Anonymous posted on Wednesday, October 13, 2004 - 2:01 am
Since "higher" Residuals (observed-expected in an EFA; and Correlations in a CFA) between Variables can be seen as responsible for the height of RSMSR respectively SRMR, can you tell me something about a cutoff or boundary for the values in "Residual Observed-expected" or in the "Residual Correlations"? So that i can perhaps say: Ok, this or that variable is quite often involved, so I should take it out of the analysis to get lower RMSR or SRMR.
bmuthen posted on Wednesday, October 13, 2004 - 4:51 pm
It is hard to say, but if you consider < 0.05 as a reasonable decent RSMSR/SRMR cutoff, and noting that this is the average residual in a correlation metric, then it seems a good idea to keep an eye on residuals in the correlation metric that exceed 0.05.
Anonymous posted on Wednesday, December 29, 2004 - 12:43 am
I have got a question regarding the residual correlations. I have run several models and the modification indices output tells me to add a residual correlation between two items of one of my latent variables. The fit indices grow significantly when adding this residual correlation. But how can I theoretically explain this? Is a residual correlation a sign of a misspecified model, of ommitted variables or just of very similar scales? Is it quite normal to add residual correlations or is it seen as a weekness of a model?
Residual covariances are legitimate parameters but should only be added if they make some substantive sense. In factor analysis, they might represent a minor methods factor, for example, due to similar wording of items or another minor factor that was not well-represented and therefore was not found in an EFA.
Anonymous posted on Tuesday, April 26, 2005 - 6:48 am
I wonder if it seems appropriate to make a residual covariance between the items " I am in good health" and "times a week I am doing sports"? Althought, they are not similar in wording, they somehow include each other.