Ch. 11, Part 1: do pp. 253 (under the "Type of Model" tab, check Custom, then select binary distribution, logit link function - Run the model under the "Estimation" tab, change the "Scale Parameter Method" to "Deviance"; to get Figs. 11-5 and 11-6, rerun the model but change "Scale Parameter Method" back to "Fixed", then under the "Statistics" tab, set "Analysis Type" to "Type I and III"; to get Fig. 11-8, rerun the model but under the "Predictors" tab, click Options, and change the "Category Order" of the factors to "Descending". This is the logit regression model.
Ch. 11, Part 2: Run the model on p. 264, but under the "Type of Model" tab, check "Poisson Loglinear" (or you could check Custom, then select Poisson distribution, log link function, which would be the same thing). This is the Poisson regression model. Since it is for count/duration data, not an offset variable was required, as it is whenever you have rate rather than just count data.
Ch. 11, Part 3: Run the model on p. 268. but under the "Type of Model" tab, check "Binary Probit" (or you could check Custom, then select binary distribution, probit link function). This is the probit regression model.
Ch. 11, Part 4: Run the model on p. 269, but under the "Type of Model" tab select Custom, then select Poisson distribution, log link function and set the parameter to 0 (selecting "Poisson Loglinear" above defaults to ancillary parameter=1); then under the "Statistics" tab, check that you want the Lagrange multiplier test). This is the Poisson regression model for count data. There is no offset variable as with rate models.
Ch. 11, Part 5: Run the model on p. 272. First, as explained under "Specifying the Analysis", you must use Data, Aggregate, to create a new dataset in which the variable N-Break represents the counts in a 3-way table whose axes are vote, college, and male. Switch to the new dataset before running the menu choices shown at the bottom of p. 272. In the "Type of Model" tab, select "Poisson loglinear". This shows that loglinear analysis can be implemented through GZLM, albeit with fewer options than Analyze, Loglinear in the SPSS menus. Norusis does not ask for it, but run the Part 5 model a second time, using the Options button under the Predictors tab to change the order to "Descending", which will switch reference categories.
Ch. 11, Part 6. Run the model on p. 274. In the "Type of Model" tab, select "Gamma with log link". This is the gamma regression model.
Ch. 12, Part 1: Do the run on p. 281. For the "Type of Model" tab, select Custom, binomial distribution, logit link function. Under the "Statistics" tab, set "Analysis Type" to "Type I and III". Also under "Statistics", check that you want "exponential parameter estimates" [to get Exp(b)] and check that you want the "Working Correlation Matrix" printed.
Ch. 12, Part 2: Now go back to the "Repeated" tab and change the working correlation matrix from "Unstructured" to "Exchangeable", and run the analysis over.
Ch. 12, Part 3: Now go back to the "Models" tab and remove Age and Sex from the model, leaving only Treatment (and the intercept). Also go to the Repeated tab and change the "Working Correlation Matrix" back to "Unstructured". Run the analysis a third time.
Ch. 12, Part 4: Go back to the "Models" tab and add Age and Sex back into the model. Then go to the "Repeated" tab and change the "Covariance Matrix" setting from the default "Robust Estimator" to "Model-based estimator". Run the model a fourth time.
Note: Pp. 290-296 contain a second GEE model, for repeated-measures Poisson regression. You may wish to do this as exercise but we will not go over this output in class.