Variants of Estimating an IRT-Based Actor-Partner Interdependence Model (APIM) with R

The Actor-Partner Interdependence Model (APIM) is a complex regression-based model capturing the simultaneous effects within a dyad. In the present study, we explore several variants of estimating how self-esteem (assessed with Rosenberg’s self-esteem scale) predicts relationship satisfaction (measured with the Relationship Assessment Scale) within each partner and between partners. The main focus of the study was to apply IRT models using the free and open software R. For that purpose, we contrast (1) a simple sum score-based path analysis and (2) two structural equation modeling approaches (one based on the covariance matrix for interval data one based on the polychoric correlation for ordinal data) to (3) three Item Response Theory (IRT) models, the Partial Credit Model (PCM), the Generalized Partial Credit Model (GPCM), and the Graded Response Model (GRM), each in a uni- and a multidimensional approach. For obtaining the person parameter estimates of the IRT models, we applied maximum likelihood estimation and three plausible values (PV) methods (10 PVs, median of 10 PVs, and 1 PV). Comparing the various modeling variants, we found the GPCM to fit best, slightly outperforming the GRM and clearly outperforming the PCM. The APIM coefficients varied considerably and indicated in some instances even contradicting effects. The IRT models excelled the other approaches for their detailed item analysis features revealed characteristic weaknesses of the scales, which would be difficult to detect otherwise.