The Effects of Using Partial or Uncorrected Correlation Matrices When Comparing Network and Latent Variable Models.

Dennis Mcfarland

doi:10.3390/jintelligence8010007

Abstract

Network models of the WAIS-IV based on regularized partial correlation matrices have been reported to outperform latent variable models based on uncorrected correlation matrices. The present study sought to compare network and latent variable models using both partial and uncorrected correlation matrices with both types of models. The results show that a network model provided better fit to matrices of partial correlations but latent variable models provided better fit to matrices of full correlations. This result is due to the fact that the use of partial correlations removes most of the covariance common to WAIS-IV tests. Modeling should be based on uncorrected correlations since these represent the majority of shared variance between WAIS-IV test scores.

Highlights

The structure of cognitive test batteries, such as the WAIS-IV or the WJ-IV are typically modeled with factor analytic procedures (Dombrowski et al 2017) or covariance structural modeling (Benson et al 2010)
The EBICglasso routine evaluates a series of L1 penalties on the partial correlation matrix and selects the one that results in the lowest Bayesian Information Criterion (BIC)
This study shows that a network model provides very good fit to matrices of partial correlations between the WAIS-IV subtests but very poor fit to matrixes of uncorrected correlations

Summary

Introduction

The structure of cognitive test batteries, such as the WAIS-IV or the WJ-IV are typically modeled with factor analytic procedures (Dombrowski et al 2017) or covariance structural modeling (Benson et al 2010). Psychometric network models represent correlation matrices as a graph in which each variable is a node and each correlation an edge Both Kan et al (2019) and Schmank et al (2019) used reduced partial correlation matrices of WAIS-IV test scores. These partial correlation matrices were further reduced in the sense that “spurious or false positives” edges were removed by use of model selection procedures in the qgraph package from R (i.e., they created sparse correlation matrices with many values set to zero) Both Kan et al (2019) and Schmank et al (2019) compared model fit indices for network models with those for traditional latent variable models and found the former to provide a better fit. It is not possible to rule out the possibility that differences in model fit are due to the fact that disparate correlation matrices are compared

Participants

Procedures

Results

Discussion

Conclusions

Methods