We then develop a Bayesian stats model that takes raw behavioural observations as input, and can recover the generative model’s parameters. We finally illustrate how to use our models to ask causal questions (eg effect of sex on behaviour)

Instead, we first develop a generative model describing what pairs of individuals (dyads) do over time. Dyads differ in how long they stay in, and how they transition between behavioural states. This results in varying sequences of behavioural interactions that are structured like empirical data.

To test causal theories about why animal societies look the way they do, we often rely on observations collected as time-series. For example, with focal-animal sampling, observers follow individual animals and record what they are doing, and who they’re interacting with, in continuous time.

New paper!

We propose a framework to empirically study animal social relationships by modelling social network (SN) data as time-series—that is, without the need to aggregate them over time.

www.biorxiv.org/content/10.1...

Now that it has a title page, this all becomes very official

Here, for instance, the graphical rules tell us that that controlling for Sab is necessary to correctly capture the effect of X on y. That is, we control for the “spurious” part of the dependency structure, and interpret the remaining association pattern causally.

Consider now individual-level features (age, personality, genotype) affecting how they interact with others. E.g., the young age X of individual “1” causes it to socialise more across partners 2, 3, 4—resulting in high values for y12, y13, y14 (i.e. in their inter-dependency).

For example, here, variation in sampling effort (Sab) results in yab and yba to be associated, i.e. non-independent.

Note that the graphs do *not* represent the social networks themselves, but the process that generated the social network edges y.

To illustrate this point, let’s use graphs representing causal relationships (arrows) between variables (nodes). General graphical rules tell us which variables are associated given a certain causal structures (e.g., fork, chain), and which variables to control for to block this association.