Is Predicted Data a Viable Alternative to Real Data?

Abstract

Abstract It is costly to collect the household- and individual-level data that underlie official estimates of poverty and health. For this reason, developing countries often do not have the budget to update estimates of poverty and health regularly, even though these estimates are most needed there. One way to reduce the financial burden is to substitute some of the real data with predicted data by means of double sampling, where the expensive outcome variable is collected for a subsample and its predictors for all. This study finds that double sampling yields only modest reductions in financial costs when imposing a statistical precision constraint in a wide range of realistic empirical settings. There are circumstances in which the gains can be more substantial, but these denote the exception rather than the rule. The recommendation is to rely on real data whenever there is a need for new data and to use prediction estimators to leverage existing data.