Discrete-time survival model PreviousNext
Mplus Discussion > Latent Variable Mixture Modeling >
Message/Author
 Anonymous posted on Tuesday, April 12, 2005 - 6:33 am
Hello,

I am trying to estimate a discrete-time survival model. I have 21 time points out of which respondents can change status in the last 17 (the first 5 are locked). I have estimated the model without covariates and I now want to add covariates to the model. How do I add time varying covariates? Also did I set the starting values corectly?

Thanks!

Tobias

Here is the code so far:

DATA:
file is u:\initiationdata.dat;
VARIABLE:
NAMES ARE id gender u5 u6 u7 u8 u9 u10
u11 u12 u13 u14 u15 u16 u17 u18
u19 u20 u21;
usevariables = gender u5 u6 u7 u8 u9 u10
u11 u12 u13 u14 u15 u16 u17 u18
u19 u20 u21;
categorical = u5 u6 u7 u8 u9 u10
u11 u12 u13 u14 u15 u16 u17 u18
u19 u20 u21;
MISSING = ALL(999);
IDVAR is id;
CLASSES = c(1);
ANALYSIS:
Type = Mixture Missing;
!Starts = 100 10;
Model:
%overall%
u5-u21 on gender;
%c#1%
[u5$1*0 u6$1*0 u7$1*0 u8$1*0 u9$1*0 u10$1*0
u11$1*0 u12$1*0 u13$1*0 u14$1*0 u15$1*0
u16$1*0 u17$1*0 u18$1*0 u19$1*0 u20$1*0 u21$1*0];

OUTPUT:
sampstat Tech1 Tech8;
 Tobias posted on Wednesday, April 13, 2005 - 5:04 pm
I changed the starting values to the conditional probabilities of the event to occur in each period. The model runs fine, but I do have problems with the starting values when I am adding demographic variables such as gender and race. How can I determine the appropriate starting values?

Thanks!
 BMuthen posted on Wednesday, April 13, 2005 - 11:42 pm
If you are using Version 3.12, you should not need starting values at all. I would use STARTS = 50 5;

Time-varying covariates are added by saying ut ON xt; for each value of t.

If you continue to have problems, send you input/output and data along with your license number to support@statmodel.com.
 Maarten Hermans posted on Tuesday, June 15, 2010 - 7:56 am
I am trying to reproduce the discrete time survival analysis examples in Singer & Willet (2003), chapter 12 [0], using the approach described in Muthen & Masyn (2005).

1. I can reproduce the zero order polynomial (constant) and completely general model, but I am not sure how to specify the linear, quadratic, etc. models in Mplus. I am using the syntax of Model 5 ('linear logit baseline hazard model') in Masyn (2003), but I do not obtain meaningful values (the means of the latent variables are 0). See [1] for input.

2. S&W describe the use of the complementary log-log link. In a old reply I read that cloglog is not supported in Mplus, is this still the case?

3. In R one can drop the intercept using a -1 in a formula, eg. glm(event~u7 + u8 + u9 + u10 + u11 + u12 - 1, family=binomial(link = "logit")). Is there an equivalent shorthand in Mplus?

Thanks,

Maarten

[0] http://www.ats.ucla.edu/stat/examples/alda/
[1] http://pastebin.com/RV8hPURt
 Bengt O. Muthen posted on Tuesday, June 15, 2010 - 11:13 am
1. Try fixing the variances of eta0 and eta1:

eta0-eta1@0;

2. Yes.

3. You can fix a threshold @0.
 Maarten Hermans posted on Saturday, July 24, 2010 - 2:56 am
Thanks.

(1) Regarding the baseline hazard form: if I fix the variances of eta0 and eta1 I get the message that I need to add ALGORITHM=INTEGRATION. If I add that however, Mplus will not provide output. In the DOS-output I find the error: "forrt1: severe (164): Program Exception - integer divide by zero". The data an input are online[0], I can email the licence number on Monday if needed.

(2) I am confused regarding the use of TRAINING vs. KNOWNCLASS when doing multiple group comparisons in a discrete time survival analysis. The TRAINING-statement does work (KNOWNCLASS does not), but I'm not sure why. Are there more examples/explanation available somewhere?

[0] http://kanut.studentenweb.org/DTSA/
 Linda K. Muthen posted on Saturday, July 24, 2010 - 5:56 am
Please send the input and data for number 1 along with your license number to support@statmodel.com. Send also the two outputs showing that TRAINING and KNOWNCLASS results differ. They should be the same. Please use Version 6 for all analyses.
 Luciana Quaranta posted on Sunday, January 06, 2013 - 1:38 am
Dear professors,
I am running a path model with discrete time surv analysis, mediators and shared frailty. The prop hazards assumption is violated for M on Y but not for the X and control (W) variables.

I have 15 time intervals and it makes no theoretical sense for the effect of M to fluctuate in each interval. Under a continuous surv analysis framework I would estimate a piecewise constant hazard model, with M as a time-varying effect and W and X as time constant, imposing the cutoffs of the piecewise intervals to be the points where the hazards cross. Could something similar be done for a discrete model with Mplus?

I tried:
f by u1-u15@1;
g BY u1-u4@1;
h BY u5-u15@1;
f on w;

!cī path
f on x;

!b path
g on m;
h on m;

!a path
m on x;

1) Could this work or is this implementation incorrect?
2) A problem is that I get residual variances for f g and h and also correlations between f, g, and h. Can they be excluded? Do the correlations between f, g, h influence the coefficients or standard errors of the other estimates of the path model?

If it is impossible to estimate this I could allow all effects to fluctuate in g and h, although less ideal. And I also face the problem of the correlations between g and h.

Thank you very much!

Luciana
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: