In this exercise, practice some techniques and procedures related to regression analysis, using data (more or less) from the Darr and Johns (2003) academic politics study used as part of this module's case. The data are comprised of a number of attitude measures and demographic characteristics of a relatively large (N=620) group of faculty members at a large university, obtained at the time of a major crisis in university operations. The variables are:
Demographics:
Age, Years in dept, Rank, Sex, salary, level, tch_rating, minority, admin_resp, Science, degree_tier, pub_prestige
Attitudes:
Role ambiguity, Role conflict, Task conflict, Relationship conflict, Political perceptions
Respondent perceptions of responsibility in current crisis:
Fault, blame
1. Start by bonding with your data and forming a personal relationship with it. Prepare appropriate descriptive statistics - remember, means and standard deviations (and possibly histograms) for interval variables, frequency tables for categorical and dummy variables. Try to tidy up your tables following the procedures outlined in the presentation on table-tidying. Are there any things of interest in these descriptive statistics?
2. Construct a couple of scatterplots looking at the relationships among some of the interval variables that interest you. What if anything do you discover this way?
3. Set up a regression model to predict salary (DV) from teaching ratings (IV). Be sure to request appropriate residuals plots. What are your results? What do you learn from the residuals plots? [HINT: here is a good guide on interpreting residuals plots.]
4.Since neither salary nor teaching ratings is particularly normally distributed (i.e., both are skewed), we might do better with the log transforms of these data (if you're not familiar with log transforms of data, here is a good general guide to the subject (Hopkins). Try your regression model with these transformed variables. Any better? How do you know?
5. Now let's try a multiple predictor model. Use salary as DV, and Age, Years in dept, Sex, level, tch_rating, minority, admin_resp, Science, degree_tier, and pub_prestige as IV's. How good is your prediction now/? What are the best predictors? Anything in the residuals?
6. Your prediction is good, but it's too complicated. See if you can reduce it some by using a stepwise procedure on the same model. What is the efficiency of prediction of your final model here? What predictors are left? Which have been excluded? Of those left, which are best? Anything in the residuals? Any overall comments or observations here?
7. Actually, including LEVEL as an interval variable in this last analysis is somewhat suspect, since while it has some ordinal properties it's really more of a categorical variable. So let's try another approach - separate regressions for each level. Use DATA | SPLIT FILE | COMPARE GROUPS | based on LEVEL to divide the sample into the four levels; then run the same regression (of course, without level as a predictor now. Compare the results you get for each level. Any differences in accounting for salary based on these factors between the different levels - i.e., are assistant professors' salaries set using different criteria from those used for full or chaired professors? Test and comment.
8. Try another couple of regression analyses, single or multiple, among variables of your choice. Explain why you set up the analyses you did, and what you found.
9. Any overall comments on using regression techniques?