Speaker: Andrew Yiu, University of Oxford
Title: “Semiparametric posterior corrections”
Abstract: Suppose we wish to estimate a finite-dimensional parameter but we don’t want to restrict ourselves to a finite-dimensional model. This is called semiparametric inference. An exciting aspect of this paradigm is that we might be able to leverage state-of-the-art machine learning algorithms to estimate our high-dimensional nuisance parameters and still obtain statistical guarantees (e.g. a 95% confidence interval). To achieve this, however, we will generally need to carefully tailor our inference to the target estimand. This can be problematic for nonparametric Bayesian inference, which focuses on good performance for the whole data-generating distribution, possibly at the expense of low-dimensional parameters of interest. To remedy this, we introduce a simple, computationally efficient procedure that corrects the marginal posterior of our target estimand, yielding a new debiased and calibrated one-step posterior.
A preprint can be found here
This will be a free hybrid seminar. To register to attend virtually, please click here: https://cam-ac-uk.zoom.us/meeting/register/tZIqd-quqjkiHtRsIwSM3Sec0QFfHrtRftXa