Cross-validating administrative and survey datasets through microsimulation

In this paper we cross-validate two sources of data, administrative and sample survey, addressing an issue commonly faced by analysts regarding the relative reliability and comparability of these two data sources. By way of case study, the paper uses data presently available in the Grand-Duchy of Luxembourg. While administrative data extracted from the recently implemented Social Security Data Warehouse contains information about the whole population of Luxembourg (449,000 observations) in 2003, survey data, extracted from the Luxembourg household panel PSELL3/EU-SILC for 2004 (incomes from 2003), provides a representative sample of only around 3,600 private households (9,800 individuals) living in Luxembourg. The attraction of the survey is the more detailed information it provides on incomes, family relationships and other socio-economic dimensions. Our paper first analyzes the advantages and limitations of each dataset, before outlining and addressing methodological difficulties relating to their cross-validation. Through the cross-validation that follows we conclude that the survey database performs reasonably well in capturing the relevant characteristics of the resident population and allows analyses with respect to characteristics not found in the administrative database, and vice versa. Importantly we find that even if, on average, some monetary variables are different in the two datasets, the shapes of the equivalised income distributions broadly coincide. Even so, we observe a few important discrepancies at the extremes of the curves. Finally, through use of the EUROMOD microsimulation platform, we are able to show that the discrepancies observed between these income data sources are insufficient to significantly affect the conclusions drawn from analysis of policy alternatives.