Efficient evaluation of treatment effects in the presence of missing covariate values.