Measuring anxiety by ordered categorical items in data with subgroup structure: the case of the Hungarian version of the trait anxiety scale of the state-trait anxiety inventory for children (staic-h)

Abstract

The present study suggests a modelling methodology for examining measurement invariance of ordered categorical item indicators of latent constructs such as anxiety, coping, motives etc., in research settings with few subgroups and a large sample of individuals. The Hungarian version of the Anxiety Trait scale of the State-Trait Anxiety Inventory for Children (STAIC-H) was administered to 605 boys and 975 girls of age 10–15 in 12 schools. A MIMIC model was suggested for examining measurement invariance across subgroups of schools and ages, while a multi-group analysis was recommended for investigating invariance across gender. High degree of invariance across groups was obtained for the Anxiety Trait scale in terms of item factor loadings, item thresholds and item homogeneity with respect to group contrast variables. Based on the diagnostic information obtained by the present methodology, the few item indicators showing non-invariance were discussed with reference to methodological and conceptual considerations.

Keywords:

• Measuring trait anxiety
• Subgroup effects
• Categorical items
• MIMIC model
• Multi-group analysis

Acknowledgments

We gratefully acknowledge the comments on an earlier version of this manuscript provided by two anonymous reviewers and Reinhard Pekrun.

Notes

The root mean square error of approximation (RMSEA-ϵ_a) represents an advancement in the assessment of model fit from both statistical and conceptual point of view. Browne and Cudeck (1993) argue that theoretical models are at best approximations of reality. Therefore, the null hypothesis for any structural model (i.e., that the data will fit the model perfectly in the population) will rarely be true. Rather than testing the null hypothesis of exact fit between the structural model and the population covariance matrix, RMSEA establishes a hypothesis of close fit between the model and the population. RMSEA values of 0.05 or less indicates a close fit between the structural model and the population covariance matrix per degree of freedom. Values in the range of 0.05–0.08 reflect fair fitted models. Values between 0.08 and 0.10 represent mediocre fit, while values beyond 0.10 do not represent acceptable models (Browne and Cudeck, 1993). Later MacCallum et al. (1996) emphasized the use of interval estimates in addition to point estimates of RMSEA, since interval estimates are more informative of the model fit. The Mplus version 2.01, used in the present study, reports point estimates and confidence intervals of RMSEA for models with continuous indicators but only point estimates for models with categorical indicators (Muthén and Muthén, 1998).