Understanding the affective reasoning behind empathetic responses is critical for developing reliable speech-emotion dialogue systems, yet current Large Language Models remain black boxes. To address this, we propose EmoDial-Reason, a novel dataset in which each example is paired with two reasoning paths: free-form reasoning in which the model “thinks aloud” without constraints, and template-guided reasoning that follows a structured cognitive pathway. Upon this, we explore whether explicit reasoning helps and, if so, which reasoning style yields the most benefits. Our findings emphasize that (1) explicit affective reasoning consistently enhances performance and transparency, (2) template-guided reasoning excels on easier scenarios whereas free-form reasoning is superior on complex situations, and (3) a hybrid approach that enables the model to dynamically select between template-guided and free-form reasoning achieves the best overall results. \(^1\) (Our dataset is available at https://huggingface.co/datasets/ZoeTang/EmoDial-Reason