We introduce a Bayesian hierarchical model for the analysis of compositional data, where the source data are observed as indicator measurements rather than direct proportions. We focus on an application of the model to water quality data, where we apportion stream water to three sources.