A1464
Title: On the inadmissibility of a well-known population genetic estimate and its improvement using shrinkage
Authors: Andreas Futschik - JKU Linz (Austria) [presenting]
Abstract: An important goal in population genetics is to infer key parameters describing population history from DNA sequence samples, notably the scaled mutation and recombination rates. Historically, such parameters have been estimated from summary statistics, with unbiasedness given high priority, whereas more recent likelihood-based and machine-learning approaches are often computationally demanding. It is shown how shrinkage can improve common estimators. First, work is presented demonstrating that the popular Ewens-Watterson estimator is inadmissible under the Wright-Fisher model and can be uniformly improved by shrinkage, and then investigation is made into whether other estimators, such as the maximum likelihood estimator, can also be improved. An explanation is provided for why shrinkage is helpful within a framework involving Poisson processes over random intervals, and then the potential gains from shrinkage are considered. Finally, applications to next-generation sequencing data are discussed, where sequencing errors and cost-saving pooling designs must be accounted for.
