Frequently Asked Questions for the Project

Spring Semester, 2000

Where do we turn in the projects?
Put it in my mailbox in the Stat Dept (Suite 3000, SH-DH). If that gets full, we'll put a labeled box out for you to use as well.

How can I tell if prediction intervals are significantly shorter when additional predictors are added to a regression model?
Remember that the length of a prediction interval is essentially determined by the size of the RMSE of the fitted models. So, the question is asking if the RMSE has gotten significantly smaller. To see if this has happened, notice what would occur if all of the added variables were not useful -- all of their slopes would be zero. That is, unless some of the added variables are helping, the RMSE has not gotten significantly smaller. Now, how can you tell whether any of the added variables contributes significantly?

For the final group of questions (10-14), do I need to fit three separate regression equations, one each for the city and old/new suburbs?
No, just continue with the regression model used to answer the prior group of questions. Add the variables that you need in order to answer the listed questions and any others that you think will help with the final prediction (question 14). Keep all these in one equation, and use this one equation for all of 10-14.

How do you get JMP to do the prediction intervals?
Add a 'dummy row' to the JMP spreadsheet and fill in the values of the predictors that you are using (or just fill in them all from the conditions in Question 14). With the predictors filled in, use the 'Save Indiv Confidence' item to save the upper and lower prediction limits (use the $ button at the lower left of the regression output window to get the save menu for regression). However, this does not always work for me. If it does not work for you either, use the 'Save Prediction Formula' button -- and then add the +/- 2 RMSE to the predicted value.

Location is not significant in my model, but the interaction Location * 1/Sqft is. When I try to remove Location, JMP complains. What should I do?
Follow JMP's advice and keep the Location term since it appears in an interaction term. If you have an interaction, say X1*X2, then both X1 and X2 should be predictors in your model as well -- regardless of the p-value for them.

The sign on some of my coefficients seems wrong (eg, negative rather than positive). What went wrong?
Perhaps nothing. Don't worry too much about the sign unless the estimated coefficient is significant. (Recall the intercept of the diamond price example.)

I got a negative cost for parking in the old suburbs. How can that be?
If you check, the added variable is likely not significant. Thus, as far as the statistics are concerned, the negative value is within random variation of zero. In more practical terms, the negative value might represent some form of incentive designed to attract customers to the old suburbs.

Can I work with someone in Prof. Krieger's class?
No, your team must consist of classmates in this section. I'd encourage you to talk with friends in other sections, but you will each need to submit an analysis of your own data set.