Computing Multidimensional Composite Indicators for Small Areas in Presence of Missing Variables: a Data Integration Approach

We evaluate data integration methods to estimate small area composite indicators, when some of the single indicators cannot be computed due to completely missing variables needed for their computation. The parameter is a multidimensional poverty index, where some of the required variables are not available in the population Census, which is used as the main source to compute the indicator. We propose two approaches to generate these missing variables, considering an auxiliary sample survey. Specifically, the performance of an approach based on a generalized linear mixed model is compared with a two-step imputation technique. The measurement of multidimensional poverty, also including nonmonetary dimensions is crucial and aligned with the Sustainable Development Goals defined by the United Nations. We consider Colombia as a case study, which has a recent population Census providing most of the information necessary to compute the indicator at small area level. Our methodologies can be greatly of interest of other Latin American countries having similar indices, and other countries computing poverty indicators with missing variables. The approaches are evaluated via simulations. We show an application based on the National Population Census, 2018 and the Great Integrated Household Survey 2018 of Colombia, focusing on the Antioquia region.