Multimodality

The structure algorithm starts at a random place in parameter space, and then converges towards a mode of the parameter space. (In this context, a mode can be thought of, loosely speaking, as a clustering solution that has high posterior probability.) For some data sets there may be multiple modes⁸, and the current implementation of structure does not normally cross between them in runs of realistic length. This means that different runs can produce substantially different answers, and longer runs will probably not fix this. This is mainly an issue for very complex data sets, with large values of $K$, $K>5$ or $K>10$, say (but see the example of Data Set 2A in Pritchard et al. (2000a)). You can examine the results for $Q$ to get an idea of whether this seems to be happening. A careful analysis of this type of situation was presented by Rosenberg et al. (2001), for a data set where the estimated $K$ was around 19.