Nonparametric Regression with Selectively Missing Covariates