Discussion: View Thread

  • 1.  Single response clusters and GEE

    Posted 04-12-2012 12:41
    Dear All,

    Does anyone of you know what assumptions does PROC GENMOD makes while fitting GEE logistic regression to estimate the working correlation matrix when some of the clusters only have one response.

    To be more precise there are 58 physicians (clusters) and the cluster size (number of patients being treated by each physician) varies from 1 to 39. There are 10 physicians (clusters) who only have one patient.

    We have about 17%  physicians  with a single subject  however the remaining 83%do have larger number of patients and from clinical perspective we would still like to find an estimate of within cluster correlation if possible. Although when I fit GEE with structured or some other correlation structures it does give estimates of correlation matrix but I am not sure how it is calculating it with 17% clusters of size 1. From what I gather from the link below it seems it is assuming them to be missing values:
    http://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/viewer.htm#statug_genmod_sect043.htm


    Also, the option covtest doesn't work in GEE does anyone know how to test for significance of within cluster correlation obtained from the GEE model fit.

    I would really appreciate any comments and suggestions.

    Looking forward for your help and guidance.
    Best Regards,
    Tasneem


    -------------------------------------------
    [Tasneem] [Zaihra]
    [Assistant Professor]
    [Concordia University]
    [Montreal]
    [QC]
    [Canada]
    -------------------------------------------


  • 2.  RE:Single response clusters and GEE

    Posted 04-16-2012 19:55
    hi Tasneem

    last time I checked, singleton clusters weren't used in the estimation.  but I forwarded your question to a friend and colleague John Preisser who wrote the following in reply:

    "First. clusters of size one do not contribute any information to the estimation of within-cluster correlation so the observation Tasneem makes about GENMOD is not specific to that procedure.  Second, the first-order GEE (GEE1) of Liang and Zeger does not provide variance estimation for within-cluster association parameter estimates, nor significance testing for them; it uses simple moment estimators.   Next, confidence interval estimation of within-cluster correlations would generally be of greater interest than significance testing.  If the outcome is binary, the alternating logistic regressions options in GENMOD can be used to estimate some simple structures of within-cluster association based on the pairwise odds ratios.  For estimating correlations, Prentice's (1988) GEE procedure is useful for large clusters of correlated binary data.  It provides estimates and standard errors for both parameter estimates in the marginal mean model and in the marginal pairwise correlation model.  I have developed a SAS/IML macro for Prentice's GEE which can be downloaded from my website.  Admittedly, the documentation could be better, but it is easy to use, once the Z matrix for the correlation model is set up (analogous to the X matrix for the marginal mean model). 

    "I could provide more detail if Tasneem wants to contact me after Wednesday.  As I am not a member of this section or forum, please feel free to post my response and adapt it as you see fit.  If you wish to attribute the response to me (not necessary), just put the above paragraph as written in quotes.
    "   John Preisser [jpreisse@bios.unc.edu]





    -------------------------------------------
    Stuart Gansky
    University of California, San Francisco
    -------------------------------------------