Assessing the Properties and Functioning of Model-Based Sum Scores in Multidimensional Measures With Local Item Dependencies: A Comprehensive Proposal

A common problem in the assessment of noncognitive attributes is the presence of items with correlated residuals. Although most studies have focused on their effect at the structural level, they may also have an effect on the accuracy and effectiveness of the scores derived from extended factor analytic (FA) solutions which include correlated residuals. For this reason, several measures of reliability/factor saturation and information were developed in a previous study to assess this effect in sum scores derived from unidimensional measures based on both linear and nonlinear FA solutions. The current article extends these proposals to a second-order solution with a single general factor, and it also extends the added-value principle to the second-order scenario when local dependences are operating. Related to the added-value, a new coefficient is developed (an effect-size index and its confidence intervals). Overall, what is proposed allows first to assess the reliability and relative efficiency of the scores at both the subscale and total scale levels, and second, provides information on the appropriateness of using subscale scores to predict their own factor in comparison to the predictive capacity of the total score. All that is proposed is implemented in a freely available R program. Its usefulness is illustrated with an empirical example, which shows the distortions that correlated residuals may cause and how the various measures included in this proposal should be interpreted.