Accuracy and Precision of Social Relationship Indices

Abstract

Combining interaction rates of different social behaviours into social relationship indices to represent the structure of dyadic relationships on one underlying dimension is common practice in animal sociality studies. However, the properties of these relationship indices are not well explored – mainly because, for real-world social systems, the ‘true’ value of relationships is unobservable. Here, we use simulation studies to estimate the accuracy and precision of three relationship indices: the Dyadic Composite Sociality Index, the Composite Relationship Index, and the Dynamic Dyadic Sociality Index. We simulated one year of social interactions for multiple groups of 25 individuals and 4 interaction types with different properties, and tested the impact of different focal follow regimes, data densities and sampling conditions on the representation of social relationships. Accuracy and precision of social relationship indices were strongly driven by sample size, similar to simple interaction rates. Under the assumption that there was a clear, one-dimensional relationship underlying interactions, and that different interaction types constituting an index were highly correlated, indices indeed increased accuracy over single interaction rates for small sample sizes. Including uninformative constituting behaviours (i.e., those not highly correlated with the underlying relationship dimension) reduced the accuracy of all indices. The precision of each index (i.e., whether multiple simulated focal follow regimes achieve the same dyadic values for the same data) was generally poor and was driven by the precision of the least precise constituting behaviour, making them less precise than some single interaction rates. Our results showed that social relationship indices do not remove the need to have sufficient data for each individual constituting interaction type. Index quality was defined by the least accurate and precise constituting interaction type. Indices might only be useful if all constituting interaction rates are highly correlated and if there are clear indications that one dimension is sufficient to represent social relationships in a group.