Common permutation methods in animal social network analysis do not control for non-independence.

Area of Science:

• Behavioural Ecology
• Social Network Analysis
• Statistical Methods

Background:

• Non-independence in social network data is a significant concern for behavioural ecologists.
• Permutation-based methods, such as node-label permutations and Quadratic Assignment Procedure (QAP), are commonly used to test hypotheses in social network analysis.
• Existing permutation methods may not adequately account for all forms of non-independence inherent in network data.

Purpose of the Study:

• To evaluate the effectiveness of node-label permutations and QAP in addressing non-independence in social network data.
• To compare permutation-based methods with standard parametric regression models.
• To propose alternative statistical approaches for social network analysis.

Main Methods:

• Analysis of node-label permutations for nodal regression.
• Assessment of the Quadratic Assignment Procedure (QAP) for dyadic regression.
• Comparison of permutation test results with parametric regression models.

Main Results:

• Node-label permutations do not automatically correct for non-independence due to assumptions of residual exchangeability.
• QAP only accounts for a specific type of edge non-independence.
• Permutation tests yield similar p-values to parametric models, but parametric models provide more reliable effect size estimates.

Conclusions:

• Standard parametric regression models can effectively handle non-independence in social network data.
• Replacing permutation-based methods with parametric models can reduce over-reliance on p-values.
• Adopting parametric models facilitates more reliable effect size estimation and causal inference in behavioural ecology.