All　data　for　training　and　evaluating　recommended　systems　are　subject　to　selection　biases.　In　much　research,　principled　approaches　have　been　found　to　manage　selection　biases　by　adapting　estimation　techniques　and　models　from　causal　inference.　However,　no　matter　what　kind　of　method　is　adopted,　the　problem　of　data　randomly　missing　always　exists.　This　paper　tries　to　discover　whether　the　deeper　model　can　effectively　bring　better　prediction　results　and　debias.　We　theoretically　and　experimentally　examine　whether　the　models　are　robust　or　not.　©　2023　SPIE.