Crossclassified Longitudinal Model PreviousNext
Mplus Discussion > Multilevel Data/Complex Sample >
Message/Author
 Joseph Bonito posted on Thursday, June 27, 2013 - 7:47 pm
Hi,

I have data from small group discussions that I think is cross-classified but not hierarchical. Participants took part in 3 separate discussions but were randomly assigned to different groups for each discussion. The DV is the number of speaking turns for each person for each of the three discussions. So person 1 has DV1, DV2, and DV3 but belonged to different groups each day. Ideally, I'd like variance estimates attributable to the group for each of the three discussions, as well as correlation estimates for each person's participation.

I think the CROSSCLASSIFIED routine should handle this but I can't tell from the documentation if (a) the CC command handles more than two clusters, and (b) if the clustering has to be hierarchical (e.g., the Hox example of middles schools crossed with high schools).

Thanks in advance.

Joe
 Bengt O. Muthen posted on Thursday, June 27, 2013 - 10:17 pm
Yes, sounds like Cross-classified would handle this. You can have only 2 cluster variables in CC, so not a higher clustering beyond these 2. The clustering is not hierarchical - I don't think that that Hox example is hiearchical.

You will get variance estimates corresponding to the 2 clusters. I don't see where a correlation would come from.
 Joseph Bonito posted on Thursday, June 27, 2013 - 11:21 pm
Thanks, Bengt. This begs the question of data setup and subsequent analysis. I think the data file would have to be "wide" with group ID columns for each discussion session. So person 1 would have DV1, DV2, and DV3 and then three columns for the IDs of the groups she was in. Those columns for the group IDs at T1, T2, and T3 indicate the clusters. So if I understand your response, I'm limited to just 2 cluster variables which makes the analysis problematic.

On the other hand, I could have a long file with three records for each participant, with one column for the DV, another for group (which varies within each participant because of the different groups they belonged to), and another that identifies which discussion session (1, 2, or 3). If that's true, then I'm not sure how to handle the analysis that gets different variance estimates attributable to groups at each time given there is only one column that identifies groups.

As for correlations, I want to estimate if, for example, a person that spoke frequently during the first session spoke frequently during the second session, and so on.

Thanks again for your help.

Joe
 Bengt O. Muthen posted on Friday, June 28, 2013 - 9:46 pm
I was thinking long format (your second paragraph). In line with UG ex 9.24. That's why I can't think of a correlation.

But are you sure you can't view this as a single-level analysis with multiple measures for each subject?
 Joseph Bonito posted on Saturday, June 29, 2013 - 4:25 pm
Thanks again for your help.

The problem is that a single-level analysis ignores the group effect/variance. The group effects will vary for each discussion because participants belong to different groups for each discussion. So participants are cross-classified by group, and each participant belongs to three groups, one for each discussion section.

UG 9.24 is close but I still have the same problem--3 days means three different groups, which means three different levels. The problem seems akin to an educational study with students tracked from elementary school, through middle school, into high school. That's 3 cross-classification levels, correct?
 Joseph Bonito posted on Sunday, July 07, 2013 - 3:45 pm
Hi Bengt,

After some searching I found a solution to my problem that incorporates SAS (which is what I generally use). Please see:

http://psych.unl.edu/psycrs/945/12b_Changes_in_Nesting_Example.pdf

The "acute transfer" model (Model 3) is most relevant. I'd really like to implement this solution in MPlus but not sure if it's feasible given the number of random statements needed.

Joe
 Bengt O. Muthen posted on Tuesday, July 09, 2013 - 2:11 pm
Is this acute transfer model described statistically in any paper? Is it in Raudenbush & Bryk?
 Joseph Bonito posted on Tuesday, July 09, 2013 - 4:40 pm
Hi Bengt,

Yes, that's the original citation, at least according to the website, which mentions pp. 392-393 in R&B. I'm on vacation so my copy of R&B isn't handy to verify. The model does a good job of reproducing variance attributed to groups at each time when run separately, as well as includes within person variances/covariances across time.
 Bengt O. Muthen posted on Tuesday, July 09, 2013 - 4:54 pm
So is the R&B model different from the mobility model in the JEBS article we refer to in our handout?
 Joseph Bonito posted on Tuesday, July 09, 2013 - 6:09 pm
Do you mean the last 4 slides of the "New Developments" presentation? Yes, I read through that with great interest. That prompted my original question in this thread. The only difference, from what I can tell, is that you have 2 cross classifications (middle school and high school) whereas my data has three (participants were members of different groups at each measurement period).
 Bengt O. Muthen posted on Tuesday, July 09, 2013 - 10:00 pm
I think we are currently limited to moving between two classifications.
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: