Can Large Language Models Infer Causation from Correlation?