Statistical Concepts and Networks in Causality

Abstract

The present chapter focuses on more statistical approaches to causality, including network approaches. It examines test validity models (reflective, formative, mutualistic) and causal models in testing (regularity, counterfactual, process). Test validity involves two major approaches—behavior domain theory and causal theory of measurement.

Statistical approaches to causality are important, but classic experimental approaches to causality need to be supplemented by other means when the classic approaches cannot be applied, for example, due to ethical considerations in setting up certain experimental manipulations. The classical causal model involving experimentation gives validity (and generalizability) cardinal importance, and these are difficult to target in nonexperimental designs but, through their statistical innovations, they are narrowing their limits in these regards. The supplementary statistical approaches in the study of behavioral causality include the potential outcomes model and the directed acyclic graphs (DAGs) model. The former includes a basis in hypothetical outcomes that cannot be ascertained directly, involving the SUTVA (stable unit treatment value assumption), and the latter includes the equivalent of experimental manipulations in its graph surgery/“do” operators (interventions), causal descendants, and counterfactuals. A variation of this latter approach is the ICA (integrated counterfactual approach). Some of the new approaches to statistical mediation analysis include: average causal mediation effect; left-out variables error method; latent growth curve modeling; state-space modeling; and the ignorability-based approach.

Some of the philosophical precursors to statistical and related causality not only concern interventionist/counterfactual accounts but also concepts such as NESS and INUS. The former is defined as necessary element for the sufficiency of a sufficient set and the former as insufficient but necessary components of unnecessary but sufficient causes. These are complex concepts that inform but to not direct statistical approaches to causality in psychology. Aristotle’s concept of four causes still has currency today (material, efficient, formal, final), with efficient causes considered as the equivalent of mechanisms.

Baye’s theorem is an emerging approach in the statistical approach to causality. It is subjective rather than classically frequentist. It deals with concepts such as priors, precision, likelihood, posteriors, and credibility instead of confidence intervals, and it stands in opposition to the classic approach to testing null vs. experimental hypotheses.

Other portions of the chapter deal with FACCDs (Functional Analytic Clinical Case Diagram), ecology, Granger Causality (GC), and networks. The latter is explored in much more detail in subsequent chapters on the brain, in particular. Among the notable aspects of networks discussed in the present chapter include measures of centrality and betweenness.

The chapter also covers epidemiology, with its emphasis on temporality (e.g., predisposing, precipitating, perpetuating factors), and causal webs or pies. The chapter includes new models of statistics and causality, such as the decision theoretic approach, minimal causal models, dynamic causal modeling, and convergent cross-mapping.

The chapter includes a section on posttraumatic stress disorder (PTSD) because it has been related to the concept of networks by McNally et al. (Clinical Psychological Science, 1–14, 2014). For this area of research, the authors contrasted network modeling with the approach of latent variable/constructs. Some of the network concepts applied to the data include: networks of association, concentration, and relative importance. The key measures used also related to centrality and betweenness.