Should Graph Convolution Trust Neighbors? A Simple Causal Inference Method