Data from: Is BAMM flawed? Theoretical and practical concerns in the analysis of multi-rate diversification models

BAMM (Bayesian Analysis of Macroevolutionary Mixtures) is a statistical framework that uses reversible jump MCMC to infer complex macroevolutionary dynamics of diversification and phenotypic evolution on phylogenetic trees. A recent article by Moore and coauthors (MEA) reported a number of theoretical and practical concerns with BAMM. Major claims from MEA are that (1) BAMM's likelihood function is incorrect, because it does not account for unobserved rate shifts; (2) the posterior distribution on the number of rate shifts is overly sensitive to the prior; and (3) diversification rate estimates from BAMM are unreliable. Here, we show that these and other conclusions from MEA are generally incorrect or unjustified. We first demonstrate that MEA's numerical assessment of the BAMM likelihood is compromised by their use of an invalid likelihood function. We then show that “unobserved rate shifts” appear to be irrelevant for biologically-plausible parameterizations of the diversification process. We find that the purportedly extreme prior sensitivity reported by MEA cannot be replicated with standard usage of BAMM v2.5, or with any other version, when conventional Bayesian model selection is performed. Finally, we demonstrate that BAMM performs very well at estimating diversification rate variation across the ∼20% of simulated trees in MEA's dataset for which it is theoretically possible to infer rate shifts with confidence. Due to ascertainment bias, the remaining 80% of their purportedly variable-rate phylogenies are statistically indistinguishable from those produced by a constant-rate birth-death process and were thus poorly-suited for the summary statistics used in their performance assessment. We demonstrate that inferences about diversification rates have been accurate and consistent across all major previous releases of the BAMM software. We recognize an acute need to address the theoretical foundations of rate-shift models for phylogenetic trees, and we expect BAMM and other modeling frameworks to improve in response to mathematical and computational innovations. However, we remain optimistic that that the imperfect tools currently available to comparative biologists have provided and will continue to provide important insights into the diversification of life on Earth.贝叶斯宏观演化混合分析（Bayesian Analysis of Macroevolutionary Mixtures，以下简称BAMM）是一种统计框架，采用可逆跳变马尔可夫链蒙特卡洛（reversible jump Markov Chain Monte Carlo，可逆跳MCMC）方法，以推断系统发育树（phylogenetic tree）上复杂的宏观演化动态，涵盖物种分化与表型演化过程。摩尔及其合作者近期发表的一篇文章（MEA）对BAMM提出了多项理论与实践层面的质疑。MEA的核心主张包括：（1）BAMM的似然函数（likelihood function）存在错误，因其未考虑未观测到的速率转移（rate shift）；（2）速率转移数量的后验分布（posterior distribution）对先验分布（prior）过度敏感；（3）BAMM得到的分化速率估计结果不可靠。 本研究表明，MEA得出的上述及其他结论大多不正确，或缺乏合理依据。首先，MEA对BAMM似然函数的数值评估存在缺陷，原因是其使用了无效的似然函数。其次，针对分化过程的生物学合理参数化场景而言，所谓“未观测到的速率转移”似乎并不相关。研究发现，当采用标准贝叶斯模型选择（Bayesian model selection）流程时，MEA所称的极端先验敏感性无法通过BAMM v2.5或其他任何版本的标准使用方式复现。 最后，我们证明，在MEA数据集里约20%的模拟系统发育树中，BAMM能够很好地估计分化速率变异——这类树在理论上可以可靠地推断出速率转移。由于检出偏倚（ascertainment bias），MEA剩余80%的所谓可变速率系统发育树，在统计学上与恒定速率生灭过程（constant-rate birth-death process）生成的树无法区分，因此并不适合其性能评估中使用的汇总统计量。 我们还证明，在BAMM软件所有主要过往版本中，分化速率的推断结果始终准确且一致。我们认识到，亟需针对系统发育树的速率转移模型的理论基础展开研究，并期待BAMM及其他建模框架能够随着数学与计算创新而不断完善。然而，我们仍持乐观态度：当前比较生物学家所使用的这些尚不完善的工具，已经并将继续为地球生命的分化研究提供重要见解。