

LGM with ordered categorical indicato... 

Message/Author 


Can you discuss the interpretation of the effect of covariates on ordered categorical indicator variables? In my case, there is a positive significant relationship between depression score and smoking status (a variable with 5 categories from nonsmoker to frequent smoker). 

bmuthen posted on Saturday, October 19, 2002  3:37 pm



See the two log odds given in (29) and (30) on page 342 of the Mplus User's Guide Appendix 1. For a unit change in x (Depression), the beta slope tells you how much these log odds change (increase or decrease). See also the Agresti reference given there. 

Anonymous posted on Sunday, June 01, 2003  4:45 pm



I would like to describe how suicidality changes over time in a group of people and analyze whether individuals move in and out of a binary category (having suicidal ideation) over 4 measurement points (3 posttreatment), or whether people simply stop being suicidal at some point and stay there. Also, whether being suicidal to begin with predicts changes in ideation over time. The following model runs, but I can't really figure out what it means. The sample means, for example, exceed 1, even though the values of the variable can only be 0 or 1. Is there some literature where I can find guidance on how to interpret the output (and figure out whether I did this correctly). USEVARIABLES ARE si si03 si09 si15 ; CATEGORICAL ARE si si03 si09 si15; MISSING ARE * ; ANALYSIS: TYPE = MEANSTRUCTURE; MODEL: SUIint BY sisi15@1; SUIslope BY si@0 si03@3 si09@9 si15@15; [si$1 si03$1 si09$1 si15$1] (1); !threshold [SUIint@0 SUIslope]; !fixes intercept {si@1 si03si15}; !scale factor SUIslope on SUIint; Thanks for your input! 

bmuthen posted on Monday, June 02, 2003  7:59 am



Growth modeling with binary outcomes describes individual differences in growth/decline of the probability of the outcome. It sounds like your example fits into this mold to some extent, although your questions in part seem to go beyond what this model will tell you. You have set up the model correctly. The sample means you refer to might simply be the threshold estimates which are z scores corresponding to the probability of suicide. I am not aware of any real applied writing on this type of modeling yet. A more technical account is given in Mplus Web Note #4 and also in the article below. Muthén, B. (1996). Growth modeling with binary responses. In A. V. Eye, & C. Clogg (Eds.), Categorical Variables in Developmental Research: Methods of Analysis (pp. 3754). San Diego, CA: Academic Press. (#64) You get estimates that tell you if the slope is positive or negative and if it influenced by the initial status. You can also compute etimated average probabilities from this output, although that requires a bit more knowledge. In addition, you can estimate individual growth factor scores, from which you can compute individual estimated probabilities. Some of this is covered in Day 3 of our Mplus short courses. If you have a specific (short) question, feel free to send me the output. 

Back to top 

