Back To Index Previous Article Next Article Full Text


Statistica Sinica 22 (2012), 1611-1637

doi:http://dx.doi.org/10.5705/ss.2009.191





SUFFICIENT DIMENSION REDUCTION IN REGRESSIONS

WITH MISSING PREDICTORS


Liping Zhu$^1$, Tao Wang$^2$ and Lixing Zhu$^2$


$^1$Shanghai University of Finance and Economics
and $^2$Hong Kong Baptist University


Abstract: Existing sufficient dimension reduction methods often resort to the complete-case analysis when the predictors are subject to missingness. The complete-case analysis is inefficient even with the missing completely at random mechanism because all incomplete cases are discarded. In this paper, we introduce a nonparametric imputation procedure for semiparametric regressions with missing predictors. We establish the consistency of the nonparametric imputation under the missing at random mechanism that allows the missingness to depend exclusively upon the completely observed response. When the missingness depends on both the completely observed predictors and the response, we propose a parametric method to impute the missing predictors. We demonstrate the estimation consistency of the parametric imputation method through several synthetic examples. Our proposals are illustrated through comprehensive simulations and a data application.



Key words and phrases: Central subspace, missing at random, missing predictors, nonparametric imputation, sliced inverse regression, sufficient dimension reduction.

Back To Index Previous Article Next Article Full Text