Sampling bias is a challenge for quantifying specialization and network structure: lessons from a quantitative niche model
Sampling bias is a challenge for quantifying specialization and network structure: lessons from a quantitative niche model
复制标题
DOI：
10.1111/oik.02256
复制
发表时间：
2016-04-01
期刊：
影响因子：
3.4
通讯作者：
Williams, Neal M.
中科院分区：
文献类型：
--
作者：
Fruend, Jochen;McCann, Kevin S.;Williams, Neal M.
网络方法已经成为在不断变化的世界中理解生态复杂性的流行工具。许多网络描述符直接或间接地与专业化相关，专业化是生态学的核心概念，并以不同的方式衡量。不幸的是，使用现场数据的专业化和网络结构的量化可能会受到抽样效应的影响。先前评估这种抽样效果的研究要么使用未知真实网络结构的现场数据，要么基于完全广义的相互作用模拟抽样。在这里，我们使用一个定量的生态位模型来生成代表广泛专业化的二部网络，并评估了不同网络规模的大量专业化和网络指标的潜在抽样偏差。我们表明，对于物种丰富的网络来说，样本大小是现实的，所有的指标都倾向于高估专业化（低估泛化和生态位重叠）。重要的是，这种抽样偏差取决于真正的专业化程度，对于广义网络来说是最强的。我们表明，用于经验数据的方法可能会错误地表示抽样偏差：模拟广义相互作用的零模型可能会高估偏差，而丰富度估计器可能会严重高估抽样完整性。有些网络指标在同一网络的大小子样本之间几乎没有关系，因此可能通常没有意义。小样本也高估了广义网络中专门化的种间变异。虽然必须开发应对这些挑战的新方法，但我们也确定了在抽样强度上相对公正和相当一致的指标，并确定了准确估计所需的观测数量的临时经验法则。我们的定量生态位模型可以帮助理解网络结构的变化，同时捕获采样效应和生物学意义。这需要将网络科学与基础生态学理论联系起来，并为应用生态学问题提供可靠的定量答案。
Network approaches have become a popular tool for understanding ecological complexity in a changing world. Many network descriptors relate directly or indirectly to specialization, which is a central concept in ecology and measured in different ways. Unfortunately, quantification of specialization and network structure using field data can suffer from sampling effects. Previous studies evaluating such sampling effects either used field data where the true network structure is unknown, or they simulated sampling based on completely generalized interactions. Here, we used a quantitative niche model to generate bipartite networks representing a wide range of specialization and evaluated potential sampling biases for a large set of specialization and network metrics for different network sizes. We show that with sample sizes realistic for species-rich networks, all metrics are biased towards overestimating specialization (and underestimating generalization and niche overlap). Importantly, this sampling bias depends on the true degree of specialization and is strongest for generalized networks. We show that methods used for empirical data may misrepresent sampling bias: null models simulating generalized interactions may overestimate bias, whereas richness estimators may strongly overestimate sampling completeness. Some network metrics are barely related between small and large sub-samples of the same network and thus may often not be meaningful. Small samples also overestimate interspecific variation of specialization within generalized networks. While new approaches to deal with these challenges have to be developed, we also identify metrics that are relatively unbiased and fairly consistent across sampling intensities and we identify a provisional rule of thumb for the number of observations required for accurate estimates. Our quantitative niche model can help understand variation in network structure capturing both sampling effects and biological meaning. This is needed to connect network science to fundamental ecological theory and to give robust quantitative answers for applied ecological problems.