Multiple imputations and covariates f... PreviousNext
Mplus Discussion > Growth Modeling of Longitudinal Data >
Message/Author
 Eric Wan posted on Friday, September 04, 2015 - 3:48 am
I would like about the multiple imputations and covariates for GMM.
1) Is it correct that Tech 11 (Lo-Mendell-Rubin test) isnˇ¦t available after multiple imputations? If not available, how can I conduct this test after multiple imputations?
2) Should I select the number of class by GMM without covariates? And then the covariates will be added in the final model after the final model of unconditional GMM.
3) For the conditional GMM, how can change the reference group of the class on covariates by multinomial logistic regression (c on X)? The default setting of Mplus is the last class.
Thank you very much.
 Jon Heron posted on Friday, September 04, 2015 - 1:47 pm
[1] Perhaps make your decision regarding the number of classes prior to imputation, using FIML to deal with any missing class indicator information.

[2] Whilst I have seen a paper where the number of classes was reduced following the introduction of covariates - since covariates explained some of the continuous variation - I don't think this is the norm. Assuming that you are using imputation to deal with covariate missingness, my response to [1] implies a yes to your question [2].

[3] You can play with starting values to permute the class ordering. Personally I find this doesn't always work so instead I use different OPTSEED values until I get the class ordering I want. Having said that, for some models Mplus does give you results for all the different reference groups anyway.


best, Jon
 Tihomir Asparouhov posted on Friday, September 04, 2015 - 5:07 pm
I will just add to Jon's reply. For [1] consider also this method. For each imputed data set use tech11/tech14/bic to determine the number of classes. Then select the number of classes that is selected most frequently across the different imputed data sets. If the amount of missing data is not large, the selected number of classes will not differ across the different imputed data sets.
 Eric Wan posted on Sunday, September 06, 2015 - 3:57 am
Thanks for the replies.
1) I still find uncertain for the selection of number of classes by unconditional or conditional GMM. I read a previous literature (Bengt Muthen (2004). Latent variable analysis: Growth mixture modeling and related techniques for longitudinal data. https://www.statmodel.com/download/KaplanChapter19.pdf) and it suggested that the covariates should be included to find the proper number of classes. However, I also read other literatures such as (Dunn, Jeff, et al. "Health-related quality of life and life satisfaction in colorectal cancer survivors: trajectories of adjustment." Health and quality of life outcomes 11.46 (2013): 1-8. http://www.hqlo.com/content/11/1/46 ). Firstly, they found the number of classes (K-class) by using the unconditional GMM. Then, covariates were entered into the K-class (unconditional) GMM via multinomial logistic regression to identify the predictors among covariates. Is it correct to do it? If yes, have any literatures to support this procedure?
 Eric Wan posted on Sunday, September 06, 2015 - 3:57 am
2) For FIML, how can I set it for covariates? I know that FIML is default setting for the independent variables in GMM. Is it correct to replace missing values of covariates by FIML as follows?

Analysis:
Type = mixture;
algorithm = integration;
integration = montecarlo;

Model:
%overall%
i s q
i s q ON a list of covariate
c ON a list of covariate
[a list of covariate ]

Thank you very much.
 Bengt O. Muthen posted on Sunday, September 06, 2015 - 7:07 pm
Some authors find advantages to deciding on classes without covariates. You can email Hanno Petras at AIR to get a copy of this book chapter:

Petras, H & Masyn, K. (2009). General growth mixture analysis with antecedents and consequences of change. To appear in Piquero, A. & Weisburd, D., Handbook of Quantitative Criminology.

2) Yes, but c on covariates where covariates are brought into the model likes this may give many dimensions of integration because of missingness on the covariates.
 Eric Wan posted on Monday, September 07, 2015 - 3:53 pm
Many thanks for Professor's reply.

1) I found the paper in researchgate. (https://www.researchgate.net/publication/226942423_General_Growth_Mixture_Analysis_with_Antecedents_and_Consequences_of_Change). However, Iˇ¦m not familiar with GMM so would like to clarity the advantages for deciding on classes without covariates mentioned in the paper. Is the main advantage that the models that assign all systemic variability in growth to class membership are less parsimonious but are more flexible and make fewer parametric assumptions to avoid empirical identification problem?

2) If I want to investigate the predictors for the classes after unconditional GMM, is it correct that the potential predictors only included in the class (c on predictors) but not in the intercept, linear and quadratic term (i s q on predictors)? Moreover, I would like to confirm that Mplus doesnˇ¦t allow us to change the reference group of the class on covariates by multinomial logistic regression (c on X)? The default setting of Mplus is the last class.

3) Yes, the computational time is very long. How I investigate the predictors effectively to decreases the running time? By MI?
 Bengt O. Muthen posted on Tuesday, September 08, 2015 - 11:55 pm
If you are not familiar with GMM as you say, you should first carefully read some of the many intro papers on the topic, rather than try to learn it by asking questions here where only short answers can be given.

1) The advantages are shown by Monte Carlo simulations in that paper. From a practical point of view it is much simpler to decide on number of classes without covariates. With covariates one has to decide on whether to let them influence the growth factors as well as c.

2) As I state in my chapter, the presence off significant effects from covariates to growth factors makes the unconditional method inappropriate.

The reference group can be changed by choosing starting values (using SVALUES) to move classes around.

3) Listwise present analysis or MI.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: