A social theory which is generally applicable, true in the sense that it describes observed social processes and outcomes and simple is unquestionably desirable. It is equally unquestionable that social science methodology has not successfully produced any such theory.
An alternative methodology based on the emerging approach to empirical, agent based, social simulation modelling starts from qualitative, narrative (some would say, anecdotal) evidence which is used to build detailed, empirical models. The models must be simple enough for the providers of the narratives - stakeholders and other domain experts - to be able to assess their plausibility and accuracy. A step towards simplicity and generality without losing accuracy can always be attempted by abstracting from the detail of the empirical model in a way which, for example, which substitutes abstract instantiations of classes of entities for empirically driven representations of entities. In the example of the companion paper to this, the modeller can substitute a set of abstract activities (activity-1, activity-2, ...) for specific churches, clubs, schools, etc. Then rules determining agent behaviour that depend on specific activities must be combined to specify behaviour activities in general. This step may involve fewer rules if some real activities require or motivate specialised behaviour. And model initialisation is likely to entail fewer lines of programming code to set up the activity entities since all that is required is a number of iterative loops equal to the number of activities being instantiated. There is nothing special about any of these instantiations. In such cases, both enhanced generality and simplicity result. The question of accuracy (or truth) is a different issue.
There is no reason to suppose that the informants who provided the narrative and qualitative accounts will be able to assess models at the resulting level of abstraction. However, if the empirical models have been validated by stakeholders and other domain experts and the abstract models yield the same results such as friendship networks with the same distribution of node degrees, the same patterns in the time series of cluster coefficients and the same time pattern of connectivity, then we can have some confidence (at least we have not disconfirmed) that the empirical model is a special case of a more general model. The model is more general in the sense that, at the very least, its assumptions describe conditions in which there are ranges of activities but not anything very specific about those activities. So the validated abstract model is simpler and, in conditions of application, more general than the empirical model and might also be true. To determine whether it is both more general and true will be a matter for further exloration. If other empirical models can be abstracted in the same way to yield the same model, then it would seem reasonable to have greater confidence that the abstract model is simpler and more general than the empirical models and equally true.
It seems worth noting that the approach to generalisation and validation described here is an alternative to the model-to-model (M2M) comparisons advocated by Edmonds, Rouchier, Hales and others [refs] - though by no means a mutually exclusive alternative. The fact is that model replication, though useful and often showing up the effects of highly special constructions, is not obviously less time consuming than evidence based modelling and may also require detailed interaction with stakeholders - in this case the modellers whose models are to be replicated [Axtell-Epstein-Axelrod-Cohen]. Experience shows that abstraction is easier and more straightforward than replication. If the abstract models are also simpler in the sense of having fewer behavioural rules and lines of code, then it seems likely that M2M comparisons among abstract models will be easier than replications of empirical models. Both M2M analysis and the sort of model abstraction suggested here and demonstrated in the companion paper give us information about models as formal artifacts. As a result, we know more about the sensitivities of the models to particular assumptions and special programming devices. This knowledge is important in forming our own assessments of how well and robustly our models capture the social processes (or stakeholder perceptions of the social processes) under consideration. The further advantage of model abstraction is that, as part of a coherent program of companion modelling, it might lead to a collection of canonical models (Moss, 2000) that instantiate to a wide range of empirical models. The result would be an effective body of theory embodied in general and relatively simple models the goodness or truth value of which could be assessed by their history of validation.