ML and WLSMV estimates PreviousNext
Mplus Discussion > Categorical Data Modeling >
Message/Author
 Adrienne Tin posted on Wednesday, April 12, 2006 - 7:59 am
I have a very simple regression with a categorical outcome. ML and WLSMV give very different estimates. How should I interpret them?

ML estimates of the predictor

Estimates 0.513
S.E. 0.249
StdYX 0.272

Thresholds
Outcome$1 2.049

Why is StdYX differ from the Estimates?

WLSMV estimates (same model)
Estimates 0.283
S.E. 0.143
StdYX 0.274


Why is the WLSMV estimate (0.283) so different from the ML estimate (0.513)?

Thanks in advance.
 Linda K. Muthen posted on Wednesday, April 12, 2006 - 8:06 am
In Mplus, the default for maximum likelihood is logistic regression and the default for weighted least square is probit regression. These differ by a scale of approximately 1.7.
 Adrienne Tin posted on Wednesday, April 12, 2006 - 3:05 pm
Hi, Linda, thanks for the response. Now I remember reading about WLSMV and probit model somewhere.

Could you also explain why stdYX is different from the estimate?
I read that they are different only when there are multiple dependent variables. In this simple model, there is only one dependent variable.

Thanks
 Linda K. Muthen posted on Thursday, April 13, 2006 - 9:37 am
As long as the variances of y and x are not one, STDYX will be different that the raw parameter estimate for both univariate and multivariate regression.
 Lydia August posted on Friday, September 01, 2017 - 6:11 am
1) Regarding scale setting, can you use the fixed factor method (set latent variances to 1) when using the WLSMV estimator? If so, is it appropriate to interpret either the raw or STDYX estimates (since they should be the same)?

2) I ran the same model (with ordinal indicators) using WLSMV to get fit-statistics and then with MLR to get BIC to compare non-nested models and used the fixed factor method. I'm not sure why I am getting such different parameter estimates when using the MLR estimator (they are even different in terms of significance and directionality!).

For example, here is some WLSMV output:

Estimate S.E. Est./S.E. P-Value

EFFICACY BY
E01 0.623 0.054 11.497 0.000

...

SATISFAC BY
S01 -0.027 0.067 -0.405 0.686

...

EFFICACY WITH
SATISFACTI -0.854 0.066 -12.931 0.000

And here is some MLR output:

Estimate S.E. Est./S.E. P-Value

EFFICACY BY
E01 1.507 0.347 4.340 0.000

...

SATISFAC BY
S01 1.227 0.388 3.163 0.002

...

EFFICACY WITH
SATISFACTI -0.002 0.253 -0.009 0.993
 Bengt O. Muthen posted on Friday, September 01, 2017 - 1:08 pm
1) I assume you have a categorical DV and the default Delta parameterization of WLSMV in which case the DV variance is 1. If so, the answer to your question is yes.

2) First check that your 2 runs have the same number of parameters. Then not that WLSMV uses probit and ML uses logit by default (but can use link=probit). Note also that if all loadings of a factor are negative you can switch them all to positive.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: