Measurement equivalence testing 2.0

Item does not contain fulltextProbably the most ignored assumption in the social and behavioral sciences is the assumption that measures are observed without error. Ignoring measurement error when it is present will result in conclusions that are biased to some degree. A second commonly ignored assumption is that survey measures can be used to make valid comparisons between (groups of) persons. When the unit of the scale on which we express ourselves is widely used, for example, time in minutes, between-person comparisons are most likely valid. When time is, however, expressed in subjective labels, for example, very long, long, not so long, short, it is less clear whether between-person comparisons are valid; what a short time is for one person might be not so long for another. Saris (1988) showed for different topics that people vary in using these scales. For measures to be meaningfully comparable between (groups of) persons, they should be equivalent (Meredith, 1993). Measurement equivalence is tested with measurement invariance tests. The standard procedure tests between-group constraints on the factor model using multiple group structural equation modeling (SEM) analysis (Meredith, 1993). Recently, alternatives have been developed using Bayesian approaches (Cieciuch, Davidov, Schmidt, Algesheimer, & Schwartz, 2014; Asparouhov & Muthén, 2010). However, our approach is different and has two unique aspects. 1 First, we use a different measurement model - implementing the concept of correction for measurement error in invariance testing. Second, we use a different approach to evaluate invariance across groups. This new approach to model evaluation is an answer to two issues that make testing multigroup 246SEM models complex. The first issue is the power of the test, which is influenced by sample size, normality of the data, and the size of incidental model parameters (Saris, Satorra, & van der Veld, 2009). These conditions are often found in measurement invariance testing. The second issue is the large amount of output that SEM programs produce in case of multigroup analysis, which makes it hard to oversee the results