Multiple Linear Regression Polynomial Regression Monotonic but Non-Linear • The relationship between X and Y may be monotonic but not linear. • The linear model can be tweaked to take this into account by applying a monotonic transformation to Y, X, or both X and Y. • Predicting calories consumed from number of persons present at the meal. R2 = .584 R2 = .814 Calories Log Model Persons Polynomial Regression • A monotonic transformation will not help here. • A polynomial regression will. • Copp, N.H. Animal Behavior, 31, 424-430 • Subjects = containers, each with 100 ladybugs • Containers lighted on one side, dark on the other • Y = number on the lighted side • X = temperature Polynomial Models • Quadratic: • Cubic: Yˆ a b1X b2 X 2 Yˆ a b1X b2 X 2 b3 X 3 • For each additional power of X added to the model, the regression line will have one more bend. Using Copp’s Data • Compute Temp2, Temp3 and Temp4. • Conduct a sequential multiple regression analysis, entering Temp first, then Temp2, then Temp3, and then Temp4. • At each step, evaluate whether or not the last entered predictor should be retained. R2 Linear = .137 Quadratic = .601 The Quadratic Model • The quadratic model clearly fits the data better than does the linear model. • Phototaxis is positive as temps rise to about 18 and negative thereafter. Lighted 33.7 3.37 Temp .091 Temp 2 A Cubic Model Lighted 27.6 6.64 Temp .369 Temp 2 .006 * Temp3 • R2 has increased significantly, from .601 to .753, p < .001 • Does an increase of 15.2% of the variance justify making the model more complex? • I think so. Interpretation • Ladybugs buried in leaf mold in Winter head up, towards light, as temperatures warm. • With warming beyond 12, head for some shade – the aphids are in the shade under Karl’s tomato plant leaves. • With warming beyond 32, this place is too hot, lets get out of here. A Quartic Model Lighted 29.0 4.95 Temp .072 Temp 2 .010 * Temp .00024 Temp 3 4 • R2 =.029, p = .030 • Does this small increase in R2 justify making the model more complex? • Can you make sense of a third bend in the curve. The quartic plot does not look much different than the cubic. Multicollinearity • May be a problem whenever you have products or powers of predictors in the model. • Center the predictor variables, • Or simply standardize all variables to mean 0, standard deviation 1. • For complete SPSS output, go here • Polynomial regression can also be used to conduct ANOVA.