Message/Author 


Dear Dr. Muthen, We have a longitudinal database that originally started with 252 first grade children at risk for reading failure that were randomly assigned to 3 conditions (group1  immediate small group reading intervention during the fall semester, group 2 monitoring children’s progress in regular classroom instruction during the fall semester and provide those with inadequate response to regular classroom instruction with small group reading intervention during the spring semester, and group 3  no treatment control). We have 18 weeks of progress monitoring data (using a word identification fluency measure that has very good psychometric properties; see Fuchs, Fuchs, & Compton, 2004) on all children during first grade (9 points fall semester and 9 points spring semester) along with an extensive pretest battery. We now have complete data on 180 children at the end of third grade with outcome measures on standardized measures (e.g., Woodcock Reading Mastery Test). Our interest is in evaluating whether a response to intervention approach can be used to identify groups of children with reading disability at the end of third grade. In other words, are there subpopulations of at risk children that do not respond to generally effective reading interventions during first grade and can our progress monitoring system adequately identify these children. So I have developed a model using discrete time survival mixture analysis with survival predicted by growth trajectory classes. I created a factor score at the end of third grade using a standard score cutoff of 85 on Woodcook word id, word attack, and passage comprehension (u1u3). This factor is basically whether you survive in reading. We then have growth data using cbm1cbm18 (y1y3) to monitor children during intervention in first grade. I have four known groups: group1 immediate intervention (same as group 1 above), group 2 spring intervention, group 4 no spring intervention, and group 5 controls (same as group 3above). Groups 2 & 4 make up the original group2 listed above with children that we monitored in regular classroom instruction during the fall semester. Group 2 were children that did not respond to regular instruction and thus were provided with small group reading intervention during the spring semester, and group 4 did respond and did not receive intervention. I also have 3 covariates from the pretest battery in the fall of first grade: VOC, SM, & RDN. I am asking for two classes (reading disabled and nonreading disabled) to be identified in each of the 4 known groups. The model runs well and results make very good sense. But before I get too excited I would like to know whether my code is correct for including known groups and covariates in the discrete time survival mixture analysis with survival predicted by growth trajectory classes model. Thanks, Don Compton TITLE: NRCLD Growth DATA: FILE IS C:\Documents and Settings\Don Compton\Desktop\Mplus\NRCLDSMC.dat; FORMAT IS 35F8.2; VARIABLE: NAMES ARE ID GP Y1Y18 U1U3 MFS OVS AMS WVS LCS OV RLN SM WV LC RDN; USEVARIABLES ARE Y1Y18 U1U3 SM WV RDN; CLASSES = cg (4) c(2); KNOWNCLASS = cg (GP = 1 GP = 2 GP = 4 GP = 5); CATEGORICAL = U1U3; CENTERING = GRANDMEAN (SM WV RDN); MISSING IS BLANK; AUXILIARY = ID; ANALYSIS: TYPE=MIXTURE MISSING; MODEL: %OVERALL% i s  Y1@0 Y2@1 Y3@2 Y4@3 Y5@4 Y6@5 Y7@6 Y8@7 Y9@8 Y10@9 Y11@10 Y12@11 Y13@12 Y14@13 Y15@14 Y16@15 Y17@16 Y18@17; i s ON SM WV RDN; f BY U1U3@1; c#1 ON cg#1 SM WV RDN; c#1 ON cg#2 SM WV RDN; c#1 ON cg#3 SM WV RDN; SAVEDATA: FILE IS C:\Documents and Settings\Don Compton\Desktop\Mplus\new.dat; RESULTS ARE C:\Documents and Settings\Don Compton\Desktop\Mplus\ results.dat; ESTIMATES ARE C:\Documents and Settings\Don Compton\Desktop\Mplus\ estimates.dat; SAVE=FSCORES; SAVE=CPROBABILITIES; OUTPUT: STANDARDIZED; TECH1 TECH4; PLOT: TYPE IS PLOT1 PLOT2 PLOT3; SERIES IS Y1 (0) Y2 (1) Y3 (2) Y4 (3) Y5 (4) Y6 (5) Y7 (6) Y8 (7) Y9 (8) Y10 (9) Y11 (10) Y12 (11) Y13 (12) Y14 (13) Y15 (14) Y16 (15) Y17 (16) Y18 (17); Fuchs, L. S., Fuchs, D., & Compton, D. L. (2004). Monitoring early reading development in first grade: Word identification fluency versus nonsense word fluency. Exceptional Children, 71, 721. 

