Network analysis of tail-biting in pigs : the impact of missed biting events on centrality parameters
With social network analysis, group structures of animals can be studied. However, underlying behavioural observations face problems of missing events or deviations between observers. The current study analysed the robustness of node-level network parameters based on tail-biting observations in pigs affected by missed events. Real observations of one observer were used as a gold standard to build true networks and to compare two sets of erroneous networks to them. The first set consisted of networks from different observers of the same data basis. The second set consisted of networks with a fixed error rate (random samples of the gold standard). The stability of the ranking order was used as an indication of accuracy (range 0–1; ≥0.49 good accuracy; ≥0.81 very good accuracy). Comparing observers with true networks yielded overall bad accuracy scores. Generally, outgoing network parameters (active: biting) provided better accuracy scores than ingoing network parameters (passive: being bitten). The results of sampled networks showed decreasing accuracy scores with increasing error rates. At the same error rate, longer observation periods yielded better accuracy scores. For sampled networks, differences between outgoing and ingoing network parameters were more distinct and local parameters (direct contacts) provided better accuracy scores than global parameters (direct and indirect contacts). Overall, sampled networks with 3/10 missed events yielded good to very good accuracy. As networks with more observations handle missed events better, studies of behavioural observations always need to evaluate the required accuracy and feasible workload. The current study gives insights in the accurate estimation of behavioural observations.