|
|
 |
| GMM for non-normal ordinal outcome |
 
|
| Message/Author |
|
|
| jhodel posted on Thursday, October 03, 2019 - 6:39 am
|
|
|
Dear all,
I would like to perform GMM on longitudinal data about persons undergoing clinical rehabilitation in order to explore the number of classes of recovery during rehabilitation. The study design includes max. 4 repeated unstructured measurement time points of an ordinal outcome (sum score between 0 and 100) which is non-normally distributed.
Therefore I was thinking of fitting a skew-t GMM for categorical outcomes with individually-varying time scores. Is this possible in Mplus and do you have any examples on this?
I also read into the R package lcmm, which seem to be able to perform the desired GMM with unstructured data, non-normal and ordinal outcome by using Link Functions and Latent Process Mixed Models (https://arxiv.org/pdf/1503.00890.pdf). Unfortunately, there is very few literature around which describes the differences between the approaches used in the lcmm package and Mplus for skewed and categorical outcomes. Therefore, I was wondering if you know more about the differences and advantages/disadvantages of the two approaches used in these two softwares?
Thanks a lot in advance! Best regards,
Jsabel |
|
|
|
|
Skew-t GMM works well with skewed variables where the cause for the skewness is NOT strong floor or ceiling effects. See e.g.
Muthén, B. & Asparouhov T. (2015). Growth mixture modeling with non-normal distributions. Statistics in Medicine, 34:6, 1041–1058. DOI: 10.1002/sim6388
I will send you an output from this.
I am not sure you need to treat your outcome as ordinal - it sounds sufficiently interval scaled to be considered a continuous variable.
I am not familiar with lcmm but it probably does something similar. You can try it and compare - check that you have the same number of parameters and best loglikelihood value. |
|
|
| Back to top |
|
|
