Unidimensional community detection: A Monte Carlo simulation, grid search, and comparison

Resumen

Unidimensionality is fundamental to psychometrics. Despite the recent focus on dimensionality assessment in network psychometrics, unidimensionality assessment remains a challenge. Community detection algorithms are the most common approach to estimate dimensionality in networks. Many community detection algorithms optimize an objective criterion called modularity. A limitation of modularity is that it penalizes unidimensional structures in networks, favoring two or more communities (dimensions). In this study, this penalization is discussed and a solution is offered. After, a Monte Carlo simulation using one and two factor models is performed. Key to the simulation was the condition of model error or the misfit of the population factor model to the generated data. Based on previous simulation studies, several community detection algorithms that have performed well with unidimensional structures (Leading Eigenvalue, Leiden, Louvain, and Walktrap) were compared. A grid search was performed on the tunable parameters of these algorithms to determine the optimal trade-off between unidimensional and bidimensional recovery. The best performing parameters for each algorithm were then compared against each other as well as maximum likelihood factor analysis and parallel analysis (PA) with mean and 95th percentile eigenvalues. Overall, the Leiden and Louvain algorithms and PA methods were the most accurate methods to recover unidimensional and bidmensional structures and were the most robust to model error. More nuanced method recommendations for specific unidimensional and bidimensional conditions are provided.

Áreas y palabras clave

Cómo citar (APA 7)

Christensen, A. P. (2022). Unidimensional community detection: A Monte Carlo simulation, grid search, and comparison. https://doi.org/10.1037/met0000692