It is possible to have both additive and multiplicative effects, but
not in the models that we used. If the response variable is log
transformed, the effects on the untransformed variable will be
multiplicative. Without log transformation, the effects are additive
Maybe it's easiest to explain with some examples. Imagine x is a
factor with 2 levels, and x1 and x2 are dummy variables (x1=1 and x2=0
for the 1st level, and x1=0 and x2=1 for the 2nd level), and the
response variable y is log transformed (or you used a log-link
function).
log(y) = a + b1*x1 + b2*x2
y = exp(a + b1*x1 + b2*x2) = exp(a) * exp(b1*x1) + exp(b2*x2)
Thus, in case of the first level, you get
y = exp(a) * exp(b1*1)
and for the second level
y = exp(a) * exp(b2*1)
If you have a continuous variable x, the interpretation of the summary
table is easiest when this x variable is log-transformed as well, thus
log(y) = a + b* log(x)
y = exp[a + b*log(x) ] = exp(a) * x^b
So different values of x results in multiplying y relative to the
'base line' exp(a).
Hope this cleared things up a bit?