Message/Author |
|
|
Dear Linda/Bengt I am wondering about the appropriate way to specify the path diagram for a latent growth curve model, in which the observed repeated measure is binary. I am using the MLR estimator so there is no error term. Should I then draw the path diagram as for example 6.1 but without the error variances? thank you, Patrick |
|
|
Yes. |
|
|
Thanks Linda, a quick follow-up; when using wlsmv with theta parameterization, would one show the error variances? Patrick |
|
|
Only in a multiple group or multiple timepoint model for some groups or timepoints. In multiple group analysis, the residual variances of categorical outcomes using the Theta parameterizaiton are fixed to one in one group and are free in the others. Likewise with multiple timepoints, the residual variances must be fixed to one at one timepoint and can be free at the others. In single group analysis with categorical outcomes and the Theta parameterization, the residual variances are fixed to one. |
|
|
Thanks again Linda could you point me towards references for the mlr estimator? I get very similar but slightly different estimates with wlsmv and mlr and would like to take a closer look at why this might be happening. I have come across quite a lot on weighted least squares but nothing on mlr. best wishes, Patrick |
|
|
MLR is discussed in TECHNICAL Appendix 8 which you can find on the website. If you are using the default logit link with MLR, I am surprised that the results are close because MLR is logistic regression and WLSMV is probit regression. If you are using LINK=PROBIT with MLR, what you say makes sense. |
|
|
Actually, I meant that the results are close in terms of pattern of magnitude and significance rather than the actual numbers. Is there any general sense in which one should prefer wlsmv or mlr for a growth curve model with repeated binary outcome? Patrick |
|
|
If you want to include residual covariances, it is easier to do with weighted least squares. With maximum likelihood, each residual covariance is one dimension of integration. When you have many factors and few items, weighted least squares is better because each factor is one dimension of integration with maximum likelihood. If you have few factors and many items, maximum likelihood may be a better choice. |
|
Back to top |