Page 234 -

232 R. Seri and D. Secchi

of the simulation outcome on the configurations of parameters can then be tested in
an ANOVA framework. An example of this type of ABM can be a model of intra-
organizational bandwagons (Secchi and Gullekson 2016), in which authors run the
simulation multiple times in order to test propositions as guide for future, probably
empirical, research.
Another distinction that might be helpful when considering the number of runs
can be drawn between models that strive at defining abstract and simple rules
of behavior for their agents and those that are more concerned with describing a
particular aspect of reality with fine degrees of details. This is the divide between
the KISS (“Keep it Simple, Stupid!”) and the KIDS (“Keep it Descriptive, Stupid!”)
principles (Edmonds and Moss 2005). While advocates of the first approach are
in line with modeling efforts of the past (Troitzsch 2017; Coen 2009), those who
indicate that ABM opens a new way stand with a more descriptive and complex
approach to modeling (Edmonds and Moss 2005). If we take that the extreme for a
KISS model is a system of deterministic equations and, on the other side, the bound
for a KIDS model is the attempt to replicate reality, there is an entire spectrum of
models (and ABMs) falling between these two extremes. However, considerations
on the number of runs are more likely to become relevant as modelers tend toward
increased complexity, without reaching the extreme of full description.
In summary, the determination of the number of runs in a simulation is warranted
every time the researcher is seeking to measure—with some degree of confidence—
whether different configurations of parameters are more or less likely to affect the
outcome.

11.3 Testing Theory: Controlling for Alpha and Beta

In this section, we provide an introduction to testing theory that can be read
independently from the rest of the paper. As an example, in this section, the term
parameter denotes an unknown characteristic of a population, and not a quantity
whose value is fixed before data are collected, as customary in ABMs. Therefore, we
suppose that the researcher has identified some parameters describing the behavior
of the population from which the data have been sampled (as a trivial example, the
mean and the variance of the population). We also assume that he/she has formulated
a null hypothesis H 0 , i.e. an assertion about the value of the parameters.
The original approach to testing, pioneered by K. Pearson and theorized by R.A.
Fisher, looks for a statistic T with the following property: when the hypothesis H 0 is
verified, the value t that the statistic T assumes in the sample is near to a fixed value,
generally identified with 0. Therefore, small values of t appear to bring support to
the null hypothesis H 0 , while extreme values of t are seen as witnessing a possible
violation of H 0 . This explains why the most sensible summary of the test, in Fisher’s
approach, is the p-value, i.e. the probability of observing, under H 0 , values of T that