Bad Compatibility Of Mumin With Multinom (Nnet) Objects Having 3 Level Response Variables?

Question

[l.duvaux]

I face a problem when trying to use the function model.avg of the MuMin package on objects resulting from the function multinom (package nnet). Although the manual list as multinom compatible, the model.avg function doesn't return the model average coefficients for the explanatory variables as I expect.

I know that the problem doesn't exist when the multinom reponse variable as 2 levels only, so I guess it depends on the structure of the resulting multinom object. However, I don't know how to fix this issue.

Here is an example code where the function model.avg works as expected:

#### MuMIn example  
library(MuMIn)  
data(Cement)  
fm1 <- lm(y ~ ., data = Cement)  
dd <- dredge(fm1)  
summary(model.avg(dd, subset = delta < 4))$coefmat

Results:

               Estimate Std. Error Adjusted SE   z value  Pr(>|z|)
(Intercept) 64.69312754 22.2354794  22.4624137 2.8800613 0.0039760
X1           1.45579791  0.2036676   0.2193043 6.6382554 0.0000000
X2           0.62502598  0.1202570   0.1291705 4.8387682 0.0000013
X4          -0.47600710  0.2215167   0.2309395 2.0611767 0.0392862
X3          -0.02153196  0.3767120   0.3924356 0.0548675 0.9562440

In this case, the 'model.avg' function returns the Model-averaged coefficients (column Estimate) for the explanatory variables (i.e. X1-X4) as expected.

Now, if I want to apply this to a multinom object (response variable with at least 3 levels):

#### application to multinomial regression
library(foreign) ; library(nnet) ; library(MuMIn)
ml <- read.dta("http://www.ats.ucla.edu/stat/data/hsbdemo.dta")
ml$prog2 <- relevel(ml$prog, ref = "academic")
test <- multinom(prog2 ~ ses + write + schtyp, data = ml)
# [...] verbosity not displayed here
dd <- dredge(test)
# [...] verbosity not displayed here
summary(model.avg(dd, subset = delta < 4))$coefmat

Results

         Estimate Std. Error  z value    Pr(>|z|)
general  2.850316   1.169132 2.437977 0.014769732
vocation 5.134921   1.173878 4.374322 0.000012181

Here, the function returns "model average coefficients" for the different levels of the response variable (i.e. general and vocation) rather than for the model average coefficients for the explanatory variables.

If this is possible, could you tell me please how to get the Model average coefficients for the explanatory variables from multinom objects?

Many thanks in advance.