Running Group-Comparison of CFA using AMOS
AMOS Files: ...¥OUPSEM¥CFA¥Group1.amw
...¥OUPSEM¥CFA¥Group2a.amw
...¥OUPSEM¥CFA¥Group2b.amw
...¥OUPSEM¥CFA¥Group2c.amw
...¥OUPSEM¥CFA¥Group2d.amw
...¥OUPSEM¥CFA¥Group2e.amw
...¥OUPSEM¥CFA¥Group2f.amw
...¥OUPSEM¥CFA¥Group3.amw
Data Files: “...¥OUPSEM¥CFA¥Female.xls” for the Female group,
“...¥OUPSEM¥CFA¥Male.xls” for the Male Group
In this example, you will learn how to conduct group comparison in CFA using AMOS. The data set is comprised of 2,204 children, after listwise deletion of missing data, on 26 items originally designed to measure children’s recent experiences with violence exposure at home, school, and neighborhood. Dr. Mark Singer has developed six subscales using these data, which confirms his original design of the items and the 1995 study using a similar set of items (i.e., 4 items fewer) on a different sample (Singer et al, 1995, JAMA). For details of the scale-development procedure, see CFA1.doc.
The objective of the current study is to investigate: whether the items comprising the six subscales operate equally between boys and girls? Is the measurement instrument fully or partially invariant between boys and girls? To which extent the items and subscales may contain construct bias?
Step 1: Development of the baseline model
Below is the path diagram of the final CFA model we developed using the whole sample. This is a near-congeneric model. Although the meaning for items 6, 3, 9 & 15 is not clear (i.e., those items have cross-loadings), we decide to retain the original model with 6 subscales. The decision was made due to two reasons: (1) this model was previously factorial-analyzed (Singer et al, 1995, JAMA); and (2) this model is also suggested as one of the two possible models using the current data.
Step 2: Run the baseline model to test
The following path diagram shows the baseline model. We tested it separately for boys and girls. Now as a first model for group comparison, we just stack the two groups’ data together to estimate the coefficients for both boys and girls simultaneously (i.e., two sets of estimations provided by one overall model).
Instructions on running group-comparison with AMOS:
· Linking data and define data for different groups: File -> Data Files -> Group Variable and Group Value
· Defining groups: Analyze -> Manage Groups (Type group name)
· Managing models: see Model 4 below.
Step 3: Run 6 different models to test
Step 4: Results show that factor loadings on both F2 and F5 can be constrained to be equal. The following model just constrains the two factors’ loadings equal between groups.
Because we cannot accept for all factors, there is no reason to further test , , and . We stop here.
Results of Chi-square difference testing:
Conclusion: Data shows that the instrument of “Recent Violence Exposure Scales” is partially measurement invariant; the instrument operates differently between boys and girl on items comprising F1, F3, F4, and F6.
AMOS offers a test of these additional hypotheses automatically (labeled as: measurement weights, structural coefficients, and measurement residuals, respectively). Below (Model 4) is from such procedure trigged by the “Multiple-Group Analysis” icon.
A Group Comparison CFA for Bowen & Guo SEM page 1