1.1 Introduction
The application of network thinking to psychometric questions has led to an eruption of network approaches in several subfields of psychology, most notably clinical psychology and psychiatry (Robinaugh et al., 2020) where the idea that mental disorders are composed of interactions between components (symptoms or other problems) in a multifactorial system is plausible (Kendler et al., 2011). In many such cases, the statistical application of network models to empirical data is motivated by theoretical concerns: substantive considerations that render the conceptualization of a construct as a network plausible (e.g., because causal connections between relevant variables stand to reason; Cramer et al., 2010; Dalege et al., 2016; Isvoranu et al., 2016; Lange et al., 2020). This entanglement of statistical modeling and substantive concerns is typical of the network literature, and arguably in part responsible for its success: the combination of systems thinking with methodological tools to analyze empirical data are important drives behind the popularity of network approaches. However, the close relation between substance and statistics can also lead to problems, because the distinctions between statistical, conceptual, substantive issues are not always clear. Thus, one may unwittingly mistake conceptual questions for statistical ones, or statistical questions for substantive ones, which may hamper progress. The present chapter aims to make the distinction between substantive network theories and statistical network models explicit, and to discuss ways in which they can be connected.
Consider the network approach to psychopathology. Network theory posits that symptoms of psychopathology are causally related (i.e., activating one symptom increases the likelihood of connected symptoms to arise; Borsboom et al., 2017). However, this is not an assumption that underlies the estimation of the popular pairwise Markov random field (PMRF) as explained in Chapter 6. The assumptions of the estimation procedure are purely statistical and hold only that the joint probability distribution can be described by a set of main effects (node thresholds) and pairwise interactions (edges). Because the network theory and the network model are not equivalent, it is important not to confuse assumptions of one for assumptions of the other. Such confusion arises, for instance, when researchers think that application of the network model requires that the network theory is true. This is incorrect: the PMRF, as network model can be successfully estimated in situations where the network theory is false; for instance, when all dependencies in the data arise from a latent common cause rather than from causal interactions between components (Fried, 2020; Marsman et al., 2017). In this case, mistaking theoretical assumptions for statistical ones may inadvertently hold back the researcher, who may incorrectly think that it is unjustified to apply a network model in cases where relations between network components are not causal in nature. The converse problem arises when researchers think that the successful application of network models indicates that network theory is true. This is incorrect because the network model only contains statistical relations, and the interpretation of such relations in terms of causality requires a stronger inference than data analysis by itself can provide (i.e., one has to provide causal assumptions as well; Pearl, 2010). Interpretation of network results in terms of network theory may, in such cases, overstep the evidence (Bringmann et al., 2019; Epskamp et al., 2017; Fried, 2020).
For these reasons, it is crucial that the student of network analysis learns to distinguish between different ways in which network thinking can be applied, and to get a clear view of the distinctions between theory and statistical modeling. This chapter aims to clarify these distinctions by discussing three ways in which network thinking may be methodologically useful. First, we discuss network approaches, which simply entails âviewingâ a construct as a network of interacting components as a way of developing oneâs thinking and creating new ways to understand and investigate the construct. Second, we describe psychometric network models (hereafter, referred to simply as network models): systems of statistical relations between variables defined on empirical data. Third, we describe network theory, where both the components of the system and their connections are substantively interpreted and specified, so that they are able to explain characteristic phenomena involving the construct. We then discuss how these conceptualizations of networks are related, identifying network models and network theory as specializations under the rubric of a network approach that, ideally, serve to inform one another. Finally, we analyze a number of strategies that have been applied to connect network theory and network models.
1.2 Network approaches
The network approach to a given empirical domain simply entails conceptualizing that domain as a network: that is, as a set of components and the relationships among those components. Central to this approach, which others have referred to as âsystems thinking,â is the notion that network behavior is closely tied to network structure (Meadows, 2008). The network approach often entails conceptualizing a given phenomena of interest as an emergent property, with the components of a network âworking togetherâ to generate emergent phenomena that feature surprising levels of organization (BarabĂĄsi, 2012; Wright & Meadows, 2012). For example, consider the complex and seemingly highly organized behavior exhibited by a flock of birds. A network approach conceptualizes such behavior as an emergent property arising from the interrelationships among the individual birds that constitute the flock. This same lens can be applied to numerous empirical domains. For example, we can apply it in the domain of psychiatry, considering whether the behavioral characteristic of the depression syndrome (i.e., chronically elevated depression symptoms) might arise from interrelationships among the components of the syndrome (i.e., the symptoms themselves). This process can be facilitated using analogical abduction (Haig, 2014): establishing a systematic correspondences between a source domain (e.g., a flock of birds) and a target domain (e.g., a set of depression symptoms) so that one can use explanatory models from the source (e.g., flocking models that explain why flight courses of birds are correlated) to better understand phenomena in the target (e.g., symptom network models that explain why depression symptoms are correlated). Indeed, because similar features are consistently observed across networks taken from a wide range of empirical domains, the network approach provides fertile ground for productive analogical abduction (BarabĂĄsi, 2012; Scheffer, 2020).
When one is initially considering the viability of network approaches, it is useful to consider the applicability of its central concepts. Can one identify at least some of the important components and links between them? Do these components behave as a coherent whole? Do the links between them offer prima facie plausible explanatory resources, in the sense that a network structure would âmake senseâ of a given behavior or pattern of findings (e.g., correlations in oneâs data)? Do we see synchronized behavior that may emerge out of a network structure? Are there signs that the system shows behavior commonly exhibited in complex systems, such as non-linear behavior in which there are sudden changes in the state of the system (Scheffer, 2020)? Systematically investigating such questions can help in assessing whether a network approach is plausible and worth pursuing (Fried & Robinaugh, 2020).
If these or other initial considerations suggest the network approach may be a suitable conceptual framework, the researcher can put on a pair of network glasses and start exploring whether network science may offer further possibilities. When embarking on such a discovery process, one typically does not make a particular choice on exactly which components are in play or how they work together; rather the researcher chooses to view an empirical domain through the âlensâ of networks. Thus, network approaches usually do not single out precisely which factors act as components in the system at the outset, nor do they typically specify precisely how they interact. Instead, these are thought of as open questions worthy of discussion and research. For example, in the network approach to psychopathology, symptoms enumerated in diagnostic criteria for a given disorder were put forward as possible components of the network structure. However, it has subsequently been argued that this viewpoint may be too narrow and that other components should be included as well (Fried & Cramer, 2017; Jones et al., 2017). Similarly, in the mutualism model of intel...