+ Reply to Thread
Results 1 to 3 of 3

Thread: GEE vs. GLMM

  1. #1
    Points: 8,343, Level: 61
    Level completed: 65%, Points required for next Level: 107

    Thanked 7 Times in 7 Posts

    GEE vs. GLMM

    hello all,

    I have a question, related to my previous post, but deserves it's own post.

    Lets say I have a binary dependent variable, with one or two independent variables and repeated measures. Lets say I also have one random effect, the center in a multicenter study.

    I wanted to ask which model is more appropriate for repeated measures with binary response, GEE or GLLM? I wanted to ask what is the difference between them, theoretically speaking. I read that one is marginal and one conditional. Can you give an answer for less expert people ? And if I add the random effect, do I have to use GLMM ?

    (in SAS lingo I want the difference between GEE with GENMOD and GLMM with GLIMMIX)


  2. #2
    Points: 15,159, Level: 79
    Level completed: 62%, Points required for next Level: 191
    victorxstc's Avatar
    Thanked 332 Times in 297 Posts

    Re: GEE vs. GLMM

    As far as I understood, both can be useful in your case. If you care more about the random effect, a GLMM might be better, and if you care less about the random effect, the GEE can be good, while GLMM can be more conservative in that case (still quite good, but less likely to give significant results). So I suggest you to apply both designs and see what is more consistent with the theory.

    If you wanted to use GEE, there is no need (or I think an option) to include the random effect. It will automatically handle that part once the model was fed properly into the software.
    "victor is the reviewer from hell" -Jake
    "victor is a machine! a publication machine!" -Vinux

  3. #3
    Cookie Scientist
    Points: 13,431, Level: 75
    Level completed: 46%, Points required for next Level: 219
    Jake's Avatar
    Austin, TX
    Thanked 584 Times in 438 Posts

    Re: GEE vs. GLMM

    Which one is more appropriate depends on which question you want to answer.

    Both GEE and GLMM are, in this context, essentially ways of conducting a logistic regression while accounting for the violation of independence we expect to be present in the data given what we know about the structure of the data. But the two methods handle this non-independence in different ways.

    The GEE approach is basically to compute the logistic regression estimates as usual (it is actually not exactly the same as usual but usually is pretty close), but then to apply a possibly substantial correction to the standard errors to account for the non-zero covariances between observed responses from the same cluster. The GEE estimates are "marginal" in the sense that the parameter estimates themselves are indifferent to the grouped structure of the data. The effect of X on Y is estimated at the "population level" across all the clusters in the population.

    The "conditional" / GLMM approach is based on computing a weighted average of all of the individual within-cluster models regressing Y on X in that cluster. The cluster-level models are weighted according to the cluster size and the within-cluster variance in the predictor variables. Standard errors for these estimates also take into account the between-cluster variance in their cluster-level parameter estimates.

    These two approaches can in some cases yield very different results, because they are answering different questions. The population-average effect might be different from the average cluster-level effect. A clear illustration of this is shown in the graph at the top of the Wiki page on Simpson's paradox: http://en.wikipedia.org/wiki/Simpson's_paradox
    “In God we trust. All others must bring data.”
    ~W. Edwards Deming

  4. The Following User Says Thank You to Jake For This Useful Post:

    bugman (12-13-2013)

+ Reply to Thread


Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts

Advertise on Talk Stats