 This paper is looking at how to estimate measures of poverty and inequality from microdata and The issue is with any form of microdata that you're going to encounter different sources of survey error and these are summarized in this table here. So what I do is separate the kind of Formants of survey error that researchers will observe in public use data and that's measurement error processing error and non-response error Which is distinct to the survey organization which deals with these other sources of error and then releases the data So as researchers then who are interested in measuring poverty and inequality We see a public use data set and we suspect forms of error that are generally going to be these three Now in 2015 the UN has to measure whether the first Millennium Development Goal Which is reducing poverty or absolute poverty by 50% between 1990 and 2015 Whether that's actually being achieved So what my objective was This is the fourth paper which has come after my PhD thesis These three papers are the PhD thesis which elaborates on this as using this as a framework for investigating microdata quality then establishing whether with these different sources of error you have bias which is ignorable or not and then lastly conducting multiple imputation when you have Forms of non-response and course response What I then do is I combine editing with imputation So when you combine editing with imputation you have to have rules for editing and rules for multiple imputation and your Imputation rules cannot Violate your editing rules. So what this means is when dealing with outliers, which I identify in this paper I combine all of these aspects I use multiple imputation and then I use a reweighting method to deal with the last of these which is outliers and I do one of two things. I report the statistics when the outliers Once they're identified I adjust a bit adjusted by reweighting the outliers down to one and when they're basically Left in the data to look at the sensitivity of the estimates to that But in so doing what I have done is combined solution methods, which are discussed here for all three sources of error So in the end what we find is that poverty and inequality measures can differ by as much as 30 percent So these are data sets from South Africa from the mid 90s through to early 2003 The methodology in these four papers is generalizable to any household survey anywhere in the world the thing that I've done here is applied it to the Sample of economically active employed individuals that self-employed and just employees So it's going to be because I'm taking such a narrow sample It's going to be the smallest form with the smallest level in which outliers are going to exert this big influence if we were looking at For example all different sources of income We may find that certain forms of error dominate the final survey statistic So the methodology then is generalizable in any sense whether you're looking at the full population of a country or subsets of that and Basically the treatment method which is using multiple imputation with an editing rule Is the one that you would need to read the paper for to fully Criticize and feedback is welcome at this email address