Node centrality measures are a poor substitute for causal inference

Abstract

Network models have become a valuable tool in making sense of a diverse range of social, biological, and information systems. These models marry graph and probability theory to visualize, understand, and interpret variables and their relations as nodes and edges in a graph. Many applications of network models rely on undirected graphs in which the absence of an edge between two nodes encodes conditional independence between the corresponding variables. To gauge the importance of nodes in such a network, various node centrality measures have become widely used, especially in psychology and neuroscience. It is intuitive to interpret nodes with high centrality measures as being important in a causal sense. Using the causal framework based on directed acyclic graphs (DAGs), we show that the relation between causal influence and node centrality measures is not straightforward. In particular, the correlation between causal influence and several node centrality measures is weak, except for eigenvector centrality. Our results provide a cautionary tale: if the underlying real-world system can be modeled as a DAG, but researchers interpret nodes with high centrality as causally important, then this may result in sub-optimal interventions.

A Conceptual Introduction to Bayesian Model Averaging

“…These measures represent the relative importance of the variables and the information flow on the network (Borgatti, 2005). Centrality can also be used to identify specific symptoms that may be the target of clinical treatment (Fried, Epskamp, Nesse, Tuerlinckx, & Borsboom, 2016; but see Dablander & Hinne, 2019). Of course, computing these indices using only the most likely model results in a ranking of the nodes that ignores any uncertainty about the network structure.…”

Out of bounds? The boundary specification problem for centrality in psychological networks

“…For example, in their review of psychopathology networks, Borsboom and Cramer (2013) recommend the use of centrality measures to identify symptoms that may cause other symptoms. However, it is unclear how a network derived from crosssectional data, any more than a path model estimated from cross-sectional data, could provide information about causal influence (Dablander & Hinne, 2019). In other cases, the focus is not on drawing causal influence, but more descriptively on identifying the most central nodes.…”

Bayesian hypothesis testing for Gaussian graphical models: Conditional independence and order constraints

“…First, our methodology allows researchers to formally test the null hypothesis of conditional independence for each relation. Partial correlation networks are thought to represent causal skeletons (but see Ryan et al, 2019;Dablander and Hinne, 2019) and an important aspect of causal inference is conditional independence (Pearl, 2009). This requires assessing evidence for the null hypothesis.…”