Parameter Recovery and Model Fit Using Multidimensional Composites: A Comparison of Four Empirical Parceling Algorithms

Abstract

Manifest variables in covariance structure analysis are often combined to form parcels for use as indicators in a measurement model. The purpose of the present study was to evaluate four empirical algorithms for creating such parcels, focusing on the effects of dimensionality on accuracy of parameter estimation and model fit. Results suggest that accuracy of parameter estimation is primarily a function of the nature and number of dimensions in a composite, and that greater dimensionality also renders parceling methods more similar with respect to accuracy of estimation. Conversely, model fit is predominantly influenced by the parceling algorithm and number of parcels formed. An integrative analysis of the degree to which model fit signals accuracy of estimation highlights the respective advantages and disadvantages of the parceling algorithms. Results suggest that a Radial parceling algorithm may offer advantages over Correlational, Factorial, or Random assignment of items to parcels. The behavior of the Radial algorithm, and iterative algorithms in general, warrant further investigation.