UC Davis

Gaussian graphical models (GGM, aka partial correlation networks) have become increasingly popular in psychological science. The underlying rationale is that a group of variables, say, items from a psychometric scale, are thought to measure a dynamic system that mutually influence and interact with one another. This is distinct from customary approaches, such as structural equation models, where observed items are indicators that are thought to share a common cause represented by an unobserved or latent variable. This dissertation includes four chapters that approach (partial correlation) network psychometrics from a frequentist perspective. The first demonstrates that the advantages of regularization have been (substantially) overstated, and further that good old fashioned, non-regularized, regression works just fine for estimating psychological networks. The second builds upon this general notion and demonstrates that, in many cases, null hypothesis significance routinely outperforms "state-of-art" methodology. The first two chapters focused on detected edges, when another supposed advantage of regularization is that it reduces overfitting, especially compared to non-regularized methods. Accordingly, in chapter 3, I set out to demonstrate that, yet again, this is not really the case in situations common to the network literature (i.e., when there are far more observations $n$ than nodes $p$). That is, although The final chapter then employs frequentist reasoning to shed light upon the topic of replicability in networks.