The Shimodaira–Hasegawa test (SH;   [38,39]), which is also based on a non-parametric bootstrap, was developed to   compare multiple topologies. For this test, the null hypothesis (H0) is that   all the trees tested are equally good explanations of the data, and the   tested hypothesis (H1) is that one or several trees are better approximations   of the data. The SH test avoids for the violation of a priori tree selection   by using a step that accounts for the contribution of the maximum likelihood   tree to the null distribution. In addition, this test is also one-sided. The   SH test can be very conservative in its rejection of the null hypothesis when   the number of candidate trees is large [40–42]. This bias can be ameliorated   by a weighted SH test [38,42]. Shimodaira [41] has also suggested the use of   a test based on bootstrap replicates of different sizes to calculate the null   distribution called the Approximately Unbiased (AU) test. This test appears   to be less biased than the KH and SH tests. A similar test was proposed by   Zarkikh and Li [43] but has been shown in simulation studies to be less   accurate than the AU test [41]. More recently, Susko [44] presented a   technique that is appropriate for maximum likelihood distances, based on a   generalized least squares (GLS) metric. This technique can use a   non-parametric bootstrap or a variance estimated from the sample average for   derivation of the correct, null distribution. This test should alleviate some   of the bias in tree topology testing as long as the correct substitution   model is selected

Shimodaira–Hasegawa检验（SH; [38,39]）也是基于非参数自举检验，旨在比较多种拓扑结构的异同。在这一检验中，零假设（H0）是假设所有被检验的系统发育树对数据的解释都是相同的，备择假设（H1）是一个或多个树更加贴合数据。SH检验通过使用一个最大似然树对空分布的贡献度的步骤（脚注①），从而避免先验树（操作人员自己选择标准树）选择带来的偏差。此外，这种检验也是单侧的。当候选树的数目很大时，SH检验可以非常保守地拒绝无效假设[40-42]。这种偏差可通过加权SH检验加以改善[38,42]。Shimodaira[41]还建议使用基于不同大小的自举检验重复来计算空分布，称为近似无偏（AU）检验。这个检验似乎比KH和SH检验的偏差更小。Zarkikh和Li[43]提出了一个类似的检验，但在模拟研究中显示，该检验的精确度低于AU检验[41]。最近，Susko[44]提出了一种基于广义最小二乘矩阵（GLS）的适用于最大似然距离的技术。此技术可以使用非参数自举检验或从样本均值估计的方差来推导正确的零分布。只要选择正确的替换模型，这个检验就应该减轻树拓扑检验中的一些偏差。