The scientific study of animal welfare involves measuring physiological, behavioural, and/or cognitive variables to infer the welfare state of animals. Such an approach implies these measures are indicators, or reflect, an unmeasured latent variable of welfare state. Drawing inspiration from recent developments in human psychology and psychiatry, in this paper we propose an alternative perspective in the form of a network theory of animal welfare. This theory posits that there is no latent variable; rather, welfare is a network system of causal interactions between and within behavioural, physiological, and cognitive components. We then describe a statistical network modelling approach motivated by network theory, in which welfare-related response variables are associated with each other after controlling for all other variables measured. In three examples using simulated data, we demonstrate how this approach can be used, and the sort of novel insights it can bring. These examples cover a range of species and research questions, which network analysis is well suited to address. We believe a network approach to animal welfare science holds promise for developing our understanding of the concept of animal welfare, as well as producing practical and meaningful information to improve the welfare of animals.
Pain associated with chronic health conditions in non-human animals is an important animal welfare issue. To identify animals in pain and develop an understanding of the mechanisms by which pain affects behaviour, it is therefore important to establish the direct behavioural effects of painful health conditions. We reanalyse data from a cross-sectional survey that considered the presence or absence of a painful condition in dogs and quantified their affective predispositions using the Positive and Negative Activation Scale (PANAS). By applying ideas from network theory, we conceptualise pain as a stressor that exerts direct effects on a network of interacting behavioural variables, and subsequently estimated a network model of conditional dependence relations. Painful health conditions were positively conditionally associated with age (posterior mean partial correlation, ρ = 0.34; standard deviation [SD]=0.05), and negatively conditionally associated with the item 'your dog is full of energy' (ρ = -0.14; SD=0.06). In turn, the energy item was conditionally associated with other PANAS items which were marginally associated with pain, such as items representing ease of excitability and persistence in play. This suggests these marginal effects might be indirectly mediated via the energy item. Further, utilising the posterior predictive distribution we estimated that the median conditional probability (95% credible interval) of a painful health condition given an answer of 'strongly agree' on the energy item was 0.08 (0.05, 0.11), which increased to 0.32 (0.09, 0.58), given a response of 'strongly disagree'. This provides a potentially clinically useful interpretation of the conditional dependencies detected in the network.
Behavioural problems in dogs are a major welfare concern for both dogs and owners, which can lead to relinquishment or euthanasia. As such, it is important to find ways to minimize the risk of development or severity of problematic behaviour. Such interventions should ideally occur early in the dog’s development and need to be widely and easily accessible. One way to implement such interventions, given modern smart phone use, is via an educational application (app). Here, we conducted a cross-sectional observational study where we sought to estimate associations between use of the Zigzag dog-training app (Zigzag Pet Care Services Ltd.) and the development and severity of problematic behaviours. We focused on a subset of the survey population (n = 367) who did no training at all (n = 194) or only used the app (did not attend formal classes or in person training; n = 173). Proportional odds ordinal logistic regression models were used to estimate odds ratios and 95% confidence intervals for a range of behaviours as a function of percentage completion of the first four chapters of the app while controlling for age, sex, health, and where the puppy was from. While sample sizes were relatively small and confidence intervals wide, 19 out of 21 odds ratio point estimates were in favour of severity being lower with increasing Zigzag app completion (the remaining 2 were marginally above 1). There was no good evidence that Zigzag increased the severity of any behaviours. On the contrary, there was reasonable evidence against the null hypothesis of no association in favour of Zigzag reducing the severity of familiar aggression, house soiling, chewing, barking, escaping, and noise fear. While causal effects cannot be claimed, overall, the reported associations are favourable and suggest that further study is warranted.