The R Book Generalized Linear Models精选.pdf

13 Generalized Linear Models We can use generalized linear models (GLMs) – pronounced ‘glims’ – when the variance is not constant, and/or when the errors are not normally distributed. Certain kinds of response variables invariably suffer from these two important contraventions of the standard assumptions, and GLMs are excellent at dealing with them. Specifically, we might consider using GLMs when the response variable is: count data expressed as proportions (e.g. logistic regressions); count data that are not proportions (e.g. log-linear models of counts); binary response variables (e.g. dead or alive); data on time to death where the variance increases faster than linearly with the mean (e.g. time data with gamma errors). 4 0 1 8 3 e e c c 6 n n a 2 a i i r r a a 4 V V 1 2 0 0 0 2 4 6 8 10 0 2 4 6 8 10 Mean Mean