Hello, I have a question. Recently you provided a very nice answer to my question about the effects of a covariate on my parallel process LGM. My question is about the practical interpretation of such effects. Namely, I found an effect of -.363 from my physical activity slope to my smoking slope. This effect was significant, p < .05. I understand that the finding means that for every unit change in physical activity, there is a decrease of .363 in the log odds of advancement to the next smoking category (smoking is an ordered polytomous variable). However, in terms of a practical interpretation for a journal, I am being asked if that can be translated into a 36% decrease in the odds of progressing to a higher category of smoking, if I exponentiate the log odds. In other words, to make this more practical, can I convert the log odds interpretations at the structural level, and covariate level, as I did in the measurement level?
bmuthen posted on Wednesday, October 30, 2002 - 4:14 pm
The odds corresponding to the log odds of .363 is 1.44 - how do you get a 36% change in the odds? Because odds are probability ratios, I am not sure percentage change is independent of which probability is the starting point for the change. You better consult the Hosmer-Lemeshow logistic regression book.
Julie Noble posted on Monday, October 01, 2007 - 7:29 am
As a neophyte to categorical SEM, I have some questions. I want to run a single group, single level SEM with a multinomial (non-ordered, categorical) outcome. Indicators and latent factors are both categorical and continuous. I couldn't find much information on SEM and multinomial outcomes in the manual; most examples refer to ordered categorical outcomes. Am I restricted to either a path model or, if I want a SEM, using a dichotomous outcome? Can you refer me to other documentation, reference code, etc.?
Continuous latent variables, factors, can have factor indicators that are continuous, censored, binary, ordered categorical (ordinal), counts, or combinations of these variable types. These latent variables can predict continuous, censored, binary, ordered categorical (ordinal), unordered categorical (nominal), counts, or combinations of these variable types. The only restriction is that nominal variables cannot be used as mediating variables.
Dependent variable scale type can be specified using the CENSORED, CATEGORICAL, COUNT, and NOMINAL options of the variable command.