I think, SD might be a sort of top-down modeling approach, with no
direct representation of agent interaction, true. However, it does not
imply that it is intended only for static social structures. An example:
markets (which are also social structures, where "nodes" might be firms
and "links" their transactions, scaled-based distance or so on...)
change if some "global" indicators as the number of organizations (aka
market "density") and market concentration change. A market with high
density and low concentration has a very different structure respect to
a market with both high density and concentration. Consequently, value
changes in those population-level variables indicate changes in the
structure.

Another point is the fact that SD and AB may produce different results
not only because the lack of representation of agent interaction. We
assume that interaction is a significant driver of social structure, but
again, Organizational Ecologists have built an impressive stock of
empirical studies through the simple assumption of selection and
replacement.  Everybody knows that even an AB model may produce
different results depending on how the agent interaction is conceived
and modeled. So, the difference in SD vs. AB results for a specific
social system always indicates that there is a lack of model
interaction, or that sometimes the interaction contributes very little
to the system behavior so that a misspecified interaction just distorts
the real results?

I am not defending SD. It is just that I do not think that AB is the
"master tool" to always study social systems and emergence, same as
linear programming is not the master tool to solve any optimization problem.

All the best,

Cesar

Scott Moss wrote:
> All of the responses so far to my posting seem to me to raise
> interesting and valid points.  Some years ago on this list we had a long
> discussion about modelling top-down or bottom-up.  The modelling
> elements of that discussion were reported as a symposium among Rosaria
> Conte, Bruce Edmonds, myself and Malcolm Sawyer in CMOT.  I don't want
> to revisit the same discussion.
>
> So can we stipulate that SD is a sort of top-down modelling approach
> with a static representation of social structure and no direct
> representation of interaction among socially embedded individuals?
>
> Nearly 10 years ago, Van Parunak demonstrated that for a particular
> social target an SD model yielded different results from those of an
> agent based model.  As I remember, Van ascribed the difference to the
> effect of agent interaction.
>
> I am not aware of any general results in this vein.
>
> If the result were general, then we could define the domain of
> application of SD social models as the set (or a subset) of cases where
> social interaction among individuals does not affect the macro level
> outcome.  Similarly, the domain of application of agent based models
> would be where social interaction does affect macro level outcomes.
>
> If it is possible to capture agent-based macro level results with an SD
> model, then there might well be efficiency gains in using such a model
> where scaling up the number of agents is computationally too expensive
> with available technology.  For this to be a reliable procedure, we
> would need some general results about -- or experience with -- the
> conditions in which macro level outcomes can be represented as if they
> did not emerge from social interaction.
>
> What is not true is the remark that stimulated my original posting.  It
> is false to say that system dynamics is the modelling method of choice
> for social systems.  You can only have confidence in its applicability
> where there is no effective social interaction among individuals and the
> social structure is static.
>
>