Abstract: Variable screening is an effective method for processing ultra-high-dimensional data. Barut et al. considered that some of the known variables are significantly related to the response variables, and propose the CSIS method based on the assumption of a linear model. This method can effectively reduce the probability of false variable selection. However, its linear model assumptions are more stringent. In actual research, the structure of the model cannot be determined in advance. Therefore, this paper proposes a conditional non-parametric independent screening method (CNIS) based on a non-parametric additive model, which does not need to make assumptions about the model structure, to increases the scope of application. At the same time, under appropriate conditions, it is proved that the screening in the first stage of the method has consistent screening properties and can retain important variables with probability 1. The variable selection in the second stage also has good consistency. The simulation results based on Monte Carlo data show that this method has better performance than the NIS method.

Key words: screening, additive model, variable selection, sure screening