Generating Empirically-Plausible Networks for Psychological Data

Simulation studies in network psychometrics have relied on data generating mechanisms that limit the levels of analysis that can be evaluated. Random Gaussian graphical models offer limited topological control, empirically-derived networks conflate the population with an estimated structure, and latent factor models produce dense covariance matrices that are theoretically incompatible with network theory. These limitations have largely confined simulations in network psychometrics to edge recovery, leaving centrality, community detection, and network similarity understudied. This talk introduces a novel data generating mechanism that addresses these gaps by pairing the Stochastic Block Model with an empirically-informed edge weight generation procedure. Rather than drawing weights from arbitrary distributions or a handful of empirical datasets, this approach characterizes edge weight distributions across hundreds of psychological datasets made available by Huth and colleagues (2025), finding that the Weibull distribution provides the best fit. A linear model anchors the Weibull's scale parameter to network features and sample characteristics, enabling simulation of edge weights that closely mirror empirical patterns while allowing systematic manipulation of community structure. Ongoing improvements of the approach will be discussed, including an improved model parameterization for more realistic edge assignments, incorporation of negative edge weights, and an additional structural mechanism, small-world networks, for constructs that do not follow block model assumptions. Together, these developments offer researchers a principled and flexible data generating mechanism, implemented in the open-source R package {L0ggm}, to evaluate network psychometric methods and measures across all levels of analysis.