How to interpret geeglm results? I have a question about the geeglm function in the GEE package in R. I am a beginner in statistics and the estimates in the output of the model are puzzling me. How can I tell that the estimates are ok and how can I recognise overdispersion?

*

*The output looks like this:
Call: geeglm(formula = logSUM ~ teplota + radiace.rezidualy + srazky2, 
family = gaussian, data = PH1, id = den, corstr = "ar1") 

Coefficients: Estimate Std.err Wald Pr(>|W|) 

(Intercept) -0.933429 0.515005 3.29 0.070 . 

teplota 0.163241 0.028613 32.55 1.2e-08 *** 

radiace.rezidualy -0.000842 0.000397 4.50 0.034 * 

srazky2 -0.153942 0.037414 16.93 3.9e-05 *** 

--- Signif. codes: 0 ‘’ 0.001 ‘’ 0.01 ‘’ 0.05 ‘.’ 0.1 ‘ ’ 1 

Estimated Scale Parameters:

         Estimate Std.err
(Intercept)    0.8  0.08

Correlation: Structure = ar1  Link = identity

Estimated Correlation Parameters:

   Estimate Std.err
 alpha    0.7  0.05

 Number of clusters:   24   Maximum cluster size: 16

Thank you for any feedback or suggestions.
 A: Can you please post the entire code for your model. Without seeing the code, I would say that you might have log coefficients. This is entirely dependent on the link function, however. If you do have log coefficients, in order to get a coefficient that is easy to understand you must exponentiate the estimate: exp(Beta). This will give you a rate ratio. Numbers over 1, for instance 1.1, would mean a 10 % increase in the response variable. Numbers less than 1, for  instance .9, would be a decrease. These numbers are related. Divide 1 by 1.1 and you will get .9 and if you divide .9 by 1 you will get 1.1. See here for more information (https://onlinecourses.science.psu.edu/stat504/node/180).
A: Since your outcome variable is logSum, you have to exponentiate the coefficients.
This is very easy to do in R. For teplota, the exponentiated coefficient would be 1.17732. In r, the code is very straight forward:
exp(.163241)

This means that, for a one unit increase in teplota, there would be a 17 % increase in the outcome variable (see here: http://www.ats.ucla.edu/stat/mult_pkg/faq/general/log_transformed_regression.htm.)
The clusters are simply the groups of the id variable, in this case den. You have 24 levels of den and the largest group is 16. As for the correlation structure, autogressive AR(1) is generally used for time series data. This is because time points are often correlated with one another. You may not need this. In order to check for autocorrelation, you need to generate an acf plot. I am not sure how to do this using geeglm, so I will use lme4 with the lmer function.
install.packages("lme4")
install.packages("itsadug")
require(lme4)
require(itsadug)
den <- factor(PH1$den)

m1 <- lmer(logSum ~ teplota + radiace.rezidualy + srazky2 + (1|den), data = PH1)
summary(m1) ## check coefficients

acf_resid(m1, main = "ACF Plot") ## if any vertical lines are above the horizontal blue dotted line you have autocorrelation.

If you have autocorrelation, as indicated by the acf plot, you can proceed with your original model. You could explore other packages, for instance, nlme or lme4. I would recommend lme4, but you will have to do a little work to get p-values that you are comfortable with. In the package description, you will have to browse the table of contents for pvalues. 
car::Anova(m1) ## for p-values

If you do not have autocorrelation, my model using lme4 would be appropriate.
If this answered your question, please mark your post as answered. If this does not answer your original question, please describe your situation in more detail.
