We introduce a new approach to the analysis of attrition in South African longitudinal surveys by supplementing the public-use data with paradata about the survey process and interview experience. The number of successfully interviewed respondents reduced from 7,073 in Wave 1 of NIDS-CRAM to 5,676 in Wave 2—a level of attrition of almost 20\%. We fit probit regression models to predict the determinants of attrition. In the fully specified model, attrition was most affected by contact effort by the survey organisation, the sample batch the respondent was in during Wave 1, employment status and whether they had a missing value for household income. Another important finding was that respondents who underwent COVID-19 tests were 3\% more likely to attrite, a trend that could negatively affect the efficacy of the survey to track COVID-19 testing behaviour in future waves. Attrition was not influenced by how often respondents previously participated or refused in NIDS, the interviewer-interviewee experience in the interview or by the respondent's observable demographic characteristics. This is a new contribution to the South African literature on attrition and shows the importance of using paradata to understand nonresponse.