1. Multiple regression - Inference for multiple regression - A case study IPS chapters 11.1 and 11.2 © 2006 W.H. Freeman and Company

2. Objectives (IPS chapters 11.1 and 11.2) Multiple regression  Multiple linear regression model  Confidence interval for β j  Significance test for β j  Anova F-test for multiple regression  Squared multiple correlation R 2

3. For p explanatory variables, we can express the y-values  y =  0 +  1 x 1 … +  p x p +e e is an error which is difference between each y-value and its mean. The statistical model for the sample data (i = 1, 2, … n) is then: Data = fit + residual y i =  0 +  1 x 1i … +  p x pi +  i The parameters of the model are  0,  1 …  p and . Where s is the standard deviation of the errors, e Multiple linear regression model

4. We selected a random sample of n individuals for which p + 1 variables were measured (x 1 …, x p, y). The least-squares regression method minimizes the sum of squared deviations, e i 2 (= y i – ŷ i ) 2 to express y as a linear function of the explanatory variables: ŷ i = b 0 + b 1 x 1i … + b k x pi ■As with simple linear regression, the constant b 0 is the y intercept. ■The regression coefficients (b 1 … b p ) reflect the change in y for a 1 unit change in each independent variable. They are analogous to the slope in simple regression. ŷ  y b 0 are unbiased estimates of population parameters  β 0 b p  β p

5. Confidence interval for β j Estimating the regression parameters β 0, … β j … β p is a case of one- sample inference with unknown population variance.  We use the t distribution, with n – p – 1 degrees of freedom. A level C confidence interval for β j is given by: b j ± t* SE b j - SE b j is the standard error of b j. - t* is the critical t for the t (n – 2) distribution with area C between –t* and +t*.

6. Significance test for β j To test the hypothesis H 0 :  j = 0 versus a 1 or 2 sided alternative. We calculate the t statistic t = b j /SE b j which has the t (n – p – 1) null distribution to find the p-value of the test.

7. For a multiple linear relationship ANOVA tests the hypotheses H 0 : β 1 = β 2 = … = β p = 0 versus H a : H 0 not true by computing the F statistic: F = MSM / MSE When H 0 is true, F follows the F(1, n − p − 1) distribution. The p-value is P(> F). A significant p-value doesn’t mean that all the explanatory variables have a significant influence on y — only that at least one of them does. ANOVA F-test for multiple regression

8. ANOVA table for multiple regression SourceSum of squares SSdfMean square MSFP-value ModelpSSG/DFGMSG/MSETail area above F Errorn − p − 1SSE/DFE Totaln − 1 SST = SSM + SSE DFT = DFM + DFE The standard deviation of the sampling distribution, s, for n sample data points is calculated from the residuals e i = y i – ŷ i s is an unbiased estimate of the regression standard deviation σ.

9. Squared multiple correlation R 2 Just as with simple linear regression, R 2, the squared multiple correlation, is the proportion of the variation in the response variable y that is explained by the model. In the particular case of multiple linear regression, the model is a linear function of all p explanatory variables taken together.

10. We have data on 224 first-year computer science majors at a large university in a given year. The data for each student include: * Cumulative GPA after 2 semesters at the university (y, response variable) * SAT math score (SATM, x 1, explanatory variable) * SAT verbal score (SATV, x 2, explanatory variable) * Average high school grade in math (HSM, x 3, explanatory variable) * Average high school grade in science (HSS, x 4, explanatory variable) * Average high school grade in English (HSE, x 5, explanatory variable) Here are the summary statistics for these data given by software SAS:

11. The first step in multiple linear regression is to study all pair-wise relationships between the p + 1 variables. Here is the SAS output for all pair-wise correlation analyses (value of r and 2 sided p-value of H 0 : ρ = 0). Scatterplots for all 15 pair-wise relationships are also necessary to understand the data.

12. For simplicity, let’s first run a multiple linear regression using only the three high school grade averages: P-value very significant R 2 is fairly small (20%) HSM significant HSS, HSE not

13. The ANOVA for the multiple linear regression using only HSM, HSS, and HSE is very significant  at least one of the regression coefficients is significantly different from zero. But R 2 is fairly small (0.205)  only about 20% of the variation in cumulative GPA can be explained by these high school scores. (Remember, a small p-value does not imply a large effect.) P-value very significant R 2 is fairly small (20%)

14. The tests of hypotheses for each b within the multiple linear regression reach significance for HSM only. We found a significant correlation between HSS and GPA when analyzed by themselves, so why isn’t b HSS significant in the multiple regression equation? HHS and HHM are also significantly correlated suggesting redundancy. HSM significant HSS, HSE not If all 3 high school averages are used, neither HSE or HSS, significantly helps us predict GPA over and above what HSM does all by itself.

15. P-value very significant R 2 is small (20%) HSM significant HSE not We now drop the least significant variable from the previous model: HSS. The conclusions are about the same. But notice that the actual regression coefficients have changed.

16. P-value very significant R 2 is very small (6%) SATM significant SATV not Let’s run a multiple linear regression with the two SAT scores only. The ANOVA test for β SATM and β SATV is very significant  at least one is not zero. R 2 is really small (0.06)  only 6% of GPA variations are explained by these tests. When taken together, only SATM is a significant predictor of GPA (P 0.0007).

17. We finally run a multiple regression model with all the variables together The overall test is significant, but only the average high school math score (HSM) makes a significant contribution in this model to predicting the cumulative GPA. This conclusion applies to computer majors at this large university. P-value very significant R 2 fairly small (21%) HSM significant

19. Warnings  Multiple regression is a dangerous tool. The significance or non- significance of any one predictor in a multiple predictor model is largely a function of what other predictors may or may not be included in the model.  P-values arrived at after a selection process (where we drop seemingly unimportant X’s) don’t mean the same thing as P-values do in earlier parts of the textbook.  When the X’s are highly correlated (with each other) two equations with very different parameters can yield very similar predictions (so interpreting slopes as “if I change X by 1 unit, I will change Y by b units” may not be a good idea.