Mitigating Propensity Bias of Large Language Models for Recommender Systems