Computer-Intensive Randomization in Systematics

There has been a sort of cottage industry in the development of randomization routines in systematics beginning with the bootstrap and the jackknife and, in a sense, culminating with various Monte Carlo routines that have been used to assess the performance of phylogenetic methods in limiting circumstances. These methods can be segregated into three basic areas of interest: measures of support such as bootstrap, jackknife, Permutation Tail Probability, T‐PTP, and MoJo; measures of how well independent data are correlated in a phylogenetic framework like PCP for coevolution and Manhattan Stratigraphic Measure (MSM) for stratigraphy; and simulation‐based Monte‐Carlo methods for ascertaining relative performance of optimality criteria or coding methods. Although one approach to assessing cospeciation questions has been the randomization of, for example, hosts and parasite trees, it is well established that in questions that are of a correlative type, the association themselves are what should be permuted. This has been applied to Brooks' parsimony analysis previously and here to the recent reconciled tree approach to these questions. Although it is debatable whether the extrinsic temporal position of a fossil can stand as refutation of intrinsic morphological character‐based cladograms, one can, nonetheless, determine the strength and significance of fit of stratigraphic data to a cladogram. The only method available in this regard that has been shown to not be biased by tree shape is the MSM and modifications of that. Another similar approach that is new is applied to evaluating the historical informativeness of base composition biases. Incongruence length difference tests too are essentially correlative in nature and comparing the behavior of “perceived” partitions to randomly determined partitions of the same size has become the standard for interpreting the relative conflict between differently acquired data. Unlike the foregoing, which make full use of the observed structure of the data, Monte Carlo methods require the input of parameters or of models and in that sense the results tend to be lacking in verisimilitude. Nonetheless, these kinds of questions seem to have been those most widely promulgated in our field. The well‐established theoretical proposition that parsimony has problems with adjacent long‐branches was of course illustrated through such methods, much to the concern and angst of systematists. That likelihood later was shown to perform worse than parsimony when those long branches might repel each other has generated less concern and angst. But then many such circumstances can be divined, like the “short‐branch‐mess” problem wherein likelihood has difficulty placing just a single long branch. Overall, then, in the interpretation of these or any other Monte Carlo issues it will be important to critically examine the structure of the modeled process and the scope of inferences that can be drawn therefrom. Modeling situations that are bound to yield results favorable to only one approach (such as unrealistic even splitting of ancestral populations at unrealistically predictable times in examination of the coding of polymorphic data) should be viewed with great caution. More to the point, since history is singular and not repeatable, the utility of statistical approaches may itself be dubious except in very special circumstances—most of the requirements for stochasticity and independence can never be met.