Note: Examples 3, 5, and 6 in Ch. 10 are covered next week.
Use the file hsb12.save, located at http://www.ats.ucla.edu/stat/paperexamples/singer/hsb12.sav. Obtain the output on pp. 201, 206, and 215-216.
In the repeated measures section, use the file willett.sav, located at http://www.ats.ucla.edu/stat/paperexamples/singer/willett.sav. Note that the variable "score" in the Norusis book is the variable "y" in the actual dataset! Likewise, the "cog" variable in the text is the "covar" variable in the dataset. Also note the data are already restructured as described on pp. 228-229 of the text. Obtain the output on pp. 230, 232, and 233 and be prepared to interpret it.
Also in the repeated measures section, obtain the output for pp. 236, 237, and 242, but we will not cover this in depth in class. For pp. 236, 237, and 242, use the file "growth study.sav" from the Norusis data disk. Note "subject" in the file is the "id" variable in the text.
- P. 236 shows how by using the "Repeated" specification on the opening dialog box (the one where you specify Subject) and not specifying anything under the "Random" button you can obtain identical results to p. 230.
- P. 237 uses a new dataset and model to generate a "Covariances of Residuals" table using the assumption of unstructured ccovariance type. You can re-run the model using assumptions of compound symmetry (p. 238), Autoregressive AR1 (p. 239), and Toeplitz. Table 10-3, p. 240, uses the AIC and BIC goodness-of-fit measures from the "Information Criteria" tables to compare assumptions. For AIC and BIC, lower is better. Toeplitz is best by AIC, but only a little better than compound symmetry, and moreover, compound symmetry is best by BIC, which penalizes for lack of parsimony (Toeplitz has more parameters), so compound symmetry is the best assumption. Note, however, that this is a data-driven decision method prone to over-fitting. Norusis does not note this, but does on p. 241 discuss likelihood ratio tests of whether a model under one assumption is significantly different from a model under another.
- P. 242 treats age as a random effect rather than as a fixed factor with four levels (8, 10, 12, 14). Finds age does predict distance and that the slope of age when predicting distance differs by gender.
- Go to Week 11 (HLM) and in the Explore section, click on the link for free student HLM software. Make sure it is installed on your computer by the time Week 11 comes around.
