| Message/Author |
|
|
|
|
| I'm running a DSEM-CFA for my dissertation and trying to use Hamaker's syntax for examining differences in factor loadings across the between and within levels. However, I'm running into an issue with finding a consistently invariant item to use as the anchor item. I'm aware of a method for identifying invariant item sets by Cheung and Rensvold (1999), but given the more complex nature of DSEM, wanted to make sure that there are not better methods for identifying the best anchor item. Please advise! |
|
|
|
|
You don't need an anchor item. The hypothesis about equality of loadings is equivalent to the hypothesis of proportional loadings when you fix the factor variances to 1.
%within%
f by y1-y4 (w1-w4); f@1;
%between%
fb by y1-y4 (b1-b4); fb@1;
model test:
0=w1*b2-b1*w2;
0=w1*b3-b1*w3;
0=w1*b4-b1*w4; |
|
|
|
|
Just a little correction on the above setup. The model should be specified as
%within%
f by y1-y4*1 (w1-w4); f@1;
%between%
fb by y1-y4*1 (b1-b4); fb@1;
to free up all the loadings. |
|
|
|
|
I'm trying to run a DSEM-CFA model with three continuous indicators:
Usevariables are TSSID Y1 Y2 Y3 ;
CLUSTER = TSSID;
Define:
Y1 = Intensity1;
Y2 = Intensity2;
Y3 = Intensity3;
Y4 = Intensity4;
Analysis:
Type = twolevel random;
estimator = bayes;
processors = 2;
biteration = (2000);
Model:
%within%
f by y1-y3(&1);
s| f on f&1;
logv | f;
%between%
fb by y1-y3*;
fb@1;
fb s logv WITH fb s logv;
The error message I'm getting is this:
*** FATAL ERROR
LAG CAN NOT BE GREATER THAN THE NUMBER OF OBSERVATIONS FOR EACH CLUSTER.
Can you help me understand this error? |
|
|
|
|
To run a time-series model / DSEM model you must have more than one observation in each cluster, and in fact we would recommend at least 6 observations per person as you have 5 subject specific parameters.
As a first analysis I would recommend that you run a simpler model
%within%
f by y1-y3(&1);
f on f&1;
f;
and exclude all clusters with 5 observations or less. You should also run the standard two level as a very first step
Model:
%within%
f by y1-y3*1;
f@1;
%between%
fb by y1-y3*1;
fb@1; |
|
|
|
|
How should I cite the recommendation that every clusters should have more observations than the number of
subject-specific parameters? Also is it recommended to always allow random effects to co-vary at level 2? Thank you very much! |
|
|
|
|
We don't have a paper that addresses that exactly and the recommendation as stated above is not clear cut. Ideally you want more observations than parameters for any model, and that carries over to the cluster specific model as well. However, in many situations, information from the rest of the population can inform the really small clusters. To be more specific. Ideally, you would want cluster sizes to be larger than the number of cluster specific parameters. If that is not attainable, we would recommend that no more than 10% of the cluster sizes violate that minimum. If that is not satisfied we would expect possible convergence problems or dramatic loss of power.
Generally, it is preferable to have all random effects co-vary at level 2. If that is causing convergence problems we would recommend a one factor model to pick up at least most of the random effect covariance. |
|
|
| Back to top |
