

Start Values in Multilevel Path Analysis 

Message/Author 

Anonymous posted on Saturday, March 05, 2005  11:27 am



Hello, I am developing a path analysis model using Mplus 3.11. In the model, I have one exogenous measure, two mediating measures and one dependent measure. I have worked through all of the steps that come from Muthen (1994). The final step is where the problem is coming in. My dependent measure is creating a nonpositive definite matrix. I have checked and it is developed correctly. So, I think that I need a start value for this measure. I have tried to use the start value suggestions from pgs. 401 to 402 in the manual. However, I keep getting the don't trust the results message because of the nonpositive definite matrix problem. So, without throwing the entire computer out of the window :), do you have an suggestions? 


This message usually does not have anything to do with start values. It is usually caused by negative residual variances, correlations greater than one, or some linear dependency among 3 or more variables. Check your output for the first two. If you can't see where the problem is, you can send the output and data along with your license number to support@statmodel.com and we can take a look at it. 


Most times that I conduct multilevel path analysis (or full SEM), with at least two dependent variables, the model fails to converge because the program reports correlations between the DVs at 1.0. Sometimes I look at the correlation between the variables in an aggregated correlation matrix and it's very low (e.g., .2), but Mplus still won't run the model because it says the relation is at 1.0. I am assuming it's because the betweengroups model has variables which usually have very high correlations. Is this a common problem for multivariate MSEMs and, if so, what can one do about it? 

bmuthen posted on Thursday, August 25, 2005  2:37 pm



It happens with some frequency. The modelestimated between correlations can get high either because betweenlevel components are strongly related or because of a misspecified model. If you have a model that is particularly puzzling to you, don't hesitate to send it to support@statmodel.com to see if light can be shed on it. 

Back to top 

