The interconnecting links (edges) between individuals (nodes) in an animal social network are often defined by discrete, directed behaviours (interactions). However, where interactions are difficult to observe, a network edge is instead defined as individuals sharing space or overlapping in time (an association). Despite an increasingly accessible toolkit to assemble and analyse animal social networks, defining associations remains a challenge in behavioural ecology. While different study systems have used different ways to validate the definition of an association, an empirical comparison of how these different methods compare is lacking. Here, we apply three methods to define social associations, by 1) strict time-window, 2) co-occurrence in a group, and 3) arrival-time, in four bird systems. We first test the ability of each method to detect individually repeatable social traits. Then we describe the structure of each network using Jaccard similarity and Mantel tests, and finally, we test the sensitivity of network structure to changing parameters within the three definitions. We found that the network structure was largely robust to changing how associations were defined, with subtle differences. We suggest that these differences are the result of an inappropriate definition of association in the context of experimental design and system ecology. Researchers in ecology and evolution should carefully consider the biological relevance of association definition prior to starting research into animal social behaviour.