This thesis examines the implications of using different correlation input matrices and estimation techniques in confirmatory factor analyses (CFAs) when analyzing ordinal, nonnormal data derived from responses of recently arrived Australian immigrants to the 12-item General Health Questionnaire (GHQ-12). The GHQ-12 is one of the most widely used instruments for determining wellbeing in populations. The response format of the GHQ-12 comprises four ordinal categories and underlying distributions of data obtained invariably do not approximate univariate or multivariate normality. Owing to these data properties, consideration should be given to the application of appropriate statistical approaches for analyzing this type of data sets. This study also investigates the extent to which the GHQ-12 is invariant across gender and cultural groups.
A three-dimensional measurement model for the GHQ-12 was initially examined for four groups of Australian immigrants who originated from Hong Kong (n = 201), Mainland China (n =213), former Yugoslavia (n = 259), and the United Kingdom (n = 428). A series of CFAs using either a Pearson�s product-moment or a polychoric correlation input matrix and employing either maximum likelihood (ML), weighted least squares (WLS) or diagonally weighted least squares (DWLS) estimation methods was conducted on the data. A comparison of the parameter estimates and goodness-of-fit statistics obtained for the different analyses provided support for using polychoric correlation input matrices and DWLS estimation in CFAs when analyzing ordinal, nonnormal data with smaller sample sizes.
Invariance tests across gender and cultural groups were conducted on a second-order measurement model for the GHQ-12, culminating in significant differences between the two Asian and two European cohorts. The GHQ-12 was invariant for immigrants from Hong Kong and Mainland China, as well as for males and females from the United Kingdom. Partial invariance of the GHQ-12 was found for immigrants from Asia, the United Kingdom, and former Yugoslavia and for Asian males and females.
Findings from the present study suggest that estimating models based on nonnormal ordinal responses using polychoric correlations with DWLS is more likely to result in a solution with higher parameter estimates and better indices of fit than other approaches. Further research should be conducted on real and simulated data to investigate the efficacy of WLS and DWLS estimation in CFAs when using polychoric correlations as the input data for varying categorical response formats, with a range of model and sample sizes.
| Identifer | oai:union.ndltd.org:ADTP/216518 |
| Date | January 2005 |
| Creators | Wang, Wei Chun, wwang@swin.edu.au |
| Publisher | Swinburne University of Technology. |
| Source Sets | Australiasian Digital Theses Program |
| Language | English |
| Detected Language | English |
| Rights | http://www.swin.edu.au/), Copyright Wei Chun Wang |
Page generated in 0.0022 seconds