No bio entered.
Schelling famously proposed an extremely simple but highly illustrative social mechanism to understand how strong ethnic segregation could arise in a world where individuals do not necessarily want it. Schelling’s simple computational model is the starting point for our extensions in which we build upon Wilensky’s original NetLogo implementation of this model. Our two NetLogo models can be best studied while reading our chapter “Agent-based Computational Models” (Flache and de Matos Fernandes, 2021 [forthcoming]). In the chapter, we propose 10 best practices to elucidate how agent-based models are a unique method for providing and analyzing formally precise, and empirically plausible mechanistic explanations of puzzling social phenomena, such as segregation, in the social world. Our chapter addresses in particular analytical sociologists who are new to ABMs.
In the first model (SegregationExtended), we build on Wilensky’s implementation of Schelling’s model which is available in NetLogo library (Wilensky, 1997). We considerably extend this model, allowing in particular to include larger neighborhoods and a population with four groups roughly resembling the ethnic composition of a contemporary large U.S. city. Further features added concern the possibility to include random noise, and the addition of a number of new outcome measures tuned to highlight macro-level implications of the segregation dynamics for different groups in the agent society.
In SegregationDiscreteChoice, we further modify the model incorporating in particular three new features: 1) heterogeneous preferences roughly based on empirical research categorizing agents into low, medium, and highly tolerant within each of the ethnic subgroups of the population, 2) we drop global thresholds (%-similar-wanted) and introduce instead a continuous individual-level single-peaked preference function for agents’ ideal neighborhood composition, and 3) we use a discrete choice model according to which agents probabilistically decide whether to move to a vacant spot or stay in the current spot by comparing the attractiveness of both locations based on the individual preference functions.
Both models simulate prosocial and proself agents – using a stochastic threshold algorithm with reinforcement learning components – playing a n-person prisoner dilemma in groups (left figure). Prosocials have a lower threshold than proselfs. Which type resides in the groups shows to be an important factor group success (fixed group model).
We introduce a dynamic feature in the second model, namely: after a pre-set number of ticks, agents can meritocratically match to other groups. We vary information about agents’ merit; their prior cooperative actions. Especially prosocial agents who reside in defective groups – i.e. bad barrels – experience difficulties to harvest their cooperative potential when information about agents’ merit is incomplete (de Matos Fernandes, Flache, & Dijkstra, forthcoming).
Our model studies a way out for prosocial agents. Notably, we incorporate a multidimensional structure in which individuals are embedded in both a group (n-person PD) and a social network (2-person PD). A consequence is that agents do not have only access to merit in the group context but also access to information derived from dyadic interactions with network partners. We model a random and homophilous social network via an random spatial graph algorithm on top of the groups (right figure).