Much of the parameter space that matters in IR is scarcely observed, and many relevant variables are latent. At the micro level, short instruments, participant fatigue, and few repeated measures per respondent can generate ceiling effects and noisy estimates; at the macro level, rare events are strategically selected into observation. In both domains, data is thinnest where identification would be most informative.

This line of research develops designs and estimators that target those sparse parts of the distribution.