|
|
 |
| Testing configural invariance with Lo... |
 
|
| Message/Author |
|
|
|
|
Hi,
I am currently analysing a longitudinal auto-regressive panel model with missing non-normal data and so I am using MLR to estimate the model parameters. I am at the stage of assessing the fit of the underlying measurement model and testing whether scale invariance holds, before I go on to analysing the structural portion of the model. As it stands when I run a model, I do not get any of the standard fit indices such as RMSEA, CFI etc, and only get a log-likelihood for the model.
Given the lack of fit indices I am aware it's possible to use the log-likelihood to compare nested models (http://www.statmodel.com/chidiff.shtml). However, I wanted to make sure the process I am intending to follow will allow me to adequately assess if I have a well-fitting measurement model or not. Will the following process provide a suitable testing procedure for examining fit and testing invariance?
i) Estimate the saturated model in which my measurement model is nested
ii) Estimate the configural invariance model and compare its log-likelihood to the saturated model
iii) If it passes (no significant difference) estimate stronger (metric/weak) invariance model and compare to with the configural invariance model
iv) Continue with above to test further more restrictive models against previous model.
Best wishes,
Thomas |
|
|
|
|
| That approach sounds ok, although I don't know what the saturated measurement model means. I assume you don't have continuous outcome. If they are categorical, you can also use TECH10. |
|
|
|
|
Hi Bengt, thanks for your response. Yes I can see that wasn't entirely clear.
By the saturated model I mean the model in which all the measures are allowed to co-vary with each other. My understanding is that all models using the same co-variates are nested within this model and that it should have 'best' fit to the data since it is just reproducing the observed variance-covariance matrix exactly.
Some outcomes are not continuous, and I believe this is why the standard fit indices are not being produced. |
|
|
|
|
| So covarying all DVs at al time points? That might be an ok model to compare to if you can get this model to work. |
|
|
|
|
| Yes, that's what I mean. Okay great, thank you. |
|
|
| Back to top |
|
|
