GMM output
Message/Author
 B Lee posted on Saturday, March 17, 2007 - 6:50 am
I am running a ZIP growth mixture model and have 2 questions:

1. In output, means of intercepts for some classes are negative even though the range of outcome values are never negative. The class intercept means don't seem to reflect the intercept point expected by the graph. Is it possible the class intercept means are just the difference from the overall intercept mean? How can I get actual group means on output?

2. In estimating a linear model with intercept and/or slope variance freely estimated, the resulting plot shows a curve rather than a line. However, a quadratic term (when added) is not significant. How does a linear term generate a curved rather than straight line? The LCGA model of same data has straight lines in plot.

Thanks!
 Linda K. Muthen posted on Saturday, March 17, 2007 - 9:11 am
1. These are in a log-rate scale.

2. The model is linear in the log rate but non-linear in the means which is what are plotted.
 B Lee posted on Saturday, March 17, 2007 - 12:28 pm
I'm not sure I understand. Are you saying that e^(intercept mean listed on output) will reflect value plotted on graph at Time 0?

This seems approximately true for intercept, but not for slope. The value of the slope plot seems to change at each time-point. How can this be calculated?
 Bengt O. Muthen posted on Saturday, March 17, 2007 - 1:20 pm
You get the estimated mean if you request RESIDUAL in the output. A dependent variable that is Poisson has mean = exp(log rate), where log rate is the dependent variable in the model, i.e. it is a linear function of the growth factors. With zero-inflated Poisson, the mean is modified by the inflated part. For ZIP literature, see, e.g., stat text books and also the Roeder et al (1999) ZIP mixture article in JASA. The zero-inflation is taken into account in the means printed in the RESIDUAL output. It is the values from the RESIDUAL output that are used on the y axis of the plot.

I don't know what you mean by "the slope plot", but I hope the above answers your questions.