In this paper we investigate the use of latent variable structured prediction models for fine-grained sentiment analysis in the common situation where only coarse-grained supervision is available. Specifically, we show how sentence-level sentiment labels can be effectively learned from document-level supervision using hidden conditional random fields (HCRFs). Experiments show that this technique reduces sentence classification errors by 22% relative to using a lexicon and 13% relative to machine-learning baselines.