From empirical to hypothetical: A model for political opinions calibrated by SAOM
Kieran Mepham, Christoph StadtfeldEmpirical network analysts and agent-based modellers alike have recently called for advances in bridging of these two fields. In our current work, we aim to take steps towards this by using a SAOM as an empirical calibration for an agent-based model of political attitude and social clustering. We model multiple political attitudes, treating them as a valenced two-mode network of individuals and topics. Then, we model the coevolution of this network with the friendship network between individuals. This model is based on a previous project applying SAOMs, in which we collected and analyzed data in the scope of the Swiss StudentLife Study, estimating the strength of micro processes of selection and influence on individuals' choices, and how these ultimately related to meso-level structural outcomes. We use this model as the basis for forward simulation.
As a consequence of the SAOM framework applied, our agent-based model assumes one-to-one influence with randomly distributed sequential updates. Selection of friends is modelled as a stochastic process, where each additional shared opinion contributes equally to a linear predictor of a logistic function, weighting the relative chances of adjusting one (potential) outgoing tie over another. Similarly, the choice to adopt a political opinion is modelled through a linear predictor which increases equally for each additional friend who already holds this opinion. Additional parameters model unobserved ties and contextual factors which may drive convergence and divergence between individuals, endogeneity in the friendship network, and friendship homophily on demographic characteristics. We vary the focal selection and influence parameters and starting network configuration to examine their consequences, and trial various metrics to see how these changes affect the consequences of micro-processes in the long run. We also offer some points on conceptual differences between this approach and more traditional agent-based models.