omitted variable bias
Definition
Omitted variable bias is the distortion introduced into a statistical model when one or more variables relevant to the phenomenon under study are excluded from estimation. In network psychometric models, this omission can produce spurious edges between nodes that would otherwise be conditionally independent, or suppress genuine edges that reflect true symptom relations. These distortions compromise the identification of central nodes within a network, meaning that conclusions about which symptoms warrant targeted intervention may be systematically incorrect. EBICglasso showed greater robustness, in the statistical sense, to omitted variable bias than LoGo-TMFG across simulated conditions, and this performance difference diminished as network size increased, while sample size had little effect on the magnitude of bias.
Sources: Henry & Ye (2024)
Related Terms
- network psychometrics (1 shared article)
- centrality (1 shared article)
- measurement error (1 shared article)
Applications
Omitted Variable Bias and Measurement Error
Omitted variable bias and measurement error are treated as two distinct but related sources of model misspecification in Gaussian Graphical Models, both capable of distorting estimated network structure. Measurement error tends to inflate edge weights and overestimate network density, whereas omitted variables introduce spurious or suppressed edges, yet the two problems converge in their consequence: incorrect identification of the most central symptoms and potentially misguided intervention strategies. EBICglasso demonstrated greater robustness to both sources of bias relative to LoGo-TMFG, with the performance gap between the two methods narrowing at larger network sizes.
Sources: Henry & Ye (2024)
Omitted Variable Bias and Network Psychometric Models
Network psychometric models, particularly cross-sectional Gaussian Graphical Models, do not include built-in protections against omitted variable bias, leaving estimated partial correlation networks vulnerable to structural distortion. When relevant variables are excluded, the conditional independence properties that underpin GGM interpretation are violated, producing a network that misrepresents the true relations among symptoms. This has direct consequences for applied psychopathology research, where inaccurate network structure can lead researchers to misidentify symptom centrality and design ineffective interventions.
Sources: Henry & Ye (2024)
