|
|
| Message/Author |
|
|
| JW posted on Sunday, November 04, 2007 - 9:32 am
|
|
|
Hi, I'm using a MLM estimator with data that is missing at random. I've imputed the following in the VARIABLE command
VARIBALE: USEVAR ARE DSRS SSSMINUS CNCEQ CESBQ;
MISSING ARE ALL (999);
In the analysis command I entered:
ANALYSIS: ESTIMATOR IS MLM;
TYPE=MISSING H1;
If I don't enter TYPE=Missing, my N drops and it seems to do a list-wise deletion. If I include the Type=missing, it uses the full sample, but I get the following error:
*** ERROR in Analysis command
Estimator MLM is not allowed with TYPE = GENERAL MISSING.
Default will be used.
1 ERROR(S) FOUND IN THE INPUT INSTRUCTIONS
How do I include the full sample, using MLM and account for missing data? |
|
|
| Paul Silvia posted on Sunday, November 04, 2007 - 10:40 am
|
|
|
For TYPE = GENERAL MISSING H1 and continuous outcomes, only ML, MLR, and MLF are available (see User Guide p. 424).
If you use the MISSING option, then the ML estimator is full-information ML, so ML or MLR is probably what you want here. |
|
|
| JW posted on Monday, November 05, 2007 - 5:37 pm
|
|
|
| Thanks. I think I'm confused as I do have non-normal data, and should use the robust maximum likelihood estimator, right? How should I handle missing data with non-normal, continuous data? |
|
|
| Paul Silvia posted on Monday, November 05, 2007 - 6:12 pm
|
|
|
One would need to know a lot about the nature of the non-normality (as well as other assumptions that may be unmet), but my preference would be to add the BOOTSTRAP option to the ANALYSIS command, a la:
TYPE = MISSING H1;
ESTIMATOR = ML;
BOOTSTRAP = 1000; (see UG p. 434).
And under OUTPUT, add:
CINTERVAL (BOOTSTRAP);
This will handle the missing data as well as give you bootstrapped standard errors and confidence intervals. |
|
|
| JW posted on Tuesday, November 06, 2007 - 5:25 am
|
|
|
Thanks! This is helpful.
James. |
|
|
| JW posted on Wednesday, November 07, 2007 - 5:11 am
|
|
|
| Actually, I do have a f/u up question. Is there a way to include missing data with MLM estimator? So, when I did the analyses with boostrapping as you mentioned above, the full N was used, but when just with MLM as an estimator and no missing data command (as the error message occurs), cases were not included and my N dropped. |
|
|
|
|
| No. Only listwise deletion is allowed with MLM. |
|
|
| Back to top |
|
|
