Quantifying the predictability of behaviour: statistical approaches for the study of between-individual variation in the within-individual variance
Quantifying the predictability of behaviour: statistical approaches for the study of between-individual variation in the within-individual variance
复制标题
DOI：
10.1111/2041-210x.12281
复制
发表时间：
2015-01-01
影响因子：
6.6
通讯作者：
Schielzeth, Holger
中科院分区：
文献类型：
--
作者：
Cleasby, Ian R.;Nakagawa, Shinichi;Schielzeth, Holger
1. 动物行为的许多方面在个体之间始终存在差异，从而催生了动物性格研究领域的不断发展。虽然生物学家长期以来一直对个体间变异感兴趣，但个体内变异的作用却很少受到关注。事实上，许多模型假设个体内部变异的程度在个体之间是相同的，尽管事实上个体的变异性可能经常不同。最近，个体内部变异性或可预测性的重要性已在动物行为领域得到认识。然而，对于如何最好地量化它缺乏共识。这种情况反过来又导致了各种不同方法的发展，旨在评估不同个体的可变性或可预测性。在这里，我们回顾了被提议作为个人可预测性代理的指数。然后，我们介绍称为分层广义线性模型（HGLM）和双分层广义线性模型（DHGLM）的现有技术作为量化可预测性的通用工具。 HGLM 和 DHGLM 是随机截距混合模型的扩展，它们利用了可以在单个总体框架内对方差变化和均值变化进行建模的事实。与其他提出的指数相比，(D)HGLM 对个体内残差变异的显式建模可以更有效地利用数据，在不平衡数据集上表现更好，并捕获更多在个体内变异建模中涉及的不确定性。此外，(D)HGLM 产生了群体范围内可预测性变化的估计量，它可以作为跨性状和研究比较的标准化效应大小。我们将此估计量称为 CVP，即可预测性变异系数。这里描述的不同方法和标准化效应大小 CVP 应该为研究动物行为的个体性开辟新的途径。由于对个体变异的正确理解是生态学和进化论许多研究的核心，因此这些方法在动物性格研究及其他领域都有广泛的应用。
1. Many aspects of animal behaviour differ consistently between individuals, giving rise to the growing field of animal personality research. While between-individual variation has long been of interest to biologists, the role of within-individual variation has received less attention. Indeed, many models assume that the extent of within-individual variation is the same across individuals despite the fact that individuals may often differ in their variability. Recently, the importance of within-individual variability or predictability has been recognized within the field of animal behaviour. However, there is a lack of a consensus on how best to quantify it. This situation, in turn, has led to the development of a variety of different methods aimed at assessing how variable or predictable different individuals are. Here, we review the indices that have been proposed as proxies of individual predictability. We then introduce existing techniques called hierarchical generalized linear models (HGLMs) and double-hierarchical generalized linear models (DHGLMs) as general tools for quantifying predictability. HGLMs and DHGLMs are extensions of random intercept mixed models that exploit the fact that variation in variances as well as variation in means can be modelled within a single overarching framework. Explicit modelling of the within-individual residual variation by (D)HGLMs makes more efficient use of the data, performs better on unbalanced data sets and captures more of the uncertainty involved in modelling within-individual variation than other proposed indices. In addition, (D)HGLMs yield an estimator of population-wide variation in predictability, which can serve as a standardized effect size for comparisons across traits and studies. We call this estimator CVP, the coefficient of variation in predictability. The different methods described here and the standardized effect size CVP should open new avenues for studying individuality in animal behaviour. Since sound understanding of individual variation is central to many studies in ecology and evolution, these methods have wide application both in the field of animal personality research and beyond.