Page 187 -

184 N. David et al.

for explaining the functioning of the target system. Nevertheless, retrodiction alone
is not sufficient to assess the validity of the candidate explanations:

– Underdetermination: Given a model able to explain a certain record of behaviours
or historical data, there will always be a different model yielding a different
explanation for the same record.
– Insufficient quality of data: In many cases it is impossible to obtain long historical
series of social facts in the target system. In the social sciences the very notion
of social facts or data is controversial, can be subjective, and is not dissociable
from effects introduced by the measurement process. Moreover, even when data
is available it may not be in a form suitable to be matched to the bulk of data
generated by simulation models.
Underdetermination and insufficient data suggest the crucial importance of
domain experts for validating the mechanisms specified in the model. A model
is only valid provided that both the generated outcomes and the mechanisms
that constitute the model are sanctioned by experts in the relevant domain. The
importance of validating the mechanisms themselves leads us to the structural
validity of the model, which neither predictive nor retrodictive validity is able to
assess alone.

9.3.2.3 Validity Through Structural Similarity

In practice, the evaluation of a simulation includes some kind of prediction and
retrodiction, based on expertise and experience. Given the implementation of micro-
level mechanisms in the simulation, classes of behaviour at the macroscopic scale
are identified in the model and compared to classes of behaviour identified in the
target. Similarly, known classes of behaviour in the target system are checked
for existence in the simulation. The former case is generally what we call the
“surprising” character of simulations in which models show something beyond what
we expect them to. However, only an assessment of the model from various points
of view, including its structure and properties on different grains and levels, will
truly determine whether it reflects the way in which the target system operates. For
instance, do agents’ behaviour, the constituent parts and the structural evolution of
the model match the conception we have about the target system with satisfactory
accuracy? These are examples of the elements of realism between the model and the
system that the researcher strives to find, which requires expertise in the domain on
the part of the person who builds and/or validates the model.

9.3.3 Validation Techniques

In this section we describe validation techniques used in social simulation. Some
are used as common practices in the literature and most of the terminology has