Page 232 -

230 R. Seri and D. Secchi

defining the “steps,” the time, or the interactions within each run through sensitivity
and convergence analysis, for example (Mungovan et al. 2011; Robinson 2014;
Shimazoe and Burton 2013).
The central assumption of what is proposed in this chapter is that the number
of runs in a simulation is often crucial for results to bear some meaning. Of
course, this is not true for all simulations and it depends on scope, nature of the
simulated phenomenon, purpose, and level of abstraction. We specify these aspects
in the following section. For now, it suffices to write that for social simulations
with a strong stochastic component, where emergence and complexity cause results
to differ even within the same configuration of parameters, knowing how many
runs are enough for differences to emerge (or not) becomes an extremely relevant
information. This is where this chapter positions itself.
We first try to indicate—very broadly—what type of simulations this approach
may apply to. Then, mediating from research on sample size determination for the
behavioral sciences (Cohen 1988;Liu 2014), we introduce statistical power analysis
and testing theory. The chapter also takes an agent-based model (ABM) with a
strong stochastic component and provides two examples that show how crucial the
issue is. At the same time, the chapter offers a practical guide on how to conduct the
computation. Implications and concluding remarks follow.

11.2 Scope and Nature of Agent-Based Models

In this chapter we identify a particular sub-group of agent-based models that are
fit for hypothesis testing. In order to frame the following discussion, we propose
a classification of the aims of ABM, with the caveat that the following discussion
may not be general or exhaustive. For a more general classification of the types of
simulation, one may refer to Chap. 3 of this handbook (Davidsson and Verhagen
2017).
Some agent-based models have the purpose of studying the emergent properties
of a system (Anderson 1972; Fioretti 2016). These properties arise when the system
as a whole displays a behavior that is not explicit in its single components, in this
case, the agents. When this is aimed at establishing whether an outcome is possible,
hence the simulation has an exploratory purpose that reflects on theory, then the
visual inspection of the trajectories of the simulated system or the computation of
some descriptive statistics is sufficient to illustrate the existence of an emergent
behavior. For example, Heckbert’s (Heckbert 2013) model of the socio-economic
system in which the ancient Maya civilization developed and disappeared can be
thought of as a simulation of this kind. As a descriptive model, it establishes whether
the conditions set in the model offer reasonable explanations of historical facts.
The study of emergent properties is also linked to another objective of some
agent-based models, namely hypothesis generation (Bardone 2016; Secchi 2015).
A researcher may run a model just to assess whether it is reasonable to suppose
that some variables have an impact on a given outcome. The hypotheses obtained