GEE or Cox regression - which one is the most appropriate ?

[Michael]

Hello,

I have a data set with 1800 respondents (adolescents) who were asked to complete the same questionnaire 5 times over a 6-year period. I am interested in modelling smoking onset (binary outcome) based on several variables that might be risk factors. Respondents are clustered within schools, so there is a need to take the correlated data structure into account. Both GEE and Cox regression allow for this. However, with GEE, SPSS does allow only for the use of one correlation structure (e.g.: same auto-regressive (AR1) structure for repeated measurement over time and within school).

As far is I know,
I) Cox (proportional hazards) regression answers the following questions: a) what is the time when the child was at risk for smoking onset (time to outcome) b) what are the predictors related to the outcome ?
II) GEE answers the following question: what are the predictors related to the outcome ? (i.e. does not ask a question about time, but can take time into account in the model).

Since the outcome is smoking onset, it can only happen once. Once the respondent report having tried smoking, it should be discarded from the analysis.

In your opinion, what do I need to consider in order to choose Cox regression or GEE for my analysis ?

Sincerely,

Michael

[Max Jasper]

Take a look at Cox segmented time-dependent covariates.

 

After 1^st smoking instance/time, risk factors are eliminated from analysis for that person using segmented time-dependent covar:

 

Time program.

Compute smoker = (T_ >= TimeFirstSmoked).

Compute agerisk = (T_ >= TimeFirstSmoked)*age.

…

…

COXREG

….

*Use risk covariates as defined above.

 

Max.

 

--
To post a new thread to MedStats, send email to MedS...@googlegroups.com .
MedStats' home page is http://groups.google.com/group/MedStats .
Rules: http://groups.google.com/group/MedStats/web/medstats-rules

[Michael]

Hi,

Thank you for your answer.
But why would chose such a "Cox segmented time-dependent covariates" (or another Cox model) instead of a GEE model (my question) ?
What criteria do you use ?

Michael

[Thomas Keller]

Could it make sense to analyse your data with recurrent event analysis
(within Cox regression)?
Each start of a smoking period is a event?

Thomas

[Michael]

Hi Thomas,

Thank you for your suggestion. For conceptual reasons, smoking initiation (= event) can happen only once. Therefore, if I undersand well what you suggest, I do not think that "recurrent event analysis" is relevant here.

Each participant has the same baseline at grade 5 and then 5 successive follow-up periods (some did not respond to all measurements) over ~6 years. Only those who have reported never having experimented with smoking are eligible to be included at baseline. I used a long datafile format with SPSS, where each participant might appear on up to 6 lines. With that structure, it is possible to model time and also time-dependent covariates. When a participant reports that he/she experimented with smoking (= smoking initiation), he/she has to be discarded from the analysis since the critical event happened.

I am having a hard time choosing between GEE or Cox regression, because I am not clear about the difference(s) between the questions asked by these two techniques.

Regards,

Michael

[Dom]

I was actually coming to the group to ask a very similar question.  In my situation, I have repeated measures (time of discharge, one month, six month, and twelve month follow-up).  I am looking at a dichotomous outcome.  I am interested in a possible interaction between two main independant variables, which I would like to model as time-varying covariates.  I found a source that discusses interaction in time-dependant covariates in Cox models.  I don't know very much about GEE, but I was wondering if someone could point me to a source that would show whether or not you could show an interaction using GEE.  (To me, it seems like there shouldn't be a problem, but my research mentor would prefer a source - and neither of us are statisticians). 

 

Thanks in advance  

 

[Michael]

Hi, you can model an interaction using GEE, which is similar to logistic regression: see Kleinbaum, D. G., & Klein, M. (2010). Logistic Regression: A Self-Learning Text (3rd ed.). Springer.

http://books.google.ca/books?id=J7E0JQweHkoC&printsec=frontcover&dq=logistic+regression+a+self-learning+text&hl=fr&ei=DV-QT_D1AamN6QHEq8mgBA&sa=X&oi=book_result&ct=book-thumbnail&resnum=1&ved=0CDYQ6wEwAA#v=onepage&q=logistic%20regression%20a%20self-learning%20text&f=false

[Michael]

Hi,

Is there anybody who knows a source regarding this question ?

Regards,

Michael

[Mike]

Hi Michael
It seems to me you have a grouped survival data set.
This can be tackled, not using a logistic link, but a complementary
log-log link, and is described by Prentice and Gloeckler, Biometrics
1978, 34: 57-67.
The complementary log-log model has some nice propertiies with regard
additivity that the logistic does not for survival type data.
You could fit this using GEE and use a robust standard error to allow
for clustering.

Mike