Leading and contributing to coordinating centers for genetic consortia, I am responsible for resolving various challenges encountered as work progresses. These have included issues such as making algorithms work fast enough for high-throughput work, making approximations accurate enough so that discoveries are not swamped by false positives, and understanding the various ways confounding can affect results - all while respecting the subtle ethical issues of sharing data and results in this setting. When new approaches are successful, communicating them to a wide audience is crucial, so that the field avoids redundancy.
I have long-standing interests in meta-analysis, the combination of results from multiple strata. As well as being practically useful, as seen in its widespread use in systematic reviews as well as in high-throughput genetics, meta-analysis research benefits from clear statistical thinking. Examples include resolving long-standing confusion about fixed- versus random-effects by being clear about estimands, providing clarity on how heterogeneity affects estimation of overall mean effects and how much (or little) exchangeability assumptions assist that estimation. Applied work in meta-analytic settings with limited data has motivated my development of better-calibrated and even exact inference when combining proportions or 2x2 tables.
I have a career-long interest in the unification of methods from frequentist and Bayesian statistical approaches. Traditionally, much of the statistical literature on this topic has been riven by arguments over which approach is 'right', but I have instead found ways to understand useful methods from both points of view. These methods include 'robust' standard errors, conditional logistic regression (via the medium of Andy Warhol), outlier-robust regression, and the humble (yet extremely controversial) two-sided p-value. As well as providing better fundamental understanding of what these methods do and do not provide, my approach also provides many creative ways in which methods can be generalized, to tackle non-standard situations.
As a statistician I am responsible for not only developing new methods, but also helping non-statisticians understand subtle issues that may affect their analyses. Teaching courses is one way to do this, but I develop less formal material too:
