|
|
 |
| LGM with a large fraction of missing ... |
 
|
| Message/Author |
|
|
|
|
I am trying to establish the equivalence of long and wide form unconditional latent growth curve with a large fraction of missing data , before moving on to further analyses. I am using exactly the same data in both analyses, but the extent of missing data is such that when I reshaping the data to the wide form, coverage drops to zero in some places.
Estimation terminates normally for the long form model, using analysis TYPE = COMPLEX, because the data are clustered by groups (grp) , and ESTIMATOR = MLR. I am concerned that, while they are continuous, they are not normally distributed. I get a Loglikelihood H0 Value = -1462.191, with a H0 Scaling Correction Factor for MLR = 1.110, which suggests that my concerns about normality have some foundation.
Estimation also terminates normally for the wide form model, with the same specifications. However, now I get a considerably lower Loglikelihood H0 Value = -1279.668, with a much higher H0 Scaling Correction Factor for MLR = 1.534. I also get the following error message:
THE COVARIANCE COVERAGE FALLS BELOW THE SPECIFIED LIMIT.
THE MISSING DATA EM ALGORITHM WILL NOT BE INITIATED.
CHECK YOUR DATA OR LOWER THE COVARIANCE COVERAGE LIMIT.
THE MINIMUM COVARIANCE COVERAGE WAS NOT FULFILLED FOR ALL GROUPS.
I have tried reducing the COVERAGE as far as zero, but I found that I lose the Scaling Correction Factor below values of .01.
Many thanks in advance. |
|
|
|
|
| The wide format can have convergence issues with large amounts of missing data. In this case, the long format works better. |
|
|
|
|
| It turned out that I could resolve the problem by collapsing the data on the clustering variable, which increased the apparent coverage. |
|
|
| Back to top |
|
|
