|
|
 |
| Time score estimation and GMM |
 
|
| Message/Author |
|
|
| socrates posted on Wednesday, February 11, 2009 - 2:49 am
|
|
|
I tried to run a GMM in my longitudinal dataset. However, even fitting a conventional or one-class (linear or quadratic) LGM did not fit well. Looking at the data, I recognized that almost all of the change is between T0 and T1 (month 0 and 1), while between T1 and T6 there was almost no further change.
To get a LGM that is able to deal with this bend in the data, I think of the following procedure:
I first estimated the time scores of T0 trough T6 in a LGM:
i s |
T0@0
T1@1
T2*2
T3*3
T4*4
T5*5
T6*6;
This step yielded the following results:
s |
T0 0.000
T1 1.000
T2 1.094
T3 1.155
T4 1.214
T5 1.155
T6 1.164
I then went on with the following specification:
MODEL: %OVERALL%
i s |
T0@0
T1@1
T2@1.094
T3@1.112
T4@1.129
T5@1.147
T6@1.164;
Here T2 is fixed on 1.094 as estimated before. For T3 trough T6, in each step the mean of the differences between subsequent time score estimations (i.e., 0.018) is added.
However, does this approach provide a sound basis (i.e., a LGM) on which I can now build GMMs? |
|
|
|
|
| That seems a bit ad hoc. How about instead using a 2-piece growth model (see our handouts for Topic 3)? You would have a first piece for the first 2 time points with only the intercept being random (the slope fixed). And a second piece for the rest. Neither piece would have estimated time scores. |
|
|
| Back to top |
|
|
