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 by 13\% relative to machine-learning baselines. We provide a comprehensible description of the proposed probabilistic model and the features employed. Further, we describe the construction of a manually annotated test set, which was used in a thorough empirical investigation of the performance of the proposed model.