

Growth curve multilevel modeling 

Message/Author 


Hey everyone, At the moment I try to create a growth curve for anxiety over seven sequenced measure points. I would like to determine the slope factor as a linear function of the passed time. Besides the Interceptparameter shall be explained through a few personspecific covariates (GE,SEB,ASK,FSSK,ERW,EXT,WI). Overall it shall become a multilevel analysis with longitudinal data. Since now I have written the following command: title: Computing the growth curve data: file = DataforLWK.dat; variable: names = PA1,PA2,PA3,PA4,PA5,PA6,PA7,GE,SEB,ASK,FSSK,ERW,EXT,WI; missing = all (99); model: interc linear  PA1@0 PA2@1 PA3@2 PA4@3 PA5@4 PA6@5 PA7@6; interc on GE,SEB,ASK,FSSK,ERW,EXT,WI; output: sampstat standardized stdyx; plot: type = plot3; series = PA1 (linear) PA2 (linear) PA3 (linear) PA4 (linear) PA5 (linear) PA6 (linear); Is this command correct? Rolph 


I’m confused by one thing: When I run the command and display the graphic "Sample & estimated means", the estimated means show a horizontal curve. But I thought that I made clear that the curve should gain linearly ("interc linear  PA1@0 PA2@1 PA3@2 PA4@3 PA5@4 PA6@5 PA7@6"). Can someone help me with that? Many thanks! 


What is the mean of linear? See TECH4 to find it. 


The estimated mean of linear is 0.003. What shall I do now? 


This value is the average slope of the growth curve. It is zero or flat  a horizontal line. Look at your observed means. They probably show this same trend. 


But my theory says that the values should rise linearly. Can't I estimate a model with, let's say, has a growth of 1 instead of zero? 


Your data does not seem to agree with your theory. The results show what is found in the data. 


Dear Dr Muthens, I have a question relating to the difference between MLM and the Structural Equation Modeling approach to LGCM in MPLUS. From my understanding, the MLM differs from LGCM because data can be inputted as real time scores rather than estimated time parameters, as it is dealt with in LGCM. Is it therefore correct that MLM can produce different results for models, which have varying times of observations, and is often preferred for this type of modeling? I am also aware the MPLUS uses a multivariate approach (treating the model as 1level  as it would be treated in LGCM), does this mean that both the MLM (type = two level random) and LGCM (type = random) produce the same results in MPLUS because they are both essentially onelevel models? I have been running a quadratic model with varying times of observations using Type = Random (LGCM) in MPLUS. If MPLUS produce different results for the MLM and LGCM methods of modeling, I am wondering if I need to also run this model using a 2level MLM to see if I produce the same results (therefore, justifying my choice of modeling)? If so, is it possible to refer me to an example of the required input for a MLM in MPLUS? Unfortunately I can't seem to find anything similar. I'm sorry for asking multiple questions at once. Thank you very much for your help! 


You will find some answers in the UG ex 9.16. The UG ex 6.12 Type=Random approach should give the same answers as with the ex 9.16 approach. If this doesn't help, send input, output, data, and license number to Support. 


Thanks very much for your quick response Dr Muthens! Just to clarify  is my understanding that MPLUS treats both MLM (Type = two level random) and LGCM (Type = random) as onelevel models (using a general latent growth framework) correct? Therefore, does it also always treat time as estimated parameters rather than real time scores? This would reason that both models produce the same results in MPLUS. Please let me know if my thinking here is incorrect. Thank you in advance for your reply. 


Type=Random using AT to handle individuallyvarying times of observations as in ex 6.12 uses a wide, singlelevel approach where the time scores are variables, not parameters that can be estimated. You can tell from the output that it is a wide approach in that you get a residual variance for the outcome at each time point. In contrast, Type=Twolevel as in ex 9.16 uses a long, twolevel approach. The time scores are variables here also. The two approaches give the same answer if (1) the time scores are the same and (2) is the residual variances are held equal in the wide approach. The time scores are parameters, not variables, only in the case where you don't use AT and take the singlelevel, wide approach that is the standard Mplus approach to growth when individuals have the same times of observation. 

Back to top 

