Educational policymakers often lack data on student outcomes in regions where standardized tests were not administered. Machine learning techniques can be used to predict unobserved outcomes in target populations by training models on data from a source population. However, differences between the source and target populations, particularly in covariate distributions, can reduce the transportability of these models, potentially reducing predictive accuracy and introducing bias. We propose using double machine learning for a covariate-shift weighted model. First, we estimate the overlap score-namely, the probability that an observation belongs to the source dataset given its covariates. Second, balancing weights, defined as the density ratio of target-to-source membership probabilities, are used to reweight the individual observations' contribution to the loss or likelihood function in the target outcome prediction model. This approach downweights source observations that are less similar to the target population, allowing predictions to rely more heavily on observations with greater overlap. As a result, predictions become more generalizable under covariate shift. We illustrate this framework in the context of uncertain data on students' standardized financial literacy scores (FLS). Using Bayesian Additive Regression Trees (BART), we predict missing FLS. We find minimal differences in predictive performance between the weighted and unweighted models, suggesting limited covariate shift in our empirical setting. Nonetheless, the proposed approach provides a principled framework for addressing covariate shift and is broadly applicable to predictive modeling in the social and health sciences, where differences between source and target populations are common.
Bargagli-Stoffi et al. (Tue,) studied this question.