The concept of causality and quantitative techniques for causal inference have been extensively discussed in sociology by a variety of scholars with different research programs and theoretical perspectives (see, among others, Marini and Singer 1988; Abbott 1998; Doreian 1999; Goldthorpe 2001; Winship and Sobel 2004; Mahoney 2008; Gangl 2010; Mahoney et al. 2013; for a historical perspective, see also Barringer et al. 2013). Among quantitative sociologists, the interest in causality issues was reinforced by the rapid diffusion in sociology of the potential outcome approach to causality (Morgan and Winship 2015), an approach in turn fostered by older contributions in statistics (for a historical overview, see Imbens and Rubin 2015: ch. 2) and economics (Heckman 2005), and reinvigorated by recent developments in computer science (Spirtes et al. 2000; Pearl 2009) as well as philosophical discussions (Woodward 2003). Extensive literature reviews documented that establishing causal claims is now one of the primary goals of many articles published in leading sociological journals, be these quantitative, qualitative, or historical studies (see Abend et al. 2013; Ermakoff 2019).
Traces of the concept of causality can also be found in the literature on mechanism-based thinking (Hedström and Ylikoski 2010) and on agent-based computational models (Bianchi and Squazzoni 2015). And these are also two research topics in rapid expansion in contemporary sociology.
As to mechanism-based explanations, it is true that the effort towards identifying generalizable fine-grained chains of small-scale events with clearly defined large-scale consequences can be traced back to the infancy of modern social sciences (see, for instance, Elster’s 2009b study of Tocqueville’s œuvre). It is also true that the notions of “social mechanism” and “generative model” were initially forged by mathematical sociologists in the 1960s (see, respectively, Karlsson 1958: 16; Fararo 1969a: 225, 1969b: 81, 84–5; see also Boudon 1979). At the same time, it seems correct to regard Hedström and Swedberg’s (1998) collection on social mechanisms, coupled with philosophical studies of research practices in biology and neuroscience (Machamer et al. 2000; Bechtel and Abrahamsen 2005; Craver 2007), as the starting point of systematic investigations on the concept of mechanism-based explanation. As a by-product, analytical sociology has progressively emerged as a distinctive style of social inquiry (see, among others, Hedström 2005; Hedström and Bearman 2009; Demeulenaere 2011a; Manzo 2014a, 2021). Discontents with analytical sociology, in turn, argued that mechanism-based explanations can in fact be framed in different ways (see, among others, Abbott 2007; Gorski 2009; Gross 2009; Sampson 2011; Little 2012; Opp 2013; for a reply, see Manzo 2010, 2014a).
As to agent-based computational models, hereafter ABMs (or ABM, for the singular form and for “agent-based modeling”), the basic principles appeared in pioneering studies in the 1960s (Hägerstrand 1965; Sakoda 1971; Schelling 1971) but their diffusion accelerated after the publication of systematic monographs such as Axtell and Epstein (1996), Axelrod (1997), and Epstein (2006). Nowadays, pleas for ABMs exist in a large variety of disciplines—including biology (Thorne et al. 2007; Chavali et al. 2008), ecology (Grimm et al. 2006), macroeconomics (Farmer and Foley 2009; De Grauwe 2010), quantitative finance (Mathieu et al. 2005), organization and marketing studies (Fioretti 2013), political science (Cederman 2005; de Marchi and Page 2014), geography (O’Sullivan 2008), criminology (Birks et al. 2012), epidemiology (Auchincloss and Diez Roux 2008), social psychology (Smith and Conrey 2007), demography (Billari and Prskawetz 2003) and archeology (Wurzer et al. 2015). Sociology is no exception (Macy and Flache 2009). Leading journals have started paying attention to ABMs (Gilbert and Abbott 2005; Hedström and Manzo 2015) and the number of applications at the core of the discipline is fast increasing (Macy and Willer 2002; Sawyer 2003; Bianchi and Squazzoni 2015).
The starting point of this book is that, although scholarship on causality, mechanisms, and ABMs is burgeoning, a systematic discussion of the conceptual and methodological connections between these three topics is still missing. A few examples suffice to document this fact.
Among social scientists, Hedström and Ylikoski (2010) reflect on the concept of both cause and mechanism, but, when they treat ABMs, the issue of the potential contribution of this method to causal inference is not addressed. Demeulenaere (2011b: 12–20) explicitly studies the connections between the concepts of mechanism and causality—making the important point that it is in fact disputable to regard the former as a substitute for the latter because, he argues, any mechanism, to work as such, must rely on causal regularities, a general argument that he also applies to the more specific problem of the explanations of individual actions (see Demeulenaere 2011c). The implications of this argument for methods for causal inference are not addressed, however. Knight and Winship (2013) criticize the way the concept of mechanism is employed within the analytical sociology literature; they propose a more precise definition of the concept, which they regard as compatible with a counterfactual view of causation, and show how this definition can be employed, using directed acyclic graphs, to identify causal relations; computational methods, however, receive no mention. Watts (2014) reflects on the notion of causal explanation in connection with a critical analysis of a specific aspect of the mechanism-based perspective, namely action theory, but, when he addresses the methodological side of the issue, experimental and statistical methods are only quickly discussed and no attention is devoted to ABMs. Finally, let me mention Gross (2018) who has interestingly argued that mechanism-based explanation with causal ambitions should pay more attention to the formal properties of causal chains but, once again, how this translates, on a methodological level, in specific procedures for causal inference is not discussed.
