Colliders: A potential biasing factor when evaluating the causal structure of an unobserved construct in a SEM?

Collider variables exist when two constructs cause variation in a third construct. Under these conditions, including the collider in a regression model estimating the association between the two constructs will introduce bias into key coefficients. The effects of regressing a latent variable upon a collider can have important implications for the evaluation of unobserved psychological constructs in applied research. Explicitly, regressing a latent variable upon a collider can potentially bias the results of the analysis and generate misinterpretations about the causal structure of an unobserved construct. Nevertheless, the biasing effects of regressing a latent variable upon a collider remains unexamined in the extant literature. The current study evaluated the bias generated from regressing a latent variable upon a collider through the employment of three simulation analyses designed from scenarios that could exist in the applied psychological literature. The results suggested that regressing a latent factor upon a collider can bias the global fit statistics and individual estimates. The magnitude and direction of the bias is conditional upon the number of latent constructs, the items related to the collider, and the construct in the SEM being regressed upon the collider. Implications and guidance for future scholarship employing CFA is provided.