Message/Author 

lamjas posted on Thursday, January 10, 2013  8:10 am



I am using TIMSS database and want to do a SEM (e.g., X > M > Y). Y is achievement. In the dataset, however, achievement is represented by five plausible values (say pv1, pv2, pv3, pv4, pv5). I understand that they should not be treated as indicators in the model. I read some threads and it seems that I need to do "multiple imputation" (In Mplus, using Type = Imputation). However, I am still very confused the appropriate procedures to deal with a SEM model with plausible values as an outcome variable. Could anyone explain step by step in layman terms? Is UG Ex 11.7 exactly the way I use imply on my model (of course, I need to change from LGM to SEM)? Thank you!! 


The second part of UG ex11.8 (pp. 405407) shows how to use DATA TYPE=IMPUTATION. More information about plausible values are found in Asparouhov, T. & Muthén, B. (2010). Plausible values for latent variables using Mplus. Technical Report. which also refers to a report by von Davier et al (2009) that is more applied. 

lamjas posted on Thursday, January 10, 2013  5:13 pm



Hi Bengt, Thanks for your quick reply. Let's say I am doing a SEM (X>M>Y) and Y is represented by five plausible values. ===== So, the first step is to do a multiple imputation and construct five datasets (UG Ex 11.5). Syntax as follows: Data: File is raw.dat; Variable: Names = x1x5 m1m5 pv1pv5 w clus; Missing = ALL (99); Weight is w; Cluster is clus; Data imputation: Impute = x1x5 m1m5 pv1pv5; Ndataset = 5; Save = mp*.dat; Analysis: Type = Complex; Estimator = MLR; Output: Tech8; ===== Then, the second step is the second part of UG Ex 11.8 and examine the model. Syntax as follows: Data: File is mplist.dat; Type = imputation; Variable: Names are x1x5 m1m5 pv1pv5 w clus; Weight is w; Cluster is clus; Analysis: Type = complex; Estimator = MLR; Model: x by x1x5; m by m1m5 y by pv1pv5; y on m x; m on x; output: tech1 tech4 standardized; ======= Are these two steps correct to do a SEM with plausible values? Any comments are welcome. Thanks again. 


No, you don't do the first step. That has already been done by TIMSS when they gave you the 5 plausible values. 

lamjas posted on Thursday, January 10, 2013  6:35 pm



Then, what if I have missing data of x1x5 and m1x5? Should I skip step 1? 


One simple approach is to let "FIML" handle that without imputation. 

Huang Wu posted on Saturday, June 09, 2018  3:32 pm



I am using PISA data for analysis. These discussions are very helpful but I am still have some questions. (1) I could not find "FIML" in the estimator; (2) PISA give us 10 plausible value for each subject. Should us create 10 dataset first (each one only have one plausible value for different subject?) or just keep them in a dataset and treated them as indicators? (3) I have read lamjas's post and he/she said we could not treat plausible as indicator but in his/her code, he/she load 5 plausible value in y. If one dataset only contains one plausible value, how can we load them together? Thanks 


1) Use ML or MLR 2) Create 10 data sets and use Data: Type=Imputation. 

Back to top 