Mlogit in Mplus PreviousNext
Mplus Discussion > Categorical Data Modeling >
Message/Author
 Andy Ross posted on Thursday, October 27, 2005 - 3:48 pm
Hello

I am running a standard mlogit regression model and need to get the following output:

1. The standard loglikelihood test so that i can see whether my final model is significantly better than a model with no predictors.

2. And the loglikelihood test for each of the predictor variables, so that i can see whether their inclusion significantly improves the model, or not.

Many thanks for your support

Andy
 Linda K. Muthen posted on Thursday, October 27, 2005 - 6:08 pm
How are you running this? The reason I ask is that you get a loglikelihood for the model being estimated automatically. Mplus does not provide number 2 automatically.
 Andy Ross posted on Friday, October 28, 2005 - 9:55 am
Many thanks for your reply.

This is the input file for my model:

VARIABLE: NAMES ARE u1-u16;
USEVARIABLES ARE u1-u5;
NOMINAL IS u2;
CATEGORICAL ARE u1 u4;

MODEL: u2#1 u2#2 u2#3 u2#4 u2#5 ON u1 u3 u4 u5;

And the output:


TESTS OF MODEL FIT

Loglikelihood

H0 Value -10483.204

Information Criteria

Number of Free Parameters 27
Akaike (AIC) 21020.408
Bayesian (BIC) 21191.368
Sample-Size Adjusted BIC 21105.573
(n* = (n + 2) / 24)


Many thanks

Andy
 Linda K. Muthen posted on Friday, October 28, 2005 - 1:37 pm
So you get number 1 from this output and number 2 is not available.
 Andy Ross posted on Monday, November 07, 2005 - 11:12 am
Many thanks for your reply.

Maybe i'm reading the output wrong but i was hoping for a significance test/p-value telling me whether the full model was significantly better than the model with no predictors.

Do i need to calculate this by hand from Loglikelihood HO value? And if so, how would i go about doing this? My apologies if this sounds a little elementary. I have asked around but everyone here is used to the standard SPSS/Stata output...

Many thanks

Andy
 Linda K. Muthen posted on Sunday, November 13, 2005 - 12:35 am
You should run two models both with the covariates in the model: (1) a model with the slopes of the covariates free and (2) a model with the slopes of the covariates fixed to zero. You can then compute the loglikelihood difference and -2 time that is the chi-square difference.
 Manuel posted on Monday, January 09, 2006 - 12:00 pm
Hello,
I have a similar question: I am
running a standard discrete time survival analysis. Without covariates I obtain a LL which is identical to that of other programs (i.e., Systat etc.).
However, after including a single continuous covariate (proportional hazard assumption) the values differ vastly - how come (Mplus: -540.28 vs. SPSS/Systat:-1167.46)? THANK YOU VERY MUCH IN ADVANCE!!!
 Linda K. Muthen posted on Monday, January 09, 2006 - 12:43 pm
When there are covariates in the model, the loglikelihood values are on a different scale. This is why to compare nested models you need the covariates in both models.
 Manuel posted on Monday, January 09, 2006 - 1:35 pm
thank you very much for your prompt reply! However, please let me rephrase my question: Should the chi-square diff test (effect vs. no effect of the covariate; covariate is part of both models but fixed to zero in the more restrictive model both models are nested) be identical to the chi-square diff test in the cox regression model (using any other program)?

Part of the reason I am asking is that Muthén & Masyn (2005, p. 36) note that Mplus can handle continuous covariates, while the traditional log-linear framework only allows categorical variables. I was not aware of that fact - but that should be reflected in the LL, shoud not it?

Thank you so much for your help!
 Linda K. Muthen posted on Monday, January 09, 2006 - 2:33 pm
I don't know if the loglikelihoods would be the same. The Mplus loglikelihood is for y given x. If the cox regression loglikelihood is for y and x, then they would be different. I think the difference in the loglikelihoods would be the same however. Also, the Cox regression model usually means continuous-time survival. Mplus estimates a discrete-time survival model. I am assuming you are doing a discrete-time survival model.

Yes, that should be reflected in the loglikelihood.
 Manuel posted on Monday, January 09, 2006 - 3:37 pm
that helps - thank you very much!
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: