Under What Conditions Can Formal Models of Social Action Claim Explanatory Power?

Abstract

This paper’s purpose is to set forth the conditions of explanation in the domain of formal modelling of social action. Explanation is defined as an adequate account of the underlying factors bringing about a phenomenon. The modelling of a social phenomenon can claim explanatory value in this sense if the following two conditions are fulfilled. (1) The generative mechanisms involved translate the effects of real factors abstracted from their phenomenal context, not those of purely ideal ones. (2) The explanatory hypotheses, which account for the effects of explanatory factors, and the purely descriptive hypotheses, which introduce conceptual simplifications and summarise complex secondary mechanisms, are relatively independent from each other with regard to the phenomenon represented. This condition subjects the model to testing by alternatives through the development of purely descriptive hypotheses in the sense of explanatory or analytical realism.

Acknowledgements

The author would especially like to thank the editor and two anonymous referees of ISPS for their very constructive comments and suggestions on an earlier version of this paper.

Notes

[1] See Stinchcombe (1991); Bunge (1997); Hedström and Swedberg (1998), and especially Boudon (1998), and Elster (1998) therein; also Goldthorpe (2000), Mayntz (2004), Cherkaoui (2005), Hedström (2005), and Manicas (2006).

[2] For instance, according to Sugden (2000), the model of inductive inference for explanation is (E₁) in the model world, R is caused by F; (E₂) F operates in the real world; (E₃) R occurs in the real world; therefore, there is reason to believe that (E₄) in the real world, R is caused by F.

[3] While in numerous models used to simulate the outcomes of the composition of multiple social actions, the interdependence between decisions is indirect—i.e. affecting situations and not decision rules as such—in agent‐based models, the effects of interaction between actors tend to influence the decision‐making processes themselves. These ones are then involved in a dynamic process of change, such as occurs with the phenomena of learning or influence. Whereas we speak of agent‐based models as soon as interdependence of actions is involved, the latter approach tends to characterise recent trends of research.

[4] See for example, Margenau (1935, 1950), Harré (1960), Achinstein (1968), Cartwright (1983, 1989), McMullin (1985), Miller (1987), Ellis (1996), Bunge (1997), Bailer‐Jones (2003), Mäki (2003), Mayntz (2004), and Cherkaoui (2005).

[5] Note that causal claims were rejected from the beginning by the supporters of positivism.

[6] This version of the nature of explanation is presented, for example, by da Costa and French (2003, 59). Explanation is supposed to rest on complex series of connections between models that are interwoven and organised into a hierarchy, running from ‘data’ models and ‘phenomenological’ models to their ‘high‐level’ theoretical counterparts. These connections are represented in terms of partial isomorphisms ‘holding between families of partial relations in the structures constituting these models’.

[7] Giere (1988, 85) also adopts a semantic interpretation of theories, according to which a theory comprises two elements: (1) a population of models, and (2) various hypotheses linking those models with systems in the real world. According to the ‘constructive realism’ that he defends, ‘theoretical hypotheses are interpreted as asserting a similarity between a real system and some, but not necessarily all, aspects of a model’ (Giere 1988, 97).

[8] Mill notes that it is not quite correct to say that a body moves in a certain way because of the action of a certain force; unless it is prevented from doing so by another force, it tends to move in the same way even when it undergoes the influence of another force (see Cartwright 1989, 177).

[9] The DBO (desire, belief, opportunity) theory maintained by Hedström (2005) is founded on the assumption that desires, beliefs, and opportunities are the primary theoretical terms upon which the analysis of action and interaction could be based. This theoretical framework is, in our view, of less scope than the concept of cognitive rationality (which more generally supposes motives, cognitive capacities, notably reflexive ones, and situations, all accounting for the actor’s reasons for acting). Making no assumption that actors behave rationally but assuming that they behave reasonably and with intention (Hedström 2005, 61), DBO theory overshadows the human meaning of cognitive rationality and imposes a reductive theoretical framework.

[10] Understood in a less specific sense than the notion of tendency in Mill (1836) or of capacity in Cartwright, where the notion plays the role of the explanans and does not send us back to actual results when different factors interact. The tendency here represents the explanandum, expressed in a theoretical or observational language that makes use of units of analysis generally situated at a higher level than units of the explanans (Stinchcombe 1991).

[11] Note that Margenau speaks of ‘correspondence rules’, instead of ‘epistemic correlations’. But as he observes, ‘logical positivism, in so far as it restricts itself to an analysis of scientific language, can never do complete justice to science; it must forever talk about propositions, whereas the scientist concerns himself with meanings that are prior to propositions’ (Margenau 1950, 57).

[12] Using Mill’s terminology, when causal factors combine ‘mechanically’, the effect of their combination is the same as the total effect of each acting separately. See, for instance, Hausman (2003, 137) and Lawson (1997, 132).

[13] For example, if we ask what can be generalised from a specific experimental result, ‘The answer requires the notion of nature: the features that are necessary are exactly those which, in this very specific concrete situation, allow the nature of the process under study to express itself in some readable way’ (Cartwright 1999, 90).

[14] In reality, tit‐for‐tat never totalises more, and over the long term never less, than the adverse strategy. Nonetheless, any defection by the adversary brings about a reaction of defection by the player that cannot be overcome except through the cooperation of the adversary.

[15] The winning version is based on a starting hypothesis that is optimistic rather than pessimistic, with regard to the reactivity of the other player. It is rational in the case of an iterative game in this sense, that even if a player does not know a priori the profile of the adversary, he or she knows that interaction with him or her will be repeated a potentially infinite number of times.

[16] The procedure is closer to that proposed by Levins (1966), which suggests developing several alternative models to test the core causal structure of a model against the development of the details left out by simplifying assumptions. Nonetheless, what is being discussed here is of greater generality. Hypotheses that are essentially descriptive do not only bring about causal simplifications through ‘magnification’, which correspond to those that Levins had in view (see also Weisberg 2006).

[17] For example, S(q, x) = q ^{(1 − a)} x^a with pq + lx = r, where x stands for consumption of housing translating the quality of the housing; q stands for the importance of other items consumed in the household; a is a measure of the preference for housing; lx stands for the annual charge corresponding to housing; p stands for the price index for other uses for income; and r stands for the household’s income. Satisfaction is supposed to be maximal when the portions of income of a household that are spent on housing and on other consumptions are equal to a and 1 – a, respectively (Lévy‐Lambert 1969).