Nonstandard Inference

Standard Wald approximations can fail after data-driven selection, for nonsmooth or optimization-defined targets, when nuisance functions are learned too slowly, or when uncertainty in an estimated design mechanism must be propagated into the final analysis. My work develops perturbation, resampling, searching, and union-based procedures that provide valid confidence statements in these nonregular settings.

Authorship notation Name supervised student # equal contribution * alphabetical ordering corresponding author