Please join us for the Statistics Seminar scheduled from 11:30 a.m. to 12:30 p.m. on Friday, November 22.
Speaker: Reza Drikvandi, Durham University, UK
Title: Enriched estimation and inference for high dimensional data
Abstract: With the advent of big data era, data with high dimensions are becoming ubiquitous. In the last two decades regularization approaches have become the methods of choice for analyzing high dimensional data. However, obtaining accurate estimates and reliable inference is challenging in high dimensional situations especially when the dimension is very large or when the data is not sparse. In this talk, I introduce an enriched method for estimation and inference after variable selection in high dimensional regression models. The novelty and advantage of this method is in exploiting the relevant information of the unselected covariates sacrificed during the high dimensional selection to enrich the selected model. In addition to modeling strong signals via the selected covariates, the enriched method uses a supervised SVD procedure for modeling weaker signals via the leading supervised principal components of unselected covariates. The enriched method enables us to obtain reliable inference for regression parameters which accounts for both the selection outcome and the supervised selection of principal components. I present theoretical guaranties for this method and show some simulations and real data example to evaluate the performance of the enriched approach for estimation and inference in comparison with recent methods including debiased lasso, conditional selective inference, decorrelated score test, supervised PCR and multi-sample splitting method.
Meeting ID: 990 2474 2995
Passcode: 18628504
