Measuring poverty persistence with missing data with an application to Peruvian panel data

We consider the estimation of measures of persistent poverty in panel surveys with missing data, focusing on the persistent poverty headcount, its duration-adjusted variant, and a related measure used by the European Union as an indicator of the risk of persistent poverty. We develop a partial identification approach to allow for data missing-not-at-random, and apply it to panel data from Peru for 2007-11. The "worst case" bounds are very wide, but we achieve much more precise identification by adding a set of weak a priori restrictions. Standard non-response weighting adjustments cannot be relied upon to remove missing-data bias.