Maria Wong posted on Wednesday, October 05, 2005 - 10:40 am
Hello, when latent intercept and slope factors are used to predict binary outcomes (e.g., onset of alcohol use at a certain age with 0= no onset; 1 = onset), does MPLUS transform the binary outcome variables to log odds? All outcomes are manifest variables.
M. Wong posted on Friday, October 07, 2005 - 7:19 am
To further clarify my question, repeated measurements of a set of continuous variables are used to estimate the latent intercept and slope factors (e.g., developmental trajectories of resiliency). Then these factors are used to predict binary outcome variables of substance use. Does MPLUS transform these outcomes to log odds?
In the regression of the binary outcome variable on the continuous growth factors, a probit regression coefficient is estimated in Mplus if the weighted least squares estimator is used. A logisitc regression coefficient is estimated if a maximum likelihood estimator is used.
continuing that conversation, when regressing a binary outcome on the intercept and slope of a single class growth model, why does the regression estimate stay constant for the intercept , but it changes for the slope, comparing models where the slope is centered at different time points. I was also expecting the regression estimate for the intercept to change given that the centering option changes the definition of the intercept. Thanks.
I restructured the data prior to this analyses into a cohort sequential design. due to some sparse data I have correlations between d1 and d4, for example, of 1.0 which is causing problems with convergence. here is the message I'm getting:
WARNING: THE SAMPLE CORRELATION OF D1 AND D4 IS 1.000 DUE TO ONE OR MORE ZERO CELLS IN THEIR BIVARIATE TABLE. INFORMATION FROM THESE VARIABLES CAN BE USED TO CREATE ONE NEW VARIABLE.
WARNING: THE SAMPLE CORRELATION OF D1 AND D5 IS 1.000 DUE TO ONE OR MORE ZERO CELLS IN THEIR BIVARIATE TABLE. INFORMATION FROM THESE VARIABLES CAN BE USED TO CREATE ONE NEW VARIABLE.
NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED.
anonymous posted on Wednesday, November 23, 2005 - 4:42 am
p.s. - Each individual has a max of 4 waves of data (usually consecutive) which is part of the sparse data problem - so few people at d1 also have data at d5...
For weighted least squares, it is recommended that each cell in the bivariate tables for pairs of variable contain several observations. I would recommend using maximum liklihood given your situation. You can start out using Monte Carlo integration and if you find that some of your growth factors have small variances, you could fix their variances to zero to reduce your dimensions of integration.
Thanks. This is very helpful. For us non-statisticians, is there an accessible reading that provides an introduction to numerical integration?
bmuthen posted on Wednesday, November 30, 2005 - 9:36 am
Numerical integration in the ML context is described in technical-statistical terms - if you want those ref's I can give them. I would think an introductory numerical analysis book would be a reasonable and somewhat less technical way to understand the basic nature of the topic - but then it is not tied to the applications in statistics. I don't have such ref's to suggest since I studied this back in Sweden. In our annual week-long Mplus course we go through some slides on the topic and those are in our handouts (see web site for how to get them).