Theory:

1. Like Phlogiston, fairies and witches, latent variables need to be abandoned
2. Psychological disorders are a 'network' of causally interacting symptoms (Borsboom, 2017)

Practice:

1. Estimate statistical models such as the Gaussian graphical model on cross-sectional data
2. Visualize the network
3. Compute so-called centrality measures for each node
4. Interpret high scoring nodes as important, suggest them as intervention targets

Examples:

• Stochl et al. (2018) studied well-being and find that 3 items are consistently the most central items
• McNally et al. (2015) studied earthquake victims suffering PTSD (use the word causal 59 times)

The dominant causal inference framework is based on DAGs

Textbook example:

• Strong correlation between number of storks (\(X\)) and number of human babies delivered (\(Y\); Matthews, 2000)
• If we condition on economic development (\(Z\)), then \(X\) and \(Y\) are independent

Three ways in which information gets lost when moving from DAGs to MRFs

1. Loss of directionality
2. Conditioning on colliders introduces new undirected edges
3. Multiple Markov equivalent DAGs can map to the same MRFs

• Correlation decreases with network density and graph size
• This is because distribution over centrality scores & causal influence differ in skew

• DAGs are a hierarchical ordered system of (linear) equations
• Thus, nodes higher up the hierarchy have more pronounced causal influence
• In contrast, MRFs drop directionality, thus obscuring this hierarchy
• Moreover, MRFs 'marry parents' which bumps up the score of leaf nodes
• This effect is stronger with increased network density, creating strong right-skew

• This propagates 'back', such that right-skew cannot occur
• Therefore, the correlation is strong even with high network density

We have illustrated the (not so?) obvious

1. Centrality measures do not straightforwardly map to causal influence
2. (Single-symptom) interventions based on centrality measures may thus be misguided

First, interpreting DAGs causally requires two important assumptions

1. The causal Markov condition (see Cartwright, 1999; Dawid, 2010a,b)
2. Faithfulness (see Ramsey, 2010)

Denny already argued in 2010 that DAGs are the wrong formalism for psychological phenomena

• But that would still be cross-sectional

• Try to combine experimental with observational data (Barenboim & Pearl, 2016)
• Use quasi-experimental techniques to establish causality (Marinescu et al., 2018)

• Continuous time modeling (Ryan et al., 2018)

• Mutualism in intelligence research (van der Maas et al., 2006)
• Panic disorder in clinical psychology (Robinaugh et al., still in progress)
• Attitudes in social psychology (Dalege, et al. 2016)