Terri Scott posted on Tuesday, September 03, 2013 - 6:04 pm
I'm testing if the 12 continuous variables in my model are moderated by 2 categorial (0, 1) variables and/or the interaction term (racexsex). Classes = c(3); Analysis: Type = Mixture; Parameterization=loglinear; Starts = 40, 2; Processors = 2; Define: racexsex=race*sex; Model: %overall% c on sex race racexsex; Output: samp Stand Tech11; Plot: Type = Plot3; series = PDS_tot (1) PCL_tot (2) YLS_off (3) MCAA_RPT (4) PTSD(5) YSR_tot (6) ARSQ_fea(7) GSES_tot (8) YLS_educ(9) DAST_27(10) YFAD_GF (11) SAI_Free (12);
Summary shows only half the observations are used (155 out of 326). Am I misspecifying something thatprevents all cases from being analyzed?
SUMMARY OF ANALYSIS Number of groups 1 Number of observations 155 Number of dependent variables 12 Number of independent variables 3 Number of continuous latent variables 0 Number of categorical latent variables 1
Milan R posted on Saturday, February 24, 2018 - 5:57 am
Hello, I am analyzing a 3-trajectory GMM with a covariate (M). There is a trajectory with a significant intercept but non-significant slope (despite its magnitude is negative). Although the mean trend of this trajectory is not significantly different from zero, this covariate has a significant and negative effect on the slope. How can I probe the effect of this covariate on individual slopes within this trajectory using Mplus? I'd like to probe simple slopes treating the covariate as a moderator, time as X, my outcomes as Y. Thank you!
If the covariate M has a significant effect on s and this regression has a >0 residual variance, I think the model is ok.
Milan R posted on Saturday, February 24, 2018 - 3:28 pm
Thank you, Dr. Muthen! How can I write out syntax in Mplus to probe the effect of this covariate on individual slopes within this trajectory? I'd like to probe simple slopes treating the covariate as a moderator, time as X, outcome as Y. Thank you!
You can do that in a twolevel run with time as level 1 and subject as level 2. We have a UG example on doing growth in a twolevel fashion. The latent class variable needs to be a Between= variable. Then add a Time*Moderator interaction.
Milan R posted on Sunday, February 25, 2018 - 1:38 pm
Dr. Muthen, I followed the UG example 10.2 to write up the model. However, two problems showed up that I could not figure out why and how to fix. After reshaping my data into long format, my model is as follows:
Define: int=time*moderator; Variable: Names are moderator sex Y time id; Usevariables are moderator sex Y time int; Classes = c(3); Cluster = id; Within=time; Between= sex moderator c; Analysis:type= twolevel mixture random; Model: %within% %overall% s1| y ON time; %between% %overall% c y ON sex moderator int; %c#1% [s1]; %c#2% [s1]; %c#3% [s1]; %c#4% [s1];
Two problems: 1. the proportions of membership in each class are different from when I used type= mixture (not multilevel approach). 2. The effects of int (time*moderator) were fixed in all three class. Thank you!
will be a within-level variable because time is within.
Regarding different class percentages, first make sure that you have the same number of parameters in a model where you don't have the moderator. For instance, the single-level model has free residual variances whereas the twolevel model holds them equal - so hold them equal in the single-level model too.
Milan R posted on Monday, February 26, 2018 - 11:44 am
Dr. Muthen, Once I switched "int" to be a within-level variable, I could not use it on the between-level. My goal is to test the effect of the moderator (which is a subject/between level variable) on s1 (individual slopes). But since s1 is between-level, I don't know how to write out the effect of int on s1 and where to place this statement to achieve my goal. Thanks again!
on the Between level. This way time and moderator form a cross-level interaction.
Milan R posted on Wednesday, February 28, 2018 - 6:40 am
Thank you, Dr. Muthen! You mentioned earlier that "Regarding different class percentages, first make sure that you have the same number of parameters in a model where you don't have the moderator. " So, to solve the issue of different class memberships in the multilevel context than in the single-level one, I freed up several parameters as follows. But the results were still not consistent with those under the single-level model. Can you help and point out if I missed to free up any other parameters? Thank you so much! ----- CLASSES = c(3); CLUSTER = id; WITHIN=time; BETWEEN= sex moderator c; %WITHIN% %OVERALL% s1| y ON time; %BETWEEN% %OVERALL% c ON sex moderator; %c#1% s1 ON moderator; y; [s1*]; s1*; %c#2% s1 ON moderator; y; [s1*]; s1*; %c#3% s1 ON moderator; y; [s1*]; s1*;
You should focus on restricting the single-level model to match the twolevel. A key difference is that single-level allows residual variances to differ across time points whereas with two-level they are the same - so hold them equal in the single-level run.