Linearizing the response
I am studying a book on generalized linear model and they suggest plotting the values of each predictor against the "linearized response". The example is using log link with Poisson distribution. They give this "linearized response" as XB +(y-mu)*(d XB/d mu) where XB is the linear predictor and mu (I think) is the predicted response. They then use a formula in the computer program which looks like XB +(y-mu) /mu. I dont understand this linearizing of the response, can anyone explain. I understand plotting the log(response) against the predictor but not this...