Message/Author |
|
|
I am doing an 2-1-1 mediation model using Preacher's syntax. I am new to this and wonder can the '2' part of the model be a binary variable (in this case trained vs. not trained) and if so does one need to enter this as a CATEGORICAL variable? thanks |
|
|
The 2 part can be categorical. It should not be specified as categorical because it is an observed exogenous variable. |
|
|
Many thanks. that is helpful. Can I please ask about sample size? I suspect you get asked this often! For the 2 variable I have two clusters (39 vs. 23). However, within clusters there are between 1 and 4 people (average about 3 per cluster). That is because the study involved project groups with maximum size 4. I have seen in a previous message that you recommend including clusters with 1 in (as long as less than 15% sample, as is this case) as it contributes to the between-subject effect. However, what are your views with such a low number per cluster?? thanks Robin |
|
|
This need not be a problem. You will see the consequences in terms of the SEs. |
|
|
many thanks again. Wonder if you know of any reference(s) supporting low cluster numbers and inclusion of clusters with one. Just knowing journals like citations to back up things! thanks Robin |
|
|
None comes to mind - maybe check the simulations done by Joop Hox' team, or email him directly. |
|
|
Hi Drs Muthen, I’m working with 2-1-1 models and haven’t yet found any examples whereby the Within y ON m is incorporated into the total or indirect effect. I’ve been through Dr. Preacher’s papers and the code on his site but could be missing something because I don’t think conflation is a concern in my case. It’s meaningful for my analyses to include the Within level y ON m in the mediation model as I expect that the relationship between m and y is not the same for all cases within a cluster. I include some individual level controls that are important for the relationship of m ON x from other multilevel analyses and are as well important for y ON m. Does the following code appropriately allow for this? For context, x is a neighborhood disadvantage, m is parent substance use, y is child substance use, and the controls are family demographics. Thanks in advance! VARIABLE: Within = m controls; Between = x mmean; DEFINE: mmean = CLUSTER_MEAN(m); CENTER m(GROUPMEAN); MODEL: %Within% y ON m (b2); y m ON controls; %Between% mmean ON x (a); y on mmean (b1); y on x (c); MODEL CONSTRAINT: NEW(IND B C TOTAL); IND = a*(b1+b2); B = b1+b2; C = c; TOTAL = (a*(b1+b2)) + c; |
|
|
This looks correct. A somewhat better approach is to use a latent variable decomposition of M (so not just for Y). This means that you don't use mmean and don't put M on the Within list. |
|
Back to top |