Data from: The summary likelihood method and its implementation in the Infusion package

In recent years, simulation methods such as approximate Bayesian computation have extensively been used to infer parameters of population genetic models where the likelihood is intractable. We describe an alternative approach, summary likelihood, that provides a likelihood-based analysis of the information retained in the summary statistics whose distribution is simulated. We provide an automated implementation as a standard R package, Infusion, and we test the method, in particular for a scenario of inference of population size change from genetic data. We show that the method provides confidence intervals with controlled coverage independently of a prior distribution on parameters, in contrast to approximate Bayesian computation. We expect the method to be applicable for at least 6-parameter models, and discuss possible modifications for higher-dimensional inference problems.

近年来，近似贝叶斯计算（approximate Bayesian computation）等模拟方法已被广泛应用于似然难以求解的种群遗传模型的参数推断任务。本文提出一种替代方法——摘要似然（summary likelihood），可对通过模拟得到分布的摘要统计量（summary statistics）中保留的信息开展基于似然的分析。我们将该方法封装为一款标准化的R语言工具包Infusion，并针对该方法开展了测试，尤其聚焦于基于遗传数据推断种群规模变化的应用场景。研究结果显示，与近似贝叶斯计算不同，该方法生成的置信区间具备可控覆盖率，且不受参数先验分布的制约。我们预计该方法至少可适用于6参数的模型，并探讨了面向高维推断问题的可行改进路径。