Message/Author |
|
Anonymous posted on Friday, January 13, 2006 - 1:43 pm
|
|
|
I'm running a growth curve model and in my model I have a modest fit that can be improved by correlating my measures of time2 and time3 together. I was wondering what does this actually mean in my model? Does this mean that time2 and time3 are essentially the same measure? Or are they correlated because they are the same question? |
|
|
Residual covarainces can represent time-varying covariates that have not been included in the model. Residual covariances are not needed just because the same outcome is measured at multiple time points. This is taken care of by the growth model itself. |
|
Vanessa posted on Tuesday, June 15, 2010 - 10:40 pm
|
|
|
Hi, I have a second-order latent growth curve model and the model fit is not great (CRT is indexed by the same 3 indicators, at each of 4 timepoints). Modification indices suggest considerable improvement if residuals of the same indicators at adjacent time points (plus some unadjacent timepoints) are allowed to covary. Although it is a second-order LGCM, is this the same issue as above? ie. that the slope/intercept should be accounting for the variance shared due to same measure at different timepoints so it must indicate that a time-varying covariate has not been included? Thanks |
|
|
Yes, this could be the same issue. |
|
Vanessa posted on Wednesday, June 16, 2010 - 8:09 pm
|
|
|
Thanks. So if it is due to a time-varying covariate that has not been measured in the study, does this cast doubt on the values of other parameters that are estimated? Is it considered permissable to allow covarying residuals? |
|
|
You can add residual covariances to the model as an alternative. |
|
Jen posted on Friday, October 08, 2010 - 12:43 pm
|
|
|
To follow up on this question, I wondered if there was a recommended citation for the practice of including residual covariances? Also, when building parallel process models, there are often strong correlations between measurements taken at the same time point (x1 & y1, for example). Would you consider it acceptable to include those correlations in the model as well, and, if so, are you aware of a citation that would endorse the practice? Thanks very much! |
|
|
I don't know of such references. The Boolen book may discuss this. Residual covariances can be significant when covariates that influence both variables are left out. It is difficult to give advice about what to include in a model without more information than can be provided on Mplus Discussion. |
|
Back to top |