Estimating Poverty and Inequality from Micro Data with Multiple Sources of Survey Error∗

Univariate estimates of poverty and inequality from income micro data are frequently plagued by multiple sources of survey error that affect both estimation and inference of parameters of interest. Observable forms of survey error include coarse data, processing error and measurement error. This paper builds on earlier work by Daniels (2012a, 2012b and 2012c) to include an explicit treatment of outliers in the labour income distribution using a multivariate outlier detection technique for identification purposes. The treatment of multiple forms of survey error includes combining both multiple imputation and reweighting. Estimates of the purchasing power parity adjusted poverty headcount (at US$1.00 and US$2.00 per day) and gini coefficient for the employed, economically active population of South Africa differ by as much as 13 percent for poverty and 20 percent for inequality over the period 1997-2003 when survey error is not corrected for compared to when it is corrected. These results suggest that researchers need to be cognisant of the potential influence of survey error on estimates and inference about the income distribution.