Abstract

The introduction of computer simulation to paleobiology ushered in a new, experimental style of reasoning. Rather than starting with observed fossil patterns and hypothesizing causal processes that may have produced them, it became possible to start with a process model, and from it to simulate a range of possible patterns. ;The MBL Model is a stochastic model of phylogenetic evolution . Computer simulations conducted with the MBL Model served as thought experiments in stochastic evolution. In the MBL work, similarities between empirical and stochastically simulated clades mounted a visual argument that stochastic processes potentially explained large fluctuations in diversity. Stanley and others countered that the similarities were an artifact of scaling. The scaling problem may have been obscured by visual bias: measures of clade shape are scale-invariant, but diversification itself is highly scale-dependent. ;Null and neutral models are frequently conflated. Null models generate hypothetical data distributions under conditions that exclude some process of interest to test whether patterns in an actual data distribution provide statistical evidence for that process. A neutral model assumes selective equivalence among all units at a specified hierarchical level in the evolving system. Neutral models are often inappropriate as null models. Pattern recognition presupposes a null model, but tacit null models are subject to persistent cognitive biases, necessitating explicit formulation of null models. Nonetheless, failure to reject the null model should not preclude further investigation of a pattern. ;Often, models are not candidates for falsification, but serve to simulate data for testing concepts or methods. Sepkoski and Kendrick , and Robeck, Maley, and Donoghue used the MBL Model to test whether mass extinction periodicity is an artifact of paraphyly in taxonomic data. Such numerical experiments, which make all assumptions explicit, can serve as templates for localizing points of disagreement. Models may fall into lineages, each generation providing lessons for the next, but often sharing operational assumptions. Artifacts of shared operational assumptions are often a threat to the validity of inferences. Results holding across models should not be accepted unless they withstand rigorous robustness analysis