Permutation tests for hypothesis testing with animal social network data: Problems and potential solutions.
Permutation tests for hypothesis testing with animal social network data: Problems and potential solutions.
复制标题
DOI：
10.1111/2041-210x.13741
复制
发表时间：
2022-01
影响因子：
6.6
通讯作者：
中科院分区：
文献类型：
--
作者：
排列测试被广泛用于测试动物社交网络数据的零假设，但当排列不能正确模拟预期的零假设时，I型和II型错误率很高。两种常见的置换类型都有局限性。前网络（或数据流）排列可以用来控制“滋扰效应”，如空间，时间或采样偏差，但只有当零假设随机的社会结构。节点（或节点标签）置换检验可以检验包含非随机社会结构的零假设，但仅当滋扰效应不影响观察到的网络时。我们展示了一种解决这些限制的可能解决方案：在进行节点置换测试之前，使用预网络置换来调整每个节点或边的值。我们进行了一系列模拟，以估计由原始数据中的社会或非社会结构的混杂效应引起的错误率。模拟数据集上的回归表明，相对于仅使用节点排列、预网络排列或具有简单协变量的节点排列，这种“双排列”方法不太可能产生较高的错误率，这些方法在至少一组模拟条件下都表现出较高的I型错误。例如，在预网络置换测试的I类错误率超过30%的情况下，双置换的错误率保持在5%。双置换过程提供了一个潜在的解决方案，所产生的问题时，测试与社交网络数据的零假设升高的I型和II型错误率。我们还讨论了可以提供强大推理的替代方法，包括拟合混合效应模型，限制节点排列，测试多个零假设和拆分大型数据集以生成复制网络。最后，我们强调的方式，不确定性可以明确地考虑，并通过分析进行。
Permutation tests are widely used to test null hypotheses with animal social network data, but suffer from high rates of type I and II error when the permutations do not properly simulate the intended null hypothesis. Two common types of permutations each have limitations. Pre‐network (or datastream) permutations can be used to control ‘nuisance effects’ like spatial, temporal or sampling biases, but only when the null hypothesis assumes random social structure. Node (or node‐label) permutation tests can test null hypotheses that include nonrandom social structure, but only when nuisance effects do not shape the observed network. We demonstrate one possible solution addressing these limitations: using pre‐network permutations to adjust the values for each node or edge before conducting a node permutation test. We conduct a range of simulations to estimate error rates caused by confounding effects of social or non‐social structure in the raw data. Regressions on simulated datasets suggest that this ‘double permutation’ approach is less likely to produce elevated error rates relative to using only node permutations, pre‐network permutations or node permutations with simple covariates, which all exhibit elevated type I errors under at least one set of simulated conditions. For example, in scenarios where type I error rates from pre‐network permutation tests exceed 30%, the error rates from double permutation remain at 5%. The double permutation procedure provides one potential solution to issues arising from elevated type I and type II error rates when testing null hypotheses with social network data. We also discuss alternative approaches that can provide robust inference, including fitting mixed effects models, restricted node permutations, testing multiple null hypotheses and splitting large datasets to generate replicated networks. Finally, we highlight ways that uncertainty can be explicitly considered and carried through the analysis.