| Message/Author |
|
|
| Scott Weaver posted on Tuesday, February 07, 2006 - 12:22 pm
|
|
|
Hello,
I am seeking clarification of a sentence on pg. 340 under 'Missing Data Analysis' of the Mplus manual (version 3): "Robust standard errors and chi-square are available for all outcomes using the MLR estimator."
This statement had led me to believe that if I specify Type=Missing and Estimator=MLR, then I would receive robust SE estimates under MAR. However, I was recently discussing this with someone who is certain that Mplus is implementing the procedure outlined in Yuan & Bentler (2000), to provide robust SE with missing data - but that Yuan & Bentler maintain that that they are assuming MCAR. Am I to assume that when data are nonnormal, the robust SE and chi-square statistic hold only under MCAR (though they may perform fairly well under MAR)?
Thanks in advance for the clarification.
Scott |
|
|
| bmuthen posted on Wednesday, February 08, 2006 - 10:52 am
|
|
|
The exact condition for which MLR is correct are these: If [Y|X] is non-normal and the missing patterns are either MCAR or MAR but only predicted by X and not Y then MLR gives valid results. This condition in between MCAR and MAR.
see for example the last two simulations in Web Note 2
http://statmodel.com/examples/webnotes/mc2c.html
http://statmodel.com/examples/webnotes/mc2d.html |
|
|
|
|
I am estimating a growth model with a continuous non-normal observed outcome that has some missing data. I am currently using the Type=Missing command in conjunction with Estimator=MLR, but this combination does not provide traditional fit indices to evaluate the fit of the growth curve.
I have looked at the model using the MLR estimator with listwise deletion and it a linear growth model fits the data well. However, it seems weird to present fit statistics for a listwise deleted model and then use an MLR estimator to handle missingness for the substantive model.
Suggestions?
Thanks!
Dustin |
|
|
|
|
| Try TYPE=MISSING H1; You should then get the fit statistics you want. |
|
|
|
|
Thanks Bengt. Beginners mistake on my part. Can you suggest a reference that discusses the theoretical and statistical differences between using the "missing" and "missing h1" commands in Mplus?
Dustin |
|
|
|
|
| Sometimes estimating the unrestricted model can take time and be difficult when there is a lot of missing data so we allow the users the option of not estimating it or not. |
|
|
| Back to top |
