The design of the income question in household surveys usually includes response options for actual income, bracketed values, "Don't Know" and "Refuse" responses. This paper conducts an analysis of these response types using sequential response models specified analogously to those in the survey participation literature. We analyse the income question in Statistics South Africa's October Household Surveys (1997-1999) and Labour Force Surveys (2000-2003). The choice of survey years coincides with a period of development of the income question during which additional response options were steadily introduced to the questionnaire. An analysis of this sort sheds light on the underlying response process, which is useful for survey planning purposes and to researchers concerned with diagnosing the item missing and partial response mechanisms for variables of interest. It was found that the probability of a bracketed response increases as income increases, suggesting that this response option plays a significant role in getting higher income earners to answer the question. However, the relationship between response type and the correlates of income are no longer consistently statistically significant when the item nonresponse subset is decomposed into "Don't Know" and "Refuse". These findings suggest that response propensity models can help reduce specification error in single or multiple imputation algorithms. This is a joint SALDRU and DataFirst working paper

Questionnaire Design, Response Propensity Models, Employment Income Classification-JEL: C81, C83, D31