[R] what are supressors doing in hierarchical partitioning??

[jones christian]

Hello,
After fitting  a model with glm (glm(A~a+b+c+d,binomial)) I want to have a closer look at the goodness of fit for each of the 4 independant variables by using "hierarchical partitioning". The results (see below) indicate that three variables (MEDIANH_HE+CLOJUL+K95) act as a supressor because their dependant contribution is negative. For two variables the total contribution is near 0 (CLOJUL+N5_97). Surprisingly all variables are highly significant in the glm model. How can that be? And do the results from  hierarchical partitioning mean that I better not use this model for prediction??? (PS: I went to a lot of effort fitting this model)
many thanks in advance
Christian




> hier.part(NACHTIGALL.x,A2HP_N5,family="binomial",gof="logLik",barplot=T)
$gfs
 [1] -131.61028 -106.38458 -131.60991 -124.60603 -131.51891 -101.70663
 [7] -102.38016 -105.11808 -124.55582 -131.51761 -124.31356  -98.96125
[13]  -98.59341 -100.65798 -124.29683  -95.51063

$IJ
                   I         J        Total
MEDIANH_HE 26.895004 -1.669303 2.522570e+01
CLOJUL      2.511079 -2.510707 3.721918e-04
K95         5.245935  1.758320 7.004255e+00
N5_97       1.447637 -1.356262 9.137552e-02

$I.perc
                   I
MEDIANH_HE 74.502107
CLOJUL      6.955965
K95        14.531814
N5_97       4.010114



--