NON-STANDARD PROBLEMS AND MASSIVE DATA SETS
Moulinath Banerjee
University of Michigan, Ann Arbor

We discuss two methods of making inference on parameters in non-standard problems where the estimates under consideration are asymptotically non-Gaussian. The first relies on zoom-in strategies that sample in stages from a large data-base. These allow precise estimates to be constructed by sampling a very small fraction of the available data. The second uses sample splitting in non-standard problems which, in addition to computational efficacy, also often buys precision over an analysis of the entirety of data at one stroke.