In many applied settings, collecting outcome data is expensive, time-consuming, or technically challenging, even when covariates are available for a large sample. Researchers must then decide which units to actually measure. A new arXiv preprint (2610.19351) formalizes this as a subdata selection problem in linear regression.

The abstract emphasizes that the selection must work under model misspecification—that is, when the linear model used for analysis is not the true data-generating process. This makes the problem more realistic, as standard optimality criteria may fail when the model is wrong.

The announcement is brief, but the title and abstract indicate the focus is on measurement-constrained settings. The paper appears to propose a method for choosing the subdata, though the abstract does not provide details on the algorithm or results. Readers interested in efficient data collection and robust design may find it relevant.