Multicollinearity Multicollinearity occurs when explanatory variables are highly

Multicollinearity • Multicollinearity occurs when explanatory variables are highly correlated, in which case, it is difficult or impossible to measure their individual influence on the response. • The fitted regression equation is unstable. • The estimated regression coefficients vary widely from data set to data set (even if data sets are very similar) and depending on which predictor variables are in the model. • The estimated regression coefficients may even have opposite sign than what is expected (e. g, bedroom in house price example). STA 302/1001 - week 11 1

• The regression coefficients may not be statistically significant from 0 even when corresponding explanatory variable is known to have a relationship with the response. • When some X’s are perfectly correlated, we can’t estimate β because X’X is singular. • Even if X’X is close to singular, its determinant will be close to 0 and the standard errors of estimated coefficients will be large. STA 302/1001 - week 11 2

Quantitative Assessment of Multicollinearity • To asses multicolinearity we calculate the Variance Inflation Factor for each of the predictor variables in the model. • The variance inflation factor for the ith predictor variable is defined as where is the coefficient of multiple determination obtained when the ith predictor variable is regressed against p-1 other predictor variables. • Large value of VIFi is a sign of multicollinearity. STA 302/1001 - week 11 3

Rainfall Example • The data set contains cord yield (bushes per acre) and rainfall (inches) in six US corn-producing states (Iowa, Nebraska, Illinois, Indiana, Missouri and Ohio). • Straight line model is not adequate – up to 12″ rainfall yield increases and then starts to decrease. • A better model for this data is a quadratic model: Yield = β 0 + β 1∙rain + β 2∙rain 2 + ε. • This is still a multiple linear regression model since it is linear in the β’s. • However, we can not interpret individual coefficients, since we can’t change one variable while holding the other constant… STA 302/1001 - week 11 4

More on Rainfall Example • Examination of residuals (from quadratic model) versus year showed that perhaps there is a pattern of an increase over time. • Fit a model with year… • To assess whether yield’s relationship with rainfall depends on year we include an interaction term in the model… STA 302/1001 - week 11 5

Interaction • Two predictor variables are said to interact if the effect that one of them has on the response depends on the value of the other. • To include interaction term in a model we simply the have to take the product of the two predictor variables and include the resulting variable in the model and an additional predictor. • Interaction terms should not routinely be added to the model. Why? • We should add interaction terms when the question of interest has to do with interaction or we suspect interaction exists (e. g. , from plot of residuals versus interaction term). • If an interaction term for 2 predictor variables is in the model we should also include terms for predictor variables as well even if their coefficients are not statistically significant different from 0. STA 302/1001 - week 11 6