Abstract
This article is concerned with the choice of structural prior density for use in a fully Bayesian approach to pedigree inference. It is found that the choice of prior has considerable influence on the accuracy of the estimation. To guide this choice, a scale invariance property is introduced. Under a structural prior with this property, the marginal prior distribution of the local properties of a pedigree node (number of parents, offspring, etc.) does not depend on the number of nodes in the pedigree. Such priors are found to arise naturally by an application of the Minimum Description Length (MDL) principle, under which construction of a prior becomes equivalent to the problem of determining the length of a code required to encode a pedigree, using the principles of information theory. The approach is demonstrated using simulated and actual data, and is compared to two well-known applications, CERVUS and COLONY.