A Comparison of Latent and Deterministic Blockmodeling with Application to Binary Substance Use Disorder Data

Area of Science:

• Network analysis
• Data science
• Psychiatry

Background:

• Substance use disorder data collection often involves item endorsement, forming bipartite networks.
• These networks are represented as two-mode binary matrices (individuals x items).
• Two-mode blockmodeling partitions individuals and items within these networks.

Purpose of the Study:

• Compare the latent blockmodel (stochastic) with deterministic blockmodeling methods.
• Evaluate performance in recovering cluster memberships for individuals and items.
• Assess suitability for sparse bipartite networks in substance use disorder data.

Main Methods:

• Latent blockmodel: a stochastic approach with a statistical foundation.
• Deterministic blockmodeling methods: used for comparison.
• Simulation study and analysis of a real-world multiple-substance use disorder dataset.

Main Results:

• Latent blockmodel and deterministic methods showed comparable performance in cluster recovery when the number of clusters was prespecified.
• One deterministic method proved unsuitable for sparse bipartite networks.
• The latent blockmodel offers a principled method for selecting the number of clusters.

Conclusions:

• The latent blockmodel is a robust and advantageous method for analyzing substance use disorder data.
• It provides a statistically grounded approach for network partitioning.
• Demonstrated effectiveness compared to deterministic methods, particularly for sparse data.