iT IS HAPPEN WHEN MORE THAN ONE POSSIBLE PREDICTOR VARIABLE.
including more than one independent variable in the regression model, makes us extend the simple linerar regression model to a multiple linear regression model.
Advantages:
Relationship between response variables and several predictors simultaneously.
Disadvantages:
Model building , interpration difficulties due to complexity.
Multiple linear regression with two predictors:
Y=beta0+beta1X1+beta2X2+epsylon
where, Y is the dependent variable.
X1,X2…Xk are predictors(independent variables)
Epsylon is the random error
beta1, beta2, beta0 are unknown regression coefficients
Example=> oil consumption:
Y=oil consumption(per month)
X1=outdoor temperature
X2=size of house(in meter square)
Model:
Y=beta0+beta1X1+beta2X2+epsylon
now beta1 is expected change in Y(oil consiumption) at one unit increase in X1(outdoor temperature), when all other predictors are kept constant, i.e. in this case the size of the house is not changed.
beta1 is estimated with beta1=-27.2 degree C
Assumptions:
The random error term epsylon is normally distributed and has mean zero. i.e. E(epsylon)=0
Epsylon has (unknown) variance sigma epsylon^2. i.e. all random errors have the same variance.
Adjusted R^2
R^2adj=1- SSE/(n-k-1)/SST/(n-1)
As for simple linear regression:
plots of residual against y prime
plots of residuals against xi
normal probability plot of residuals
plots of residuals in observation order
Cook’s distance
Studentized residuals
Standardized residuals
Dffits
Collinearity:
Can only occur for multiple regression.
Predictors explaining the same variation of the response variabl.
Oil consumption continued:
One predictor measuring house size in cm^2 and another predictor in m^2
Variance inflation factor
VIFi=1/1-Ri^2
Condition Index for collinearity:
between 10 and 30=>weak collinearity
between 30 and 100=>moderate
collinearity>100=>strong collinearity
Example of Oil consumption continued:
Assume that we would like to use outdoor temperature X1 and house size X2 as predictors. Additionally, we want to use a third predictor:
X3={1 if extra-thick walls, 0 otherwise
Model:
Y=beta0+beta1X1+beta2X2+beta3X3+epsylon
Model Selection Strategies:
Mldel ranked using R^2, adjusted R^2 or mallow’s Cp
Stepwise selection methods:
Backward, forward, stepwise selection
r^2 Selection
In a data set with 7 possible predictors, there would be 2^7-1=127 possible regression models.
For every model size(k=1,2,…..,p) look at, let say, m models, chosen
Mallow’s Cp:
Large Cp=>biased model
it’s a formula.
where MSEp=mean squared error for a model with p parametes
mean squared error for the full model
n=number of observations
