

Multilevel Modelling Questions and Is... 

Message/Author 


1) I want my primary outcome variable to be a latent factor. I am unsure where this code would go in my model, since it is neither "within" or "between" as this latent factor will be regressed on both level 2 and level 1 variables. 2) I want to add a mediator into my multilevel model. When I add in a mediation, I have to take out the random intercept. My model then runs, but I do not obtain any output. How would I be able to run a multilevel mediation and obtain output? 3) I would like to use standardized estimates, but I can only obtain stdyx once I remove the random intercept. Is there a way to get standardized output while also accounting for random intercepts? 4) Why does the multilevel model automatically run "WITH" statements even if not prompted? What would be the difference between a "WITH" between two variables at level 1 and a "WITH" between the same two variables at level 2? 5) I want to add in a variable as an aggregate at level 2 and a deviation score at level 1 (i.e. group mean centred). The model runs properly with deviation without aggregate variables and vice versa. When I add both, the model says "the log likelihood decreased in the last EM iteration." and "The model estimation did not terminate normally due to an error in the computation." I'm not sure how to approach this, as they run well separately. 


1. Examples 9.6, 9.7, and 9.8 show multilevel CFA. 2. Send the output and your license number to support@statmodel.com. 3. Send the output and your license number to support@statmodel.com. 4. There are defaults. If you don't like them, you can override them. 5. Send the output and your license number to support@statmodel.com. 


Hi Dr. Muthen, Thank you for the response. 1. I am hoping to use a latent factor as my primary outcome for an SEM where I have level 1 and level 2 predictors of the level 1 latent factor (substance use). I do not think those examples apply, as I am not trying to create the latent factor at two levels, but rather use upper level variables to predict the lower level latent factor. 4. How would I go about overriding these? If I do not override them, how do I interpret the WITH statements at each level? Currently, I have my substance use variables input as manifest variables and those are the ones that are being "WITH"'d by default. These manifest substance use variables (measured at level 1) are being regressed on upper and lower level variables Thanks in advance, Jillian 


A variable measured at the between level cannot be used in the within part of the model. A level 2 variable can predict only the between part of a level 1 variable. If the parameter is, for example, y1 WITH y2, say y1 WITH y2 @0. 


Hello Professors and Administrators I am performing a twolevelregression with 10 predictors on level 1 and different continuous outcomevariables (type=twolevel; N = 255). Besides I make dataimputation. My problem is, that one of the outcomevariables (children's attitudes towards kindergarten) is very skewed and has a strong ceilingeffect (skewness: 1.96, curtosis: 5.5). Some authors suggest that nonnormality is ok if skewness is < 2 and curtosis < 7, this means, this would be ok as long as I use robust estimators. My question to you: Do you find it appropriate zu use MLR with the scale described above (attitudes) or should I take the WLSMVestimator instead fot that dimension? Besides: The results using one or the other are different! Thanks so much for your help!!! Tamara 


With strong ceiling effect (say > 20%), I recommend using the Censored approach. MLR corrects only SEs, but your model needs to be different  hence Censored. 


Thank you very much Mr. Muthen! I confess, I have never heard about censoring, but I have found the command in the user's guide and it works (censored is VAR (a);). But what do you mean by strenghts of the ceiling effect >20%? Is there a limit, and how can I analyze it? I am actually not sure if I have a ceiling effect, i just named it like this, because of the relatively large skewness and curtosis (2.0, 5.5). Besides I realized that now Imputations don't work any more. Is there another method for missings on predictors? And last question: If the results (looking at the effects) are the same for CENSORING as with estimatore = WLSMV, can I instead use WLSMV? (the intercepts for the last are more intuitive and I can use Imputations). Thanks again! Tamara 


A ceiling effect of 20% means that you have 20% of your sample at the highest value of the outcome. Missing on predictors can be handled by including them in the model. Here is a background for this: In regression, the model is estimated conditioned on the covariates. Their means, variances, and covariances are not model parameters and no distributional assumptions are made about them. You can find their means, variances, and covariances in the descriptive statistics for the data set. Any observation that has a missing value on one or more observed exogenous covariates is eliminated from the analysis. To avoid this, the covariates can be brought into the model by mentioning their means, variances, and/or covariances in the MODEL command. This can be done for maximum likelihood estimation and Bayes. It should not be done for weighted least squares estimation. When this is done, the means, variances, and covariances of the covariates become model parameters and distributional assumptions are made about them. This is not innocuous, especially with binary covariates. Pros and cons of this approach are discussed in Chapter 10 of our RMA book. For your last question, I assume that you treat the variable as regular continuous outcomes, that is, don't mention them on any list such as Categorical or Censored. If you see no difference, there must not be much censoring. 


Thank you Prof. Muthén! I try to understand the implications of your answer. I'm very glad to have had your assistance! Tamara 

Back to top 

