My nominal GEE's, with the DV having three groups, produced the following output:

                           NominalGEE Regression Results                           
Dep. Variable:                        task   No. Observations:                  112
Model:                          NominalGEE   No. clusters:                        2
Method:                        Generalized   Min. cluster size:                  44
                      Estimating Equations   Max. cluster size:                  68
Family:                       _Multinomial   Mean cluster size:                56.0
Dependence structure:         Independence   Num. iterations:                    31
Date:                     Tue, 22 Nov 2016   Scale:                           1.000
Covariance type:                    robust   Time:                         16:14:50
                 coef    std err          z      P>|z|      [95.0% Conf. Int.]
HF[1.0]        0.9400      0.436      2.156      0.031         0.085     1.794
LFHF[1.0]     -0.6134      0.082     -7.454      0.000        -0.775    -0.452
SDNN[1.0]      0.6610      0.053     12.477      0.000         0.557     0.765
pNN50[1.0]    -1.2769      0.045    -28.094      0.000        -1.366    -1.188
HF[3.0]        1.1308      0.019     59.228      0.000         1.093     1.168
LFHF[3.0]      0.5117      0.050     10.233      0.000         0.414     0.610
SDNN[3.0]     -0.4080      0.223     -1.826      0.068        -0.846     0.030
pNN50[3.0]    -0.1943      0.386     -0.503      0.615        -0.952     0.563
Skew:                          0.4459   Kurtosis:                      -1.1039
Centered skew:                 0.4320   Centered kurtosis:             -1.0652

If I understood correctly, the nominal GEE uses one group as reference group, and then calculates the IVs' coefficients for the two other groups (group 1.0 and group 3.0 in this case). The coefficients depict a change in logged odds, given the value of the IVs. In other words, with the coefficients you calculate the chance of belonging to a particular group, as opposed to the reference group.

However, is it possible to infer the odds/chance to belong to group 1.0 as opposed to group 3.0 from this graph? Or do I have to perform a new GEE with a different reference group?

  • $\begingroup$ You have to do another model with a different reference group. I just noticed that you only have 2 clusters (i.e. only 2 different levels of the clustering variable used in the GEE model). If this is correct, you shouldn't use a GEE model at all. It is often recommended that you have at least 50 clusters. Perhaps you can just include the clustering variable as another independent variable instead? $\endgroup$
    – JonB
    Nov 23, 2016 at 8:49
  • $\begingroup$ Hi @JonB, thanks for the answer. I indeed only have two clusters (participants) in my pilot study. I first wanted to do a Friedman Chi-2 test, where I found main effects, but I wasn't able to perform post-hoc testing (stats.stackexchange.com/questions/246719/…). Could you suggest another test where I can compare the different levels of the DV? Nb. the Friedman question is about a different (ordinal) set of data, but the idea is the same I guess) $\endgroup$ Nov 23, 2016 at 9:15
  • 1
    $\begingroup$ You can perhaps do a multinomial logistic regression? $\endgroup$
    – JonB
    Nov 23, 2016 at 9:25
  • $\begingroup$ @JonB thank you, it seems to be a lot easier. Final question, is the multinomial regression also fit for ordinal data? $\endgroup$ Nov 23, 2016 at 10:08
  • $\begingroup$ Just found the answer to my own question. You lose order of the ordinal variable with a multinomial regression (theanalysisfactor.com/…). An ordered logistic regression, however, has not yet been implemented in Python (github.com/statsmodels/statsmodels/issues/807). Only available Python alternative is the GEEOrdinal. $\endgroup$ Nov 23, 2016 at 10:19


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.