Looking for SAS code for Joint Model between survival and longitudinal analysis model

Accepted Solution Solved
Reply
Contributor
Posts: 22
Accepted Solution

Looking for SAS code for Joint Model between survival and longitudinal analysis model

Dear all,

I'm now handling with longitudinal data analysis.

I want to build up a joint model of survival and mixed effect model, but I haven't found an available SAS program for this issue.

My outcome variable is Hypertension (value 0 1), covariates are composition of both categorical and continuous vars. Most of them are measured yearly for 5 times.

Is there anybody having experience with this statistical method and SAS program or macro for this analysis?

 

 


Accepted Solutions
Solution
‎04-15-2016 10:13 PM
Valued Guide
Posts: 2,106

Re: Looking for SAS code for Joint Model between survival and longitudinal analysis model

Trang,

 

I'm pretty rusty on time-dependent covariates, so I can just refer you to, say, Paul Allison's BBU on survival analysis.

 

I can comment on treating sodium as continous versus quantiles.  If a predictor is reasonalby linear in the hazard space, it will be a better predictor than quantiles for a couple of reasons.  For one, with more granularity in the measure, you can get better estimates.  The second is that you are not really looking at quantiles as a test; you are looking at a 4-level categorial data and that dilutes the effect.  I'm not sure how you can test for monotonicity.  You could test for linear or quadratic effects, but that takes additional assumptions on the quantiles that may not be supported by the data.

View solution in original post


All Replies
Valued Guide
Posts: 2,106

Re: Looking for SAS code for Joint Model between survival and longitudinal analysis model

Looking at your previous query, the granularity of the data (time recorded in whole years) might allow you to do the analysis completely in PROC MIXED.

 

That said, I'm not sure that a longitudinal data analysis is appropriate for hypertension as an outcome.  For "essential" hypertension (i.e. the common type that adults acquire), once the medical diagnosis is made, it continues for the person's lifetime.  The hypertension may be in control, but the diagnosis stays.  In that case, diagnosis and time to diagnosis are appropriate for a survival analysis (though the huge number of ties will reduce the power).

 

If your true outcome variable is "blood pressure in control", then a longitudinal analysis may be appropriate.

Contributor
Posts: 22

Re: Looking for SAS code for Joint Model between survival and longitudinal analysis model

Dear Doc_Duke,

Thank you very much for your reply. I agree with you. Can I ask you some more questions?

 

  1. If I use Cox model with time-dependent covariates (for instance,  Sodium is my main interest as a predictor for hypertension, and it varies over time so I want to use the most recent measure of sodium, i.e measure at the time of censor or incident hypertension, in my model), is there any differences in the interpretation of this model compared to the standard Cox model (using baseline value of covariates)?
  2. Another problem for my Cox model is that when I treat Sodium as a continuous variable (in grams), the results is highly significant.But when it's put into the model as quartiles, the results become no more significant across categories of quartiles, with lowest quartile as a references (Q2 vs Q1, Q3 vs Q1, Q4 vs Q1). How can we explain for this issue?

 

Thank you very much for your support.

Trang

Solution
‎04-15-2016 10:13 PM
Valued Guide
Posts: 2,106

Re: Looking for SAS code for Joint Model between survival and longitudinal analysis model

Trang,

 

I'm pretty rusty on time-dependent covariates, so I can just refer you to, say, Paul Allison's BBU on survival analysis.

 

I can comment on treating sodium as continous versus quantiles.  If a predictor is reasonalby linear in the hazard space, it will be a better predictor than quantiles for a couple of reasons.  For one, with more granularity in the measure, you can get better estimates.  The second is that you are not really looking at quantiles as a test; you are looking at a 4-level categorial data and that dilutes the effect.  I'm not sure how you can test for monotonicity.  You could test for linear or quadratic effects, but that takes additional assumptions on the quantiles that may not be supported by the data.

Post a Question
Discussion Stats
  • 3 replies
  • 555 views
  • 1 like
  • 2 in conversation