Message/Author 

Lois Downey posted on Thursday, June 18, 2009  11:03 am



Before the release of Mplus 5.21, I ran a survival analysis that produced a loglikelihood of 292.574 and a parameter estimate of .638 (p=.007) for my predictor of interest. With Mplus 5.21, the same model generates a loglikelihood of +6346.678 and a parameter estimate of .102 (p=.173) for my predictor of interest. Most other models that I have compared from before and after the Mplus 5.21 release look identical. The substantial change in this model may reflect the fact that I have a relatively small sample (n=488) and a reasonably large number of free parameters (17). Do you find the difference in results believable, and would you advocate accepting the Mplus 5.21 result? 

Lois Downey posted on Thursday, June 18, 2009  2:45 pm



As an amendment to my earlier posting, I would note that at least part of the problem may be the result of the fact that I have observations with survival time of 0. Do the two versions of 5.21 handle survival times of 0 differently? 


Please send the two outputs, your data, and your license number to support@statmodel.com. 

Michael posted on Wednesday, March 07, 2012  12:52 pm



I am attempting to use a discretetime survival model to examine initiation of substance use. I have a single, dichotomous outcome variable (substance use initiation), assessed at 5 time points. I would like to include both timeinvariant (e.g., gender) and timevarying (e.g., stressful experiences) covariates, also assessed at the 5 time points. The central question I am interested in is whether recent stressful experiences represent a risk factor for substance use initiation. Following are the VARIABLE and MODEL statements I am working with. VARIABLES ARE SUW1SUW5 !Substance use initiation, coded 0 (no), 1 (yes), and 999 (missing or yes at a previous wave) Gender !coded 0 = female, 1 = male STW1STW4 !Past year stressful experiences (continuous) MODEL: f by SUW1@1 SUW2@2 SUW3@3 SUW4@4 SUW5@5; f@0; f on Gender; SUW2 on STW1 (A); SUW3 on STW2 (A); SUW4 on STW3 (A); SUW5 on STW4 (A); Have I included the timevarying covariates correctly? I constrained them to be equal across the waves, as it does not seem necessary to unconstrain them for what I am doing. Your feedback is greatly appreciated. 


It looks like you are using the timevarying covariate correctly but I wonder about your BY statement. It should be f by SUW1SUW5@1; See Example 6.20 in the user's guide. 

Michael posted on Thursday, March 08, 2012  3:44 pm



Thank you for your response. The @15 was an error from reducing the statement to post here. I was thinking survival, but it seems that my fingers liked growth better. Thanks again. 


Hello, I may have a situation that is similar to Michael's. I have 5 time points of a discrete outcome, and 2 time varying covariate that corresponds to each of the 5 outcome time points. Here are my questions. 1) Is it possible to have interaction terms between the time varying covariates? e.g. OutcomeT2 on CovariateA_T1 CovariateB_T1 CovariateA_T1*CovariateB_T1; 2) There are theoretical reasons to believe that each time varying covariate should be correlated with itself at the next time point (CovariateA_T1 CovariateA_T2 etc). How would that be written into the survival analysis syntax? What I would really like to try to do is have a path analysis that ends in a survival function, with the caveat that the path model varies over time. Can that be accomplished in mplus? Or would such a scenario be better modeled in mplus but using a different framework like SEM or a multilevel model with a within subjects time factor for the first level? 


1) sure 2) All covariates are correlated  but this marginal part of the model is not estimated in the model (can be estimated by sample statistics). Use wide format (singlelevel) analysis to get max flexibility  for instance allowing path coefficients varying over time. 

Back to top 