多水平模型(英文原著) chap7.doc

Chapter 7 Discrete response data 7.1 Models for discrete response data All the models of previous chapters have assumed that the response variable is continuously distributed. We now look at data where the response is essentially a count of events. This count may be the number of times an event occurs out of a fixed number of ‘trials’ in which case we usually deal with the resulting proportion as response: an example is the proportion of deaths in a population, classified by age. We may have a vector of counts representing the numbers of events of different kinds which occur out of a total number of events: an example is given in chapter 3 where we studied the number of responses to each, ordered, category of a question on abortion attitudes. Statistical models for such data are referred to as ‘generalised linear models’ (McCullagh and Nelder, 1989). A 2-level model can be written in the general form (7.1) where is the expected value of the response for the ij-th level 1 unit and f is a nonlinear function of the ‘linear predictor’ . Note that we allow random coefficients at level 2. The model is completed by specifying a distribution for the observed response . Where the response is a proportion this is typically taken to be binomial and where the response is a count taken to be Poisson. Equation (7.1) is a special case of the nonlinear model studied in chapter 5 and we shall be using the results given there. It remains for us to specify the nonlinear ‘link’ function f. Table 7.1 lists some of the standard choices, with logarithms chosen to base e. In addition to these we can also have the ‘identity’ function , but this can create difficulties since it allows, in principle, predicted counts or proportions which are respectively less than zero or outside the range (0,1). Nevertheless, in many cases, using the identity function produces acceptable results which may differ little from those obtained with the nonlinear functions. In the following sections we co