
The study utilizes data from 17,617 households across 21 villages in the Agincourt HDSS. The dataset includes 28 survey items, ranging from dwelling materials to the ownership of livestock and modern appliances. This region, a former "homeland" in northeast South Africa, is characterized by rapid infrastructure development alongside persistent economic hardship, making it an ideal test case for nuanced SES clustering.
Key Findings
The MFA-MD model identified a three-cluster solution as the most interpretable and statistically sound representation of the population:
- Cluster 1 (Wealthy/Modern): This group is characterized by high ownership of luxury goods (TVs, video players), the use of electricity for cooking, and modern flush sanitation.
- Cluster 2 (Intermediate/Striving): This cluster represents a transitional stage. These households often own basic modern assets like refrigerators and keep poultry, but they lack high-end luxury goods and continue to rely on wood for cooking.
- Cluster 3 (Poor): This group is defined by rudimentary dwelling materials, "bush" toilets, and a lack of modern assets.
One of the most significant insights is the importance of nominal data. By properly modeling cooking fuel, the authors captured the specific transition from wood to electricity—a nuance often lost in 1D indices that simply "binarize" such variables. The results demonstrate that SES is not just about having "more" or "less," but about unique "asset signatures" that define different social realities.
Conclusion and Future Directions
The MFA-MD framework provides a principled, objective method for defining socio-economic strata without relying on arbitrary cut-offs. While the current model provides a powerful static snapshot, the authors suggest that future work should extend this to longitudinal data to track how households move between clusters over time. Additionally, addressing the computational intensity of MCMC through rank-likelihood approaches could make the model more scalable for even larger demographic datasets. Ultimately, this approach offers a more sophisticated lens through which to view the complexities of rural poverty and development.
Summary updated