Yang An posted on Wednesday, May 17, 2006 - 7:29 am
I am trying to use CFA to get factor scores, fixing variance of latent variables to 1. But usually, this is done in cross-sectional study. In our study, we have repeated measures for up to 10 waves. I am not sure how to handle this. I tried two appoaches: 1. use all repeated data we have,fit one CFA model. 2. use only wave 1 data to fit the model, and for subsequent waves use the same loading and intercepts.
Both methods above are kind of flawed. But I can not think of better way to do this.
Boliang Guo posted on Wednesday, May 17, 2006 - 7:35 am
this should be a measurement invariance study. if you did not test the structure, I think you should test it first using each time's data, if yes, I think you could conduct the ME/I test to see the to which extent your model invariant across measurement waves. if you have strong theory about the ME/I across measurement wave, then, I think you jsut get the factor score by each time's data.
I agree that you need to test for measurement invariance across time. If you have measurement invariance across time, I would do a multiple indicator growth model. I would not use factor scores.
Yang An posted on Thursday, May 18, 2006 - 7:19 am
So when we test for measurement invariance across time, do we only test if the factor loadings are the same? How do we deal with the intercepts of the obverved variables and the correlation between latent variables. Do we keep them the same? Also, when we do the test, how do we control for TYPE 1 error? Is there a simutaeous test available? Thank you very much.
pls refer paper Vandenberg, R. J. & Lance, C. E. (2000). A review and synthesis of the measurement invariance literature: Suggestions, practices, and recommendations for organizational research. Organizational Research Methods, 3, 4-69.
The lack of independence due to several variables being measured on the same person is taken care of by multivariate modeling of the data.
Amery Wu posted on Thursday, June 05, 2008 - 3:35 pm
I am following up on Amery's question above.
When I think of measurement invariance across time I usually think of testing this with a single group model. For example, imagine I am investigating the measurement of a 10 item scale that has been administered twice to the same respondents. This single group model would then have two latent variables with item manifest varibles each; the latent variable would be correlated and particular errors would be correlated reflecting the same scale being administered twice (e.g, item 1 on time 1 correlated to item 1 on time 2).
Now, if one, however, approaches this from a multi-group approach one is not able to model the correlated errors, nor the correlated latent variables. Would this not bias the Chi-squared tests of measurement invariance? I am still thinking it through but it seems to me that treating a repeated measures measurement invariance problem as multi-groups would lead to biased results.
Any thoughts on this would be much appreciated.
Bruno D. Zumbo, Ph.D. University of British Columbia
Measurement invariance across time cannot be done using multiple group analysis because the groups would not be independent. Our short course handouts are available on the website. See Topic 3, multiple indicator growth for an example of testing measurement invariance across time.
Can you point me to any additional reading on how to handle measurement invariance over time? Specifically, I need to understand this issue of the lack of independence of the repeated measures within the CFA framework.