Abstract:
We uncover and describe major data quality problems in South Africa’s 2022 Census microdata, which suggest substantial undocumented imputation concentrated in approximately 17-20% of households. Using the public 10% microdata sample, we first reproduce and extend existing research which shows that the distribution of person weights in the 10% sample is unusual and extreme, while the household weight distribution is more typical. We then show that the most extreme person weights (vanishingly small weights of less than 1) are all found in a subset of households which share peculiar item non-response patterns. These households have anomalous and highly implausible characteristics, both for the household as a whole and for the individuals in these households. Close inspection of these households suggests that these implausible characteristics are due to problematic and undocumented imputation. The effects of this imputation are substantial and are reflected in official Census statistics as well as the 10% sample: for example, the reported increase in the prevalence of multiracial households from 1.7% in Census 2011 to 11.4% in Census 2022 is almost fully explained by implausible imputation in the subset of households we identify. We discuss the possibility that these households are wholly imputed and speculate as to what could have motivated such a procedure. 

About the presenter: 
Joshua Budlender is a Postdoctoral Research Fellow at SALDRU. His work is in Labour and Development Economics, using Applied Microeconometrics methods. He has a particular interest in exploring the role of firms in the development process and in determining wages and employment. He has also done economic advisory work on social grant policy, and other interests include the economic history and long-run effects of racial discrimination in South Africa, and evaluating the developmental impacts of policy successes and failures in South Africa today. He received his PhD in Economics from the University of Massachusetts, Amherst in September 2024.