I ran an LCGA model on drinking data that we collected on a weekly basis for 26 consecutive weeks. I'm using an ordered categorical dv (with eight categories consisting on the number of days in a week in which participants had 5+ drinks). I used the categorical option and also tried some models using the count option for poisson models. One model I have tried has only intercepts and slopes with two known classes (gender) by three unknown classes. All time points are fixed and are at equal intervals. When I look at the plots or when I plot the graphs myself using the graph data, I notice that the estimated values for the linear trends do not fit perfectly on a line (but they are close). For example, the estimated probabilities for a category may go from .455, .451, .447 but then .444 (i.e. deviations from a perfect line by about .001 for some of the time points)... with minor fluctuations from a perfect line in either direction. Is this simply due to rounding in the calculation of the probability estimates? (2) I get a message that the chisquare test cannot be computed because the frequency table for the latent class indicator model part is too large. Is this because I have a large number of time points? Thanks you. I wanted to mention that I found the course online lectures very helpful. Thank you for that resource.
Thank you for your response. I did try some GMM models and will pursue these further. I ran a GMM model with ordered categorical variables and encountered "LIKELIHOOD VALUE WAS NOT REPLICATED..." I increased the number of starts to 200 5 but still didn't get a proper solution. Will it help if I increase the number of starts further?
Is there a specific way to determine whether the GMM model is better than the LCGA model? I assume that there are no fit indices that can be compared across these models? I'm guessing that I should go with the GMM model if the variances of the latent growth intercepts or slopes are in fact significant? Thank you.
It is sometimes necessary to increase the random starts further than you have. I would try that. It is sometimes the case that the intercept growth factor has a significant variance while the slope growth factor does not in which case the slope factor variance can be set to zero.
To judge which model fits best, see which gets the best BIC and Loglikelihood. See the following paper which also looks at the residuals in TECH10 to determine model fit:
Kreuter, F. & Muthen, B. (2007). Analyzing criminal trajectory profiles: Bridging multilevel and group-based approaches using growth mixture modeling. Conditionally accepted for publication in Journal of Quantitative Criminology.
This paper is available on the website under Papers.
Howie Lim posted on Monday, October 20, 2008 - 2:33 pm
A question about time points(age) in using MPLUS
Dear Drs. Muthen,
I'm doing an analysis of trajectory of risk behavior among older men in a study of men who are HIV-positive and HIV-negative men. This is a cohort study where by participants have their clinical visit at every six month and their risk behaviors are assessed and data collected. I plan to analyze the trajectory of sexual history of the older men aged 50 at the latest visit. This will include data in the past 10 years (or 20 visits). The age of participants during the past 10 years range from 40 to 65. Each visit includes participants of different ages.
To examine the aging effect on sexual behavior, I've restructed the data so that everyone starts at age 40. The risky sex behavior was coded as a categorical (ordinal) data range from 1(low risk) to 5(high risk). The total number of time point (age) is actually 50 because participants come in every 6 month, and the age starts from 40, 40.5, 41, 41.5, 42, 42.5..until 65. I plan to run both LCGA and GMM to determine if there exist distict groups of trajectory of risk behavior. Can MPlus run the analysis with 50 time points? Do you think this is an appropriate analysis?
Thank you so much of your time. I find the discussion board very useful