HW 1

Comments
  1. The only mistake people made was an assumption that if year is significant and corps is significant, than model with year and corps is significant.  Usually this is true, but not always, and in this home work you were specifically asked to "Test the hypothesis that neither year or corps is significant in predicting the number of deaths".
  2. No general problems here.

    3. A few people try to run loops of 10 -30 iterations which is a good idea, but running it till convergence in deviance is better idea, and only one person did it.  In any case, you HAVE TO CHECK IF CONVERGENCE ACHIEVED. Here is my version of that loop:

    data<-read.table(file="/afs/stat.lsa.umich.edu/users/faculty/faraway/.public_html/data/horsekick.data",header=T)
    attach(data,pos=1)
    y<-deaths
    mu<-y
    fit<-lm(log(mu)~year + corps,weights=mu)
    fit.dif<-fit$fit*0+1

    while(sum(abs(round(fit.dif,2)))>0){
      fit.old<-fit$fit
      z<-fit$fitted+(y-exp(fit$fitted))/exp(fit$fitted)
      w<-exp(fit$fitted)
      fit<-lm(z~year + corps,weights=w)
      fit.dif<-fit.old-fit$fit
    print(sum((fit$fit)*y-exp(fit$fit)-gamma(y+1)))
      cat("\n")
    }
    fit.glm<-glm(y~year+corps,family=poisson)

    fit$coef
    fit.glm$coef
    sum(abs(fit.glm$coef-fit$coef))

    sum((fit$fit)*y-exp(fit$fit)-gamma(y+1))
    sum(log(fit.glm$fit)*y-fit.glm$fit-gamma(y+1))

    Note that I used data file located in Pf. J. Faraway directory, and didn't make another copy of it.  Also I didn't have to add 0.1 to log(mu) because I put log(mu) right in to the model, and Splus assign weight 0 to all NAs.