Title: Sparse logistic regression on functional data

Abstract: Motivated by a hemodialysis monitoring study, we propose a logistic model with a functional predictor, called the Sparse Functional Logistic Regression (SFLR), where the corresponding coefficient function is {\it locally sparse}, that is, it is completely zero on some subregions of its domain. The coefficient function, together with the intercept parameter, are estimated through a doubly-penalized likelihood approach with a B-splines expansion. One penalty is for controlling the roughness of the coefficient function estimate and the other penalty, in the form of the $L_1$ norm, enforces the local sparsity. A Newton-Raphson procedure is designed for the optimization of the penalized likelihood. Our simulations show that SFLR is capable of generating a smooth and reasonably good estimate of the coefficient function on the non-null region(s) while recognizing the null region(s). Application of the method to the Raman spectral data generated from the heomdialysis study pinpoint the wavenumber regions for identifying key chemicals contributing to the dialysis progress.