Transcript Topic 3: Simple Linear Regression

Topic 3: Simple Linear
Regression
Outline
• Simple linear regression model
– Model parameters
– Distribution of error terms
• Estimation of regression parameters
– Method of least squares
– Maximum likelihood
Data for Simple Linear
Regression
•
•
•
•
Observe i=1,2,...,n pairs of variables
Each pair often called a case
Yi = ith response variable
Xi = ith explanatory variable
Simple Linear Regression
Model
•
•
•
•
Yi = b0 + b1Xi + ei
b0 is the intercept
b1 is the slope
ei is a random error term
– E(ei)=0 and s2(ei)=s2
– ei and ej are uncorrelated
Simple Linear Normal Error
Regression Model
•
•
•
•
Yi = b0 + b1Xi + ei
b0 is the intercept
b1 is the slope
ei is a Normally distributed random
error with mean 0 and variance σ2
• ei and ej are uncorrelated → indep
Model Parameters
• β0 : the intercept
• β1 : the slope
• σ2 : the variance of the error term
Features of Both
Regression Models
• Yi = β0 + β1Xi + ei
• E (Yi) = β0 + β1Xi + E(ei) = β0 + β1Xi
• Var(Yi) = 0 + var(ei) = σ2
– Mean of Yi determined by value of Xi
– All possible means fall on a line
– The Yi vary about this line
Features of Normal Error
Regression Model
• Yi = β0 + β1Xi + ei
• If ei is Normally distributed then
Yi is N(β0 + β1Xi , σ2)
(A.36)
• Does not imply the collection of Yi
are Normally distributed
Fitted Regression Equation
and Residuals
• Ŷi = b0 + b1Xi
– b0 is the estimated intercept
– b1 is the estimated slope
• ei : residual for ith case
• ei = Yi – Ŷi = Yi – (b0 + b1Xi)
e82=Y82-Ŷ82
Ŷ82=b0 + b182
X=82
Plot the residuals
Continuation of pisa.sas
Using data set from output statement
proc gplot data=a2;
plot resid*year vref=0;
where lean ne .;
run;
vref=0 adds horizontal line to plot at zero
e82
Least Squares
• Want to find “best” b0 and b1
• Will minimize Σ(Yi – (b0 + b1Xi) )2
• Use calculus: take derivative with
respect to b0 and with respect to b1
and set the two resulting equations
equal to zero and solve for b0 and b1
• See KNNL pgs 16-17
Least Squares Solution
b1
(X  X )(Y  Y )


 (X  X )
i
i
2
i
b 0  Y  b1 X
• These are also maximum likelihood
estimators for Normal error model,
see KNNL pp 30-32
Maximum Likelihood
Yi ~ N   0  1X i , 
2

1  Y   X 
  i 0 1 i
2


2
1
fi 
e
2
L  f1  f 2   f n (likelihood function)
Find  0 and 1 which maximizes L
Estimation of σ2
2
ˆ
 ( Yi  Yi )  e i
2

s 
n2
n2
SSE
 MSE

df E
s
s  Root MSE
2
Source
Model
Error
Corrected Total
dfe
Analysis of Variance
Sum of
Mean
DF Squares Square F Value Pr > F
1
15804
15804 904.12 <.0001
11 192.28571 17.48052
12
15997
Root MSE
Dependent Mean
Coeff Var
4.18097 R-Square
0.9880
693.69231 Adj R-Sq
0.9869
0.60271
s
Standard output from Proc REG
MSE
Properties of Least Squares
Line
• The line always goes through (X, Y)
•  ei   (Yi  (b0  b1 X i ))
  Yi   b0   b1 X i
 nY  nb0  nb1 X  n((Y  b1 X )  b0 )
0
• Other properties on pgs 23-24
Background Reading
• Chapter 1
– 1.6 : Estimation of regression function
– 1.7 : Estimation of error variance
– 1.8 : Normal regression model
• Chapter 2
– 2.1 and 2.2 : inference concerning  ’s
• Appendix A
– A.4, A.5, A.6, and A.7