A0256
Title: Anytime valid and asymptotically efficient inference driven by predictive recursion
Authors: Vaidehi Dixit - University of Missouri (United States) [presenting]
Ryan Martin - North Carolina State University (United States)
Abstract: Data peeking is a common problem, and sampling till significance is obtained leads to erroneous conclusions and irreproducibility. Given this criticism, new approaches are desired. E-processes are a relatively new tool that quantify evidence against a null hypothesis, such that optional stopping is allowed, and combining evidence across multiple studies is permissible. Generally, consider comparing two classes of candidate models, where error rate control is desired at any stopping time. The focus is on the novel e-process construction that leverages the so-called predictive recursion (PR) algorithm. The resulting PRe-process gives anytime valid inference uniformly over stopping rules and is shown to be efficient in the sense that it achieves the maximal growth rate asymptotically, under the alternative relative to the mixture model being fit by PR. Specific applications of this methodology are presented, namely testing for monotonicity and log-concavity.
