Sarah Racz posted on Tuesday, February 05, 2013 - 12:54 pm
Dear Drs. Muthen, I am conducting an observed (continuous covariate; e.g., SES) by latent variable (intercept and slope) interaction in a latent growth curve model using the XWITH command. When running the analyses I get the following warning:
"WARNING: THE MODEL ESTIMATION HAS REACHED A SADDLE POINT OR A POINT WHERE THE OBSERVED AND THE EXPECTED INFORMATION MATRICES DO NOT MATCH. AN ADJUSTMENT TO THE ESTIMATION OF THE INFORMATION MATRIX HAS BEEN MADE. THE CONDITION NUMBER IS -0.419D-02. THE PROBLEM MAY ALSO BE RESOLVED BY DECREASING THE VALUE OF THE MCONVERGENCE OR LOGCRITERION OPTIONS OR BY CHANGING THE STARTING VALUES OR BY USING THE MLF ESTIMATOR."
I have decreased the mconvergence and logcriterion options and increased the starting values, which did not resolve the warning. I also have weights in my data, which means I cannot use the MLF estimator (however, the warning does go away if I remove the weights and use the MLF estimator). Even with this saddle point warning, I receive output which is interpretable and includes standard errors (although not standardized results). Does this mean I can ignore the error? Or do I need to try something else?
Thank you so much for your assistance! -Sarah Racz
Sarah Racz posted on Tuesday, February 05, 2013 - 2:19 pm
Great - thank you Bengt for the quick reply!
Meike Slagt posted on Monday, December 15, 2014 - 12:25 pm
I’m estimating a parallel process model, with the intercept of the second process centered on the final time point. I’m testing whether the interaction between an exogenous variable (x) and the slope and intercept of the first process (i1 and s1) predict the intercept and slope of the second process (using XWITH).
In my output I get a warning, stating that the model has reached a saddle point. However, model estimation terminated normally, and I understand from earlier posts that I could just go ahead and use this output.
However, I did try and get rid of these warnings, and I found two ways to do so: 1.) using an MLF estimator and 2.) using algorithm = integration. I understand the 1st option is not recommendable, because my sample size is rather small (N=245), and it will overestimate SE’s. What about the 2nd option? The SE’s get even bigger than when I use an MLF estimator, and as a result, latent interactions that were significant are no longer significant.
Do you recommend using numerical integration (=huge SE’s, even bigger than in MLF) or using the original output (=with the warning that an adjustment to the estimation of the information matrix has been made, see article on saddle points)?