Page 185 -

182 N. David et al.

description as the goal of simulating social complexity. In truth, what could be
more contentious in assessing the Schelling model is the extreme simplicity used
to describe the domain of social segregation. The descriptive power of agent-based
models (ABMs) makes them suited to model social complexity. Computational
modelling corresponds to a process of abstraction, in that it selects some aspects of a
subject being modelled, like entities, relations between entities and change of state,
while ignoring those that may be considered less relevant to the questions that are of
interest to the model builder. The expressiveness of ABMs allows the researcher to
play with intuitive representations of distinct aspects of the target, such as defining
societies with different kinds of agents, organisations, networks and environments,
which interact with each other and represent social heterogeneity. By selecting
certain aspects of social reality into a model, this process of demarcation makes
agent-based modelling suited to represent sociality as perceived by researchers and
often by the stakeholders themselves.
The descriptive power of simulation is on par with the diversity of ways used
for informing the construction and validation of models, from theoretic approaches
to the use of empirical data or stakeholder involvement. At any rate, measuring
the goodness of fit between the model and real data expressed with data series is
neither the unique nor a typical criterion for sanctioning a model. The very idea of
using a diversity of formal and informal methods is to assess the credibility of the
mechanisms of the model as good descriptions of social behaviour and interaction,
which must be shown to be resilient in the face of multiple tests and methods, in
order to provide robust knowledge claims and allow the model to be open to scrutiny.

9.3.2 Broad Types of Validity

When we speak about types of validity we mean three general methodological per-
spectives for assessing whether a model is able to reproduce expected characteristics
of an object domain: validation through prediction, validation through retrodiction
and validation through structural similarity. Prediction refers to validating a model
by comparing the states of a model with future observations of the target system.
Retrodiction compares the states of the model with past observations of the target
system. Lastly, structural similarity refers to assessing the realism of the structure
of the model in terms of empirical and/or theoretical knowledge of the target
system (see also Gross and Strand 2000). In practice, all three approaches are
interdependent and no single approach is used alone.

9.3.2.1 Validity Through Prediction

Validation through prediction requires matching the model with aspects of the target
system before they were observed. The logic of predictive validity is the following:
if one is using a predictive model—in which the purpose of the model is to predict