bmuthen posted on Tuesday, February 28, 2006  9:54 pm



Interesting study with very rich data. I wonder if you have pretest data not only on the pretest battery you mention, but also on the repeated measures on the word identification fluency outcome that you study change with. That would help identify classes without influence of the interventions. I don't see this as a discretetime survival analysis because you don't seem to be studying time to an event. Rather, you are forming a factor based on 3 binary indicators of reading failure in grade 3 predicted by growth classes  which is fine. Do you find that 2 classes for "c" is the best choice? The data may point to more classes. Both the repeated measures and the factor indicators contribute to the classification so there's plenty of information to form classes. Note also that in addition, the pretest battery might be used to form pretest classes. 

Jinseok Kim posted on Wednesday, June 04, 2008  4:27 am



Hi, I am trying to compute a mixture model with two time to an event variables (y1, t1, and y2, t2 ) and two categorical variables (y3, y4). For example, let's say y1 and t1 being first time smoking and y2 and t2 being first time alcohol drinking and y3 and y4 being whether they were dropped out of school and involved in gang activities, respectively, and I am trying to create a latent class of delinquent behaviors using these indicators. Can you guide me where I have to start to estimate something like this model using mplus? Thanks. 


See Examples 8.15 and 8.16. See also Asparahouv, Masyn, and Muthen (2006) which is on the website. 

Jinseok Kim posted on Wednesday, June 04, 2008  6:37 pm



Thanks Linda for your prompt feedback. I looked at both examples and the paper. Now, I wonder if there is any way that I can incorporate two different time to an event variables as I described above in the same mixture model. Any suggestions? Thanks. 


Just put the two processes in the same MODEL command. 


Hi, I'm trying to run two discretetime survival models, one fixing the residual variance of F to zero, the other freeing it (as in a frailty model). My program for the frailty model works (estimation terminates normally and estimates make sense), but when I add the line "F@0" to the program I get the following error: THERE IS NOT ENOUGH MEMORY SPACE TO RUN THE PROGRAM ON THE CURRENT INPUT FILE. THE ANALYSIS REQUIRES 4 DIMENSIONS OF INTEGRATION RESULTING IN A TOTAL OF 0.50625E+05 INTEGRATION POINTS. ... Do you know why fixing the variance results in an integration problem of this size when leaving it free does not? Snippets of my input program are below for the nonfrailty version of the model is below (not enough space to paste in whole program). Thanks, Shawn Bauldry ... CLASSES = c(1); ANALYSIS: TYPE = MIXTURE; COVERAGE = 0.00; ALGORITHM = INTEGRATION; MODEL: %OVERALL% F BY sex15sex25@1; F@0; RelAtt BY rh1re3 rh1re7; RelBel BY rh1re4 rh1re6; PosSex BY rh1mo1 rh1mo7 rh1mo8; NegSex BY h1mo2 h1mo3 h1mo4; rh1mo1 WITH h1mo2; RelAtt WITH RelBel PosSex NegSex; RelBel WITH PosSex NegSex; PosSex WITH NegSex; F ON w1age female hisp black asian other medu RelAtt RelBel PosSex NegSex; 


Hard to say. Please send your input, output, data and license number to support@statmodel.com. 


Dear Dr Muthen, I am reading your paper on survival analysis 'Muthén, B. & Masyn, K. (2001). Discretetime survival mixture analysis. (#92)'and I would like to ask few questions. 1) In the Recidivism analysis: Singleclass model using all covariates, proportionality, and constant hazard, why fixing Thresholds to 0? 2) linear trend on the logit baseline hazard I would like to know how to write the input if I want to test linear trend on the logit baseline hazard for Recidivism analysis, for example. 3) Logistic regression with timevariant and invariant covariates In the line of the model proposed by Singer and Willet, is this model correct? Usevariables are u13 finaidempb13; Categorical = u13; Classes = c(1); Analysis: Model: %overall% u13 on finaidempb13; %c#1% u13 on finaideduc; u13 on empb1 (1); u13 on empb2 (1); u13 on empb3 (1); u13 on empb4 (1); u13 on empb5 (1); u13 on empb6 (1); u13 on empb7 (1); u13 on empb8 (1); u13 on empb9 (1); u13 on empb10 (1); u13 on empb11 (1); u13 on empb12 (1); u13 on empb13 (1); Thank you very much in advance. 


