Abstract:
|
This report defines an optimality criteria for supersaturated designs that is based on an underlying inference methodology, the asymptotically optimal confidence regions for high-dimensional data Van de Geer et al. [The Annals of Statistics, 42(3):1166-1202, 2014], which has been shown to have good theoretical properties, as well as expected coverage on simulated data. The criteria is defined for confidence intervals of single components from a supersaturated design. Supersaturated designs with a better criteria should produce shorter desparsified lasso confidence intervals while maintaining the expected coverage. An optimal condition for single component confidence intervals is also derived.
|