Message/Author 

Valentin posted on Tuesday, December 03, 2002  1:01 pm



I would like to examine the factor structure and the reliability (testretest) of a 5item observation scale (continuous variables)with data collected from a randomized clinical trial of Methylphenidate with ADHD children. So observation data are garthered a) before and after placebo and b) before and after ritalin 1) Could someone please help me with the proper design (CFA or something else)and the Mplus syntax to use for this purpose? 2) My sample size is rather low (N=81) and I don't know if even the SEM framework is suitable? Thanks in advance 

bmuthen posted on Tuesday, December 03, 2002  7:48 pm



This sounds like an analysis of 2 groups (placebo/ritalin), and 2 time points (before/after), each time point having 5 variables. This is then a 2x5 = 10 variate outcome in 2 groups. I think of this as a longitudinal factor analysis model in 2 groups. A similar setup for one group and 3 time points, adding growth modeling, is show in Example 22.4 in the Users' Guide, pp. 218219. With only 81 observations you can't have many parameters. Assuming measurement invariance across time, your model has for a given group 4 loadings, 5 intercepts, 2 factor variances, 1 factor covariance, and 1 factor mean (at the second time point), a total of 13 parameters. The other group is different at least wrt to the factor mean at the second time point. With proper randomization, the factor variance at the first time point should be invariant across the groups. So, your sample size might be ok. You can do Monte Carlo simulations to find out. Hope this helps. 


Is the example (22.4) mentioned by Bengt above still in the users guide? I cant seem to find it. Can I perhaps find it somewhere else? Thanks in advance 


The current number for that is Example 6.14. You can also find multiple indicator growth in the Topic 4 course handout on the website. 


Dear Linda and Bengt, I am currently working on a testretest analysis and I am getting the message THE SAMPLE COVARIANCE IS SINGULAR. and NO CONVERGENCE. SERIOUS PROBLEMS IN ITERATIONS. CHECK YOUR DATA, STARTING VALUES AND MODEL. When I look at the sample statistics, this can be explained, because there is very little difference between the item characteristics at t0 and t1. I still would like to run the model. Is there an alternative method to run this model? 


Do a Type=Basic run for the variables and see if you have correlation = 1 for any pairs. Deleting a variable in such a pair seems like the only way to go. 


Thank you for your answer. There are several pairs with a correlation of 1. I have been thinking about this problem and the fact that the variables need to be deleted is a bit problematic, because I really want to know the testretest reliability of the scale. I though maybe I could run two separate CFA's and save the factor scores and subsequently test the correlation between these factor scores. Would this be an acceptable alternative? 


I would not recommend this. Check what SEMNET thinks. 

Back to top 