Hi Linda & Bengt, I have foster children nested within case-workers. I want to examine outcome variables of legnth of care (continuous) and home placement (dichotomous), predicted by case-worker race. Case workers are either white, black, or hispanic. Do you think it's better to model these races (at level 2) as either 3 dichotomous variables, a single categorical variable, or use a multigroup analysis (none of my hypotheses involve mediation)? If multigroup, how would do I tell Mplus to hold variable means constant across groups so that I may do chi-square difference tests?
If case-worker race is a covariate, then you would need to create two dummy variables to represent the three races. Covariates can be binary or continuous not nominal.
If you have enough observations to do multiple group analysis, see Chapter 13 for a description of the Mplus defaults for multiple group analysis and examples of specifying equalities in multiple group analsyis.
Hi Linda, I've read chapter 13 a few times and I only ask about specifying equalities because my model would be both multilevel and multigroup. So, when I say TYPE = MEANSTRUCTURE, should I add MULTILEVEL RANDOM.
This problem relates to a much larger issue (at least for me). I love your program, but I find your manual somewhat confusing. It seems to me that instead of providing an underlying logic which would allow users to infer the program's functioning and language, you have opted to try and give a plethora of modeling options with code provided in each case. While this is really fantastic for getting started, I am having difficulty progressing beyond an initial learning stage and can't intuit how to create new and dynamic models, such as a multilevel random coefficient multigroup model (this is exactly the opposite case of LISREL, where getting started is a bear, but movement beyond this stage is pretty straight forward). Is there anything you suggest to help me understand Mplus with greater depth?
I suggest starting with a simple model such as one you are familiar with in LISREL and translating it into the Mplus language. Then move to more complicated models until you feel you master the language. Mplus can estimate so many more type of models than LISREL that it is difficult to go into detail about each one. Specifying equalities is the same throughout the program. If you send an input to email@example.com and state clearly which parameters you want equal, I will be happy to tell you how to do this. The fact that your model is multilevel and multiple group does not change how equalities are specified. If you use RANDOM also, I think you need to use the KNOWNCLASS option for multiple group.
Anonymous posted on Tuesday, July 26, 2005 - 8:37 pm
hi linda, i want to analyze data using stata 7.0, the data will be captured from the construction workers about the frequency and severity of accidents in their work place, what type of data will i be collecting, numerical or categorical, which tests will i use to test the hypotheses?
bmuthen posted on Tuesday, July 26, 2005 - 11:54 pm
Well, why not handle Stata questions too in Mplus Discussion ... If frequency is collected as counts you could treat the outcome as Poisson or zero-inflated Poisson distributed. Or, if you find it sufficient to work with frequency ranges such as 0, 1 or 2, 3-5, more than 5, then you could treat the outcome as categorical (ordered polytomous). Severity could be either continuous or categorical (ordered polytomous). I assume you have some sort of covariate such as age or number of years on the job and want to regress these outcomes on the covariate, in which case testing if the regression slope is zero is straightforward. A joint analysis of frequency and severity as outcomes is a little more complex given their interrelatedness, but that doesn't sound necessary. All this said without knowing what you research questions are. And by the way, these analyses can be handled in Mplus.
Anonymous posted on Wednesday, July 27, 2005 - 8:08 am
hi linda, thank you very much for your response, anyway what is mplus?