As opposed to standard methods of association which rely on measures of central dispersion, entropic measures quantify multivalued relations. This distinction is especially important when high fidelity models of the sensed phenomena do not exist. The properties of entropic measures are shown to fit within the Bayesian framework of hierarchical sensor fusion. A method of estimating probabilistic structure for categorical and continuous valued measurements that is unbiased for finite data collections is presented. Additionally, a branch and bound method for optimal sensor suite selection suitable for either target refinement or anomaly detection is described. Finally, the methodology is applied against a known data set used in a standard data mining competition that features both sparse categorical and continuous valued descriptors of a target. Excellent quantitative and computational results against this data set support the conclusion that the proposed methodology is promising for general purpose low level data fusion.