Message/Author 

Sophie posted on Tuesday, March 22, 2016  3:57 am



I specified a latent growth curve model to test the effectiveness of an intervention. As the intervention was provided to mothers, and I am analyzing data at the child level (91 mothers with 133 children; ICC ranged between 3788%), I need to control for the multilevel structure (I am not interested in the difference between/within change). What variable do I need for this and how can I specify this in my lgcm? Thanks in advance! 


Your cluster variable is mother id. Using that, you can either do a Type=Complex analysis which keeps your model the same as what you have or you can do a Type=Twolevel analysis in line with UG ex 9.12. 


Dear Drs. Muthen, I ran a twolevel (classes, students) latent growth curve model(free slope). My goal is to look at crosslevel and betweenlevelinteraction effects of the variable “cos”. Input: VARIABLE: ……… between = cos; cluster = class; ANALYSIS: type = TWOLEVEL Random; ALGORITHM=INTEGRATION MODEL: %within% iw sw  t1@0 t2 t3 t4 t5 t6 t7 t8@1; t1  t8 (1); sw@0 ! because of negative (nonsignificant) residual variances of the within slope b  sw ON iw %between% ib sb  t1@0 t2 t3 t4 t5 t6 t7 t8@1; t1_tv  t8_tv@0; ibxcos  IB XWITH cos; b sb ON cos; sb ON ibxcos; Questions: Does it make sense in this model that I… 1. …constrained residual variances to be equal over time on the within level? 2. …fixed the residual variances @0 on the between level? 3. How are the slope between (sb) and the intercept between (ib) to be interpreted? Are their factor scores affected by all points of measurement (as in linear growth modeling)? Or are they affected by the first (and the last) point(s) of measurement only? Thank you so much! 


12. Yes 3. ib and sb are the betweenlevel parts of the growth factors. Factor scores are affected by all points. 


Dear Dr. Muthen, thank you so much for your quick answer. My goal is to use the between level slope growth factor (sb) as a dependent variable. It is supposed to represent reading skill competence growth of school classes. Do you think that the freely estimated slope model (as I used it) is an appropriate way to represent skill growth? Or would it make more sense to use a linear or quadratic model for my purpose? If so, could you name some reasons for this? Model fit (when I modeled the data in a nonhierarchical way/type=complex) was best for the freely estimated slope model, it was not as good for the linear slope model and even worse for the linear slope model. Thank you so much again! Frank 


First, note that you should hold free time scores equal across the two levels; otherwise the mean structure and the covariance structure are different and not in line with growth modeling. I would try to use linear or quadratic growth because of easier interpretations. 

Frank Egloff posted on Tuesday, February 14, 2017  10:09 am



Dear Dr. Muthen, thank you so much! Now the question is between linear and quadratic modeling, but I have difficulties with testing multilevel quadratic lgms (yet without covariates): Ignoring the multilevel structure of my data to model in a single level way is no problem: The quadratic factor mean is sign. and fit compared to a linear model is improved. Still: Accounting for the multilevel structure I get error messages (each concerning different kinds of measures to be modeled): 1. NONZERO DERIVATIVE OF THE OBSERVEDDATA LOGLIKELIHOOD 2. NONPOSITIVE DEFINITE FISHER INFORMATION MATRIX %within% iw sw qw t1@0 t2@1 t3@2 t4@3 t5@4 t6@5 t7@6 t8@7; t1  t8 (1); %between% ib sb qb t1@0 t2@1 t3@2 t4@3 t5@4 t6@5 t7@6 t8@7; t1  t8@0; 1. Is sth. wrong with my input? 2. What may have caused the errors? 3. Do I have to model in the same way on both levels (quadratic on both levels)? 


Send your output to Support along with your license number. 


Dear Dr. Muthen, If I have independent observations across three years in 3040 different countries, and I wished to examine change over time on a macrolevel. Could it be appropriate to nest individual cases within years, nest years within countries, and then run a latent growth curve on the higherlevel country betweenlevel, by using clustermeans for each year of observations? Or does the independent observations on the withinlevel make this type of modelling impossible or invalid? Additionally, would I have to specify to Mplus I am interested in clustermeans, or will that be run automatically. Warm regards, Alex 


If the same subjects are measured at the 3 time points, you can let the outcome for the 3 years be 3 columns in your data so that you do growth modeling in wide form. If you want country to be treated as a random mode of variation rather than a fixed mode  where fixed leads to multiplegroup modeling  then country would represent level 2 in a twolevel analysis. See UG ex 9.14. But if you have different subjects at the different time points, growth modeling is not possible. 

Back to top 