
Message/Author 

Elina Vaara posted on Monday, November 11, 2013  12:03 am



Hi! I have longitudinal NMAR data (n in the first round about 2500). There is severe (about 30%) dropout rate from 1st measurement timepoint, and I wish to separate three agegroups and do analysis for them separately. I have earlier done CFA with categorical variables, but it does listwise deletion. I have 30 correlated variables, which I plan to use in longitudinal BSEM with Mplus, and check for measurement invariance longitudinally and between genders. I am thinking is this a good way to do my analysis: even if Bayes estimation gives me more flexibility, I was wondering if the amount of missing data is still a concern? I have found papers on Bayes and papers on missing data, but still I am not convinced that I could rely on BSEM results with nonignorable missingness I have. Could you give me advise on this? 


Bayes and maximum likelihood both require MAR and have the same concerns. For NMAR, see the following paper which is available on the website: Muthén, B., Asparouhov, T., Hunter, A. & Leuchter, A. (2011). Growth modeling with nonignorable dropout: Alternative analyses of the STAR*D antidepressant trial. Psychological Methods, 16, 1733. 

Elina Vaara posted on Sunday, November 17, 2013  11:45 pm



Thank you! I have your article now, but here are a few more questions: For now I should see how the latent factors are associated in time. I wish to see change/association of different timepoints for about 5 factors (from CFA or such for ordinal variables). Growth curves model just one aspect at a time? So should I use (continous) latent factors as outcomes in five separate growth analyses? I would see possibility to model dropout (ex. pattern mixture approeach) and 5 longitudinal factors at the same time. 


Start with one growth curve at a time. It is possible to do all 5 at a time, but you should first do one at a time. 


Please, I would like to check if this BSEM syntax is correct. Thanks a lot! ANALYSIS: ESTIMATOR = BAYES; PROCESS = 2; FBITER = 100000; ALGORITHM = GIBBS(RW); MODEL: f1 by Evoc* Reco ERey; f2 by CRey* OriL; f3 by Sem* Nome Leit Esc; f4 by Aprend* Flue Tril; f1f4 ON Age Educat (reg1reg8); f1(f1var); f2(f2var); f3(f3var); f4(f4var); Age(idvar); Educat(escvar); [Age](id); [Educat](es); Age with Educat(idesc); Aprend with FlueTril(res1res11); Flue with EvocTril(res12res21); Evoc with RecoTril(res22res30); Nome with SemTril(res31res37); Sem with CReyTril(res38res43); CRey with OriLTril(res44res48); OriL with EReyTril(res49res52); ERey with EscTril (res53res54); Leit with EscTril (res56res57); Esc with Tril (res58); MODEL PRIORS: res1res58 ~N(0, 0.01); f1varf4var ~IG(0.001,0.001); reg1reg8 ~N(0, 0.01); idvar ~IG(0.001,0.001); escvar ~IG(0.001,0.001); id ~N(0,100000); es ~N(0,100000); idesc ~N(0, 0.01); 


The best way to check an input is to run it and see if you get what your expect. 


Thanks Dr. Linda! I am sorry I was not clear enough: my doubts are about the MODEL PRIORS. My data fits well in frequentist analysis (SEM), but not in Bayesian (BSEM), maybe because PRIORS mistakes. It runs, but PPpvalue is about 0.01. I would like to use Bayesian because the small sample. I tried to use OpenBUGS software, but I think MPlus is much easier (however, Bayes analysis is more complicated to me). May I use the svalues obtained in OUTPUT command as PRIORS? Are they more informative that those ones I posted? Thanks for your attention! 


If you are a beginner Bayesian, I would go by the Mplus defaults and not give priors. Setting such specific priors is not for the beginner Bayesian. If you haven't already, you should study my paper on our website Muthén, B. (2010). Bayesian analysis in Mplus: A brief introduction. Technical Report. Version 3. Click here to view Mplus inputs, data, and outputs used in this paper. and also carefully look at the Mplus scripts for that paper as well as for the paper Muthén, B. & Asparouhov, T. (2012). Bayesian SEM: A more flexible representation of substantive theory. Psychological Methods, 17, 313335. Click ""download paper"" below for the latest version of October 21, 2011. Download the 2nd version dated April 14, 2011. Click here to view the seven web tables referred to in the paper and here to view Mplus inputs, data, and outputs used in this version of paper. Download the 1st version dated September 29, 2010 containing a MIMIC section and more tables, and the corresponding Mplus inputs, data, and outputs here. The seven web tables correspond to tables 8, 10, 17, 18, 19, 20, and 21 of the first version. Note also that priors need to be chosen in relation to the scale of the variables. 


Thank you very much for your detailed answer! Certainly, I will read your papers (and cite them)! 

Tom Young posted on Tuesday, February 06, 2018  2:24 am



Dear Dr Muthen, I am wondering whether I could have some advice on residual variances in a BSEM... I am in the process of looking at the factorial validity of a 73 item questionnaire. Some of the factor loadings are not great (below 0.60), so I am thinking of taking them out. But I am wondering whether there is any guidance on residual variances on variables? Is there a threshold as such for them, as in a value being too high or low, or acceptable? I cant seem to find any references that show this. Kind Regards Tom Young 


I don't know that there is a literature on this. Although there is some discussion of loading sizes in Cudeck, R., & O’Dell, L. L. (1994). Applications of standard error estimates in unrestricted factor analysis: Significance tests for factor loadings and correlations. Psychological Bulletin, 115, 475– 487. doi: 10.1037/00332909.115.3.475 

Back to top 

