The Challenge of Using LLMs to Simulate Human Behavior: A Causal Inference Perspective