Philosophical investigations exhibit a similar pattern. Several articles scrutinize the connection between the concepts of causal and mechanistic explanation (Glennan 1996, 2002; Woodward 2002, 2013; Casini et al. 2011; Menzies 2012; Williamson 2013); however, the discussion of what techniques would support the connection between methods for causal inference and strategies for mechanistic explanation is limited, and, in any case, either does not contemplate ABMs at all (for some illustrations, see Steel 2004; Reiss 2009; Mouchart and Russo 2011; Hoover 2012) or only points to this technique without any detailed analysis (see, for instance, Kaidesoja 2021a, b). Analogously, the rare philosophical contributions specifically discussing the potential relevance of ABMs for supporting causal explanations—some denying such relevance (Grüne-Yanoff 2009a), others arguing in favor of it (Elsenbroich 2012; Casini 2014)—only rely on specific models, lack a systematic discussion of theories of causal inference and social mechanisms, and never systematically confront ABMs with other methods for causal inference (for an example of the latter limitation, see Anzola 2020).
In this book, my intention is to show that a comprehensive analysis of causality, mechanisms, and ABMs at the same time has the potential of modifying our understanding of causal inference within quantitative social sciences.
1 The Book’s Question
The absence of systematic discussions on the possible three-way connections between causality, mechanisms, and ABMs left unsolved an important question: in what sense and under which conditions, if any, can ABMs contribute to causal inference? This book has the ambition to provide a principled answer to this question by scrutinizing, and systematically connecting, scholarship on (methods for) causal inference, social mechanisms, and ABMs
Answering this question is complicated by the variety of views that exist on the status of an ABM itself. If many social scientists (see, among others, Hummon and Fararo 1995b; Epstein 1999, 2006: chs. 1–2; Axtell et al. 2002; Sawyer 2004; Cederman 2005; Tesfatsion 2006; Manzo 2014a) and philosophers of social science (Ylikoski and Marchionni 2013: §3) agree on ABM’s ability to act as a methodological lever for mechanism-based explanations, disagreement remains on the extent to which an ABM can also be used to persuade an audience that the postulated mechanism is operating in the real world.
Let us first consider philosophers’ views on this point. Grüne-Yanoff (2009a) studied a famous, empirically grounded ABM of a particular historical phenomenon (the so-called artificial Anasazi model by Jeffrey Dean and collaborators) and, on this specific basis, generally denied that this method can help causal inference. On the one hand, he argued, “full” causal explanations cannot be provided by an ABM because it is not possible to obtain empirical evidence on all the elements constituting the mechanisms designed by the model; on the other hand, Grüne-Yanoff continues, even potential explanations, meaning “possible causal histories”, cannot be supported by an ABM because the method lacks an internal criterion to select among all the possible options.
Casini (2014) defends a different view. He scrutinized two specific highly stylized ABMs aiming at the generation of stylized macroscopic targets, and argued that even hyper-simplified ABMs can in fact produce potential explanations that increase our understanding of the mechanism at work, thus contributing to identifying actual causal forces. This is possible, Casini (2014: 665) claimed, as long as the model is proved to be “credible-with-respect-to-the target”, which, according to him, depends on the following three conditions: “(i) the soundness of theoretical principles, psychological assumptions, and functional analogies; (ii) the robustness of the results across changes in initial conditions and parameter values; and (iii) the robustness across changes in modeling assumptions.” According to Casini, robustness analysis is especially important because this operation allows it to be shown that “the mechanisms represented by the models are different ‘tokens’ of the same type” (ibid.: 666), which suggests that the model at hand captures realistic, not accidental or artificial, features of the actual causal story.
Through a different language, Ylikoski and Aydinonat (2014) defended a similar argument. By studying a very abstract ABM, they argued that systematic variations of the initial model’s assumptions (i.e. robustness analysis) lead to a “clusters of models”. These clusters can contribute to causal analysis in two ways. One the one hand, they help us to see what is essential in the original model, thus helping to formulate hypotheses on the “core” features of the actual causal stories; on the other hand, clusters of models help us to formulate alternative explanations, which allow us to better locate a given causal story to be empirically tested within the set of possible alternative causal stories.
These philosophical analyses led to different conclusions as to the value of ABMs for causal reasoning but they shared an implicit assumption: as long as an ABM cannot generate on its own all the required empirical data to corroborate the assumptions on which it was built, the ABM can at best guide, or complement, methods for causal inference relying on empirical data but it cannot ascertain on its own actual causal stories. However, as noted by Elsenbroich (2012) in her response to Grüne-Yanoff’s critique, “(…) insisting on complete knowledge of microphenomena for a causal explanation makes causal explanation impossible in the social sciences. This is also not a problem of ABM but of social science as a whole.” I will develop this important point further when I discuss data-driven methods for causal inference and show that these methods, too, can make causal claims only contingently on assumptions that they cannot test empirically on their own (see Chapter 5).
Social scientists formulated similar views on the basis of similar implicit assumptions. Macy and Sato (2008), in response to the critique of lack of realism of an ABM that they proposed, defended their model by claiming that “[t]he computational model can generate hypotheses for empirical testing, but it cannot ‘bear the burden of proof’”, thus implicitly proposing a division of labor according to which the ABM is a tool for theoretical exploration while experimental and statistical methods for observational data are better suited, and necessary, to support causal inference. Quantitative scholars relying on this type of method overtly expressed skepticism about “(…) the utility of many simulation-based methods of theory construction” (Morgan and Winship 2015: 341). As clearly visible in Morgan’s (2013) overview of the most recent developments in the field of causal inference, the ABM is simply not considered as a potential player in this game.
However, other scholars have noted that an ABM can communicate with empirical data in various ways, which in principle makes it capable of pinpointing real-world mechanisms underlying the dependence relationship between variables (Hedström 2005: ch. 6; Manzo 2007; Bruch and Atwell 2015). This view has again been recently attacked by Törnberg (2019) who argued that, because of the contingent nature of the social realm, it is impossible to know how a simulated mechanism will behave in the real world. For this reason, he claimed, “whether these mechan...
