STA 290 Seminar Series
DATE: Thursday March 1st, 4:10pm
LOCATION: MSB 1147, Colloquium Room
SPEAKER: Kyunghee Han, Post-Doctoral Scholar, Department of Statistics, UC Davis
TITLE: “Errors-in-variables problem in nonparametric additive models”
ABSTRACT: As a real situation in data analysis, we may face non-negligible levels of measurement errors in sample but we usually apply statistical methods as if there was no measurement error. However, the naive approach entails a biased result, so called ``attenuation to null''. This challenge let us refine true signals from data and devise an unbiased estimator which should achieve efficient estimation as well. In this talk, we will discuss errors-in-variables problem in nonparametric additive regression models. There are two representative scenarios of covariate contamination. First, we consider that predictor variables are observed with additive measurement errors. It is called the classical deconvolution problem and we will introduce a kernel method how to apply a deconvolution technique to additive models. We have another case such that measurement errors are asymptotically negligible as sample size increases. The latter usually occurs when predictors are available only after reconstruction from sample, so that it is important to study asymptotic property of the final estimator. We will demonstrate an example of the second scenario with functional additive models and its application to bike sharing data.
KEYWORDS: Nonparametric function estimation, kernel smoothing, additive regression model, deconvolution, errors-in-variables problem, measurement errors.