Functional linear regression for discretely observed data: from ideal to reality
Authors:
- 28 September 2022
- journal article
- research article
- Published by Oxford University Press (OUP) in Biometrika
- Vol. 110 (2) , 381-393
Abstract
Summary: Despite extensive studies on functional linear regression, there exists a fundamental gap in theory between the ideal estimation from fully observed covariate functions and the reality that one can only observe functional covariates discretely with noise. The challenge arises when deriving a sharp perturbation bound for the estimated eigenfunctions in the latter case, which renders existing techniques for functional linear regression not applicable. We use a pooling method to attain the estimated eigenfunctions and propose a sample-splitting strategy to estimate the principal component scores, which facilitates the theoretical treatment for discretely observed data. The slope function is estimated by approximated least squares, and we show that the resulting estimator attains the optimal convergence rates for both estimation and prediction when the number of measurements per subject reaches a certain magnitude of the sample size. This phase transition phenomenon differs from the known results for the pooled mean and covariance estimation, and reveals the elevated difficulty in estimating the regression function. Numerical experiments, using simulated and real data examples, yield favourable results when compared with existing methods.Keywords
Funding Information
- National Natural Science Foundation of China (11931001, 11871080)
- National Key R&D Program of China (2020YFE0204200)
This publication has 23 references indexed in Scilit:
- Minimax and Adaptive Prediction for Functional Linear RegressionJournal of the American Statistical Association, 2012
- Uniform convergence rates for nonparametric regression and principal component analysis in functional/longitudinal dataThe Annals of Statistics, 2010
- p-Values for High-Dimensional RegressionJournal of the American Statistical Association, 2009
- Consistency of restricted maximum likelihood estimators of principal componentsThe Annals of Statistics, 2009
- Prediction in functional linear regressionThe Annals of Statistics, 2006
- Properties of principal component methods for functional and longitudinal data analysisThe Annals of Statistics, 2006
- On Properties of Functional Principal Components AnalysisJournal of the Royal Statistical Society Series B: Statistical Methodology, 2005
- Functional linear regression analysis for longitudinal dataThe Annals of Statistics, 2005
- Functional Data Analysis for Sparse Longitudinal DataJournal of the American Statistical Association, 2005
- The shrinkage of the coefficient of multiple correlation.Journal of Educational Psychology, 1931