Hi folks, I am trying to develop a reasonable (not 'perfect') logistic/binary model and would appreciate critiques/suggestions/references. I'll start with an example of data:
Suppose I have 10 cases, each with a binomial outcome (with n_1, ..., n_10 trials) and let's say 2 predictors, gender and age. For example case one has n_1=7 trials and 2 successes, case 2 has n_2=1 trial and 0 success, etc. Suppose the binary outcomes are legal decisions (say success is 'guilty') and each of the 10 cases corresponds to 10 judges. Now, formally, it doesn't matter if I view it as n=n_1+...+n_10 bernoullis or 10 binomials. Viewing it as bernoulli doesn't, at first, seem right - since you have to consider "judge effect". But I can rationalize it by viewing it as Binomial. It seems reasonable to assume that a new case/judge with the exact same value of gender and age can have the same estimated logit. Somehow though, it seems like I should still take into account a "judge effect", by some nested model. The quick and dirty thing I first did was just include judge as a nominal predictor. But that doesn't seem to make sense. I am trolling for someone to poke holes in my thinking and would appreciate any feedback/ref's. Or if what I'm doing seems "o.k.", that's helpful too. Many thanks. ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