1) I am not quite sure which part of the paper or an Mplus output you refer to, but a single intercept/threshold/level parameter is needed for this and if you choose to put it into an intercept, then a threshold is not also needed. 2) A linear trend can for example be done via Model constraint. So if beta1betaT are labels given for the intercepts in the MODEL command, you write Model constraint: new(int slope); beta1 = int; beta2 = int+slope; beta3 = int+2*slope; etc 3) Yes that looks correct. 


I'm working on a research project examining estimating discretetime survival models as SEMs. One of the models I estimate includes three latent variables as covariates and I do not allow a random effect in the hazard component of the model. I estimate this model using a ML estimator and it requires numerical integration over three dimensions. In order to understand what is happening with this model (and others like it) it would be helpful to see the likelihood function that is being maximized. I looked through the User's Guide, Muthen and Masyn (2005), and the discussion boards, but I did not see any examples of hazard models that include latent covariates. Would you be able to recommend a paper or some other source, published or unpublished, that derives the likelihood function for such models? 


Please see the Asparouhov, Masyn, Muthen (2006) paper on our web site under Papers, Survival Analysis. This also gives references to work by Larsen in Biometrics. 

Yan Wang posted on Friday, July 31, 2009  1:14 pm



Dear Linda: I have a followup question on the model combining Latent Class Analysis (LCA) and Discretetime Survival Analysis (DTSA). In this model, I tried to use the class membership to predict the hazard. Now the model works, but how can I assess whether the hazard function differs across the two classes? Should I assess whether the mean of f is equal to 0 in the first class (BTW:the mean of f is fixed to be 0 in the second class)? The p value for f in class one is 0.824. Does that mean the classmembership does not affect the hazard probability? Thank you very much in advance! The input is as follows: ***************************************** usevariables are u1u7 y1y5; Missing are all (9999) ; categorical are u1u7; cluster=clr; classes = c(2); ANALYSIS: Type = Complex Mixture Missing; ALGORITHM=INTEGRATION; model: %overall% f by u1u7@1; f @0; %c#1% [y1y5]; %c#2% [y1y5]; Output: Tech1; Tech8; 


Yes, the nonsignificance of the mean of f means that class membership does not affect the hazard probability. 

Nan Zhang posted on Monday, April 26, 2010  8:21 pm



Dear Dr. Muthen This morning I posted a question about continuoustime survival analysis using a parametric proportional hazards model with a factor influencing survival(example 6.23). I did remove observed variable x1 from the categorical list as you suggested and the model works now. But I have another question in 6.23, covariate x is supposed to be a continuous variable. What about my covariates are ordinal or multinomial variable? How can I specify it since I also can not include it in the categorical list? Thanks so much, Best, Nan 


In regression, covariates can be binary or continuous. In both cases, they are treated as continuous. If you have an unordered categorical covariate, you need to create a set of dummy variables. 

Nan Zhang posted on Monday, April 26, 2010  9:44 pm



Thanks for the quick response! How about the covariate is ordered categorical variable? Best, Nan 


You can leave it and treat it as continuous or create a set of dummy variables. 

Nan Zhang posted on Monday, April 26, 2010  11:31 pm



Got you! Thanks so much. Another question related to survival analysis: Can Mplus handle left truncated data? Best, Nan 


No. 


Hello, I have a question regarding the comparison of regression coefficients and variance components estimated by Mplus in twolevel discretetime survival models that are parametrized as logistic regressions: suppose I wish to compare regression coefficients and variance components from the null model (where only time dummies are included) with coefficients and variance components of a model where some extra covariates have been introduced aside the time dummies. Because the variance of the latent variable is rescaled when estimating the second model, coefficients and variance components from the second model cannot be directly compared to those of the first model. My question is: is there a way in Mplus to fix the scale of the latent variable (for instance to what it is in the null model) in order to be able to do a proper comparison of the regression coefficients and variance components in nested models? Thank you very much in advance for any advice! 


Please send the two outputs you want to compare and your license number to support@statodel.com. 

Back to top 