Pseudo-Factor Analysis with Item Intercepts

We extend the pseudo-factor analysis method by adding a mean structure to obtain pseudo-intercepts. We do so by substituting theitem response mean vector based on human responses with a vector of scalars obtained by projecting item embeddings via dot productsonto semantic intensity axes capturing the difference between low and high scale anchors. Across six scales measuring the trait domainsof the DSM-5 alternative model for personality disorders in occupational settings, the mean correlation between empirical and pseudo-intercepts was r = .54, while the mean correlation between empirical and pseudo-intercepts obtained via deliberate projection ontorandomly incorrect semantic axes was r = −.35. The inclusion of embedding-based mean proxies in pseudo-factor analysis can potentiallyenhance scale development and allow mean and covariance structure modeling without empirical data. This may open the possibility ofexamining relative item locations, multiple-group pseudo-CFA/SEM modeling, including measurement and structural invariance analysesacross languages and tranformer architectures. Topics such as systematic methods to identify semantic anchors and risks including ‘anchor hacking’ are discussed.