Crossclassified twin model
Message/Author
 Siny Tsang posted on Monday, May 11, 2015 - 4:42 pm
Hello,

We are trying to run a multi-level mediation on some twin data. The mediator (M) and outcome (Y) variables are measured at the individual level (L1), whereas the X variable is measured at the census level (L2).

I was able to set up the model using TYPE = CROSSCLASSIFIED as follows, but we are not quite how to include the twins into the model. I understand that we cannot use multiple grouping with CROSSCLASSIFIED; is there another way to approach this?

%WITHIN%
y m;
y ON m;
%BETWEEN census%
x y m;
m ON x (a);
y ON m (b);
y ON x (cp);
%BETWEEN twinid%
x y m;
m ON x (a);
y ON m (b);
y ON x (cp);

model constraint:
new (indirect direct);
indirect = a*b;
direct = cp;
 Bengt O. Muthen posted on Monday, May 11, 2015 - 6:42 pm
You don't have to use up one level for the twin nesting - instead you can handle them in a wide format. Or, perhaps you are saying that the twins split into different census places?

If the latter (they split), perhaps you want to first try to do twin modeling in a 2-level fashion to get the hang of it.

If the former (they don't split), why not do it as 2-level with twins done in wide format. Then you can have groups (MZ, DZ).
 Siny Tsang posted on Wednesday, May 13, 2015 - 9:20 am
Dr. Muthen,

Yes, the twins may not always be in the same census places, which is why we modeled using TYPE = CROSSCLASSIFIED.

Is there a way we can further the MZ/DZ in this model?

Our model is now set up as

VARIABLE: ...
CLUSTER IS census twinid;
ANALYSIS: TYPE = CROSSCLASSIFIED;
ESTIMATOR = BAYES;

MODEL:
%WITHIN%
y m;
y ON m;
%BETWEEN census%
x y m;
m ON x (a);
y ON m (b);
y ON x (cp);
%BETWEEN twinid%
x y m;
m ON x (a);
y ON m (b);
y ON x (cp);

model constraint:
new (indirect direct);
indirect = a*b;
direct = cp;
 Tihomir Asparouhov posted on Thursday, May 14, 2015 - 12:27 pm
You can run multiple group by setting up the groups in wide format. For two groups instead of having Y, you would have Y1 and Y2 and the data will have one value and one missing value depending on which group it is in.

You can add MZ/DZ modeling by expressing the ACE components as dependent parameters in model constraints.

I would recommend you drop the path analysis and settle down the univariate model for Y as a first step.

I would also recommend a totally different approach (as a cross verification) where you run a two level model first with census as the cluster variable - estimate the effect of Census (this is the between part of the variable) subtract that from the variable and then proceed with standard twin modeling.
 Siny Tsang posted on Sunday, May 17, 2015 - 5:10 pm
Dr. Asparouhov,

Would you mind clarifying the last part of you response? Perhaps I am missing something, wouldn't the "effect" of census be a variance? How do we subtract that out? Or do you mean that we should subtract the census mean from the observations of Y?
 Tihomir Asparouhov posted on Monday, May 18, 2015 - 2:45 pm
Yes it is a variable. It is the between part of the variable. In a model like this

%within%
y
%between%
y !<- this is the clusters effect

You have to request factor scores using

savedata: file is 1.dat; save=fscore;

The between part of Y is the variable B_Y. If you are using the Bayes estimator the variable B_Y gets multiple values and further analysis should be based on type=imputation, see example 13.13