Download MULTPLE REGRESSION-I

MULTPLE REGRESSION-I
Regression analysis examines the relation between a single dependent variable Y and one
or more independent variables X1, …,Xp.
SIMPLE LINEAR REGRESSION MODELS
First Order Model with One Predictor Variable
Yi  0  1X i1   i , i=1,2,…,n
 0 is the intercept of the line, and 1 is the slope of
the line. One unite increase in X gives 1 unites
increase in Y.
 i is called a statistical random error for the ith
observation Yi. It accounts for the fact that the
statistical model does not give an exact fit to the
data.
 i cannot be observed. We assume:
o E(i)=0
o Var(i)=2 for all i=1,…,n
o Cov(i, j)=0 for all ij
 Response/Regression function and an example:
EYi   0  1X i1
EYi   9.5  2.1X i1
The response Yi, given the level of X in the ith
trial Xi, comes from a probability distribution
whose mean is 9.5+2.1 Xi. Therefore,
Response/Regression function relates the
means of the probability distributions of Y (for
given X ) to the level of X.
 0, 1, and  are unknown parameters and have to
be estimated from data. We use b0, b1, and e for the
estimates of 0, 1, and .
 Fitted Values and an Example:
Ŷi  b0  b1Xi1
is called fitted values
Ŷi  9.6  1.9Xi1
 Residual and an Example:
ei  Yi  Ŷi
is called the residual value
ei  108  (9.6  1.9 * 45)  12.9
 i  108  104  4
o The residuals are observable, and can be used to check
assumptions on the statistical random errors i.
o Points above the line have positive residuals, and points below the
line have negative residuals
o A line that fits the data well has small residuals.
o We want the residuals to be small in magnitude, because large
negative residuals are as bad as large positive residuals. Therefore,
we cannot simply require  ei  0.
o In fact,
e
i
 0. (derivation on board).
o Two immediate solutions:
| e | to be small
Require  e to be small
- Require
-
i
2
i
o We consider the second option because working with squares is
mathematically easier than working with absolute values (for
example, it is easier to take derivatives).
 Find the pair (b0, b1) that minimizes e (least
square estimates)
o We call SSE= e the Residual Sum of
Squares
o We want to find the pair (b0, b1) that
minimizes
SSE= e = (Y  b  b X )
2
i
2
i
2
i
2
i
0
1
i
o We set the partial derivatives of SSE with
respect to b0, b1 equal to zero:
SSE

0
Normal equation (1) :

 (Y b
i
0
 X (Y b
i
i
i
0
 b1X i )  0
 b1X i )  0
Normal equation (2) :

 (1)(2)(Y b
0
SSE

1
 (X )(2)(Y b
i
i
0
 b1X i )  0
 b1X i )  0
Then the solution is (derivation on board);
b 0  Ŷ  b1X
b1 
 (X  X )(Y  Y )
 (X  X )
i
i
2
i
 Properties of the residuals
o  e  0. since the regression line goes through
the point (X, Y) .
o  X e  0 and  Ŷ e  0  The residuals are
uncorrelated with the independent variables
Xi and with the fitted values Ŷ . (prove it on
board)
i
i i
i i
i
Why  e  0. ,  X e  0 and  Ŷ e  0 ?
In fact, from normal equation (1) and (2), we
can immediately tell  e  0 and  X e  0 . Since
 Ŷ e  (b  b X )e  b e  b  X e , we can easily
know that  Ŷ e  0 ?
i
i i
i i
i
i i
0
1
i
i i
i
0
i
i i
1
i i
o Least square estimates are uniquely defined
as long as the values of the independent
variable are not all identical. In that case the
numerator  (X  X)  0 (draw figure).
 Point Estimation of Error Terms Variance 2
o Single Population: Unbiased sample
variance estimator of the population
variance
2
i
Yi     i , i  1,  , n


  i    ,   i ,  j  0
2
2

s2 
2


Y

Y
 i
 
n 1
E s2  2
o Regression model:Unbiased sample variance
estimator of the population variance
Yi   0  1X i   i , i  1,, n


E i   0,  2  i    2 ,   i ,  j  0
SSE
n2

EMSE  2
MSE 
 Inference in Regression Analysis:
SS XY
b1 
Estimator of 1
SS XX
Eb1  1
Mean of b1
2
2
Variance of b1
 b1 
SS XX
Estimated variance
MSE
s 2 b1 
SS XX
t-Score and 1- CI of 1
t* 
b1  1
~ t n  2 
sb1
1   CI :
b1  t 1   2 ; n  2 sb1
s 2 b1 
MSE
SS XX
 Estimating the Mean Value at Xh:
EY      X
Estimator
Yˆ  b  b X
Mean
E Yˆ   EY 
Variance
 1 X  X  
 Yˆ     

SS
Estimated variance
 n

 1 X  X  
s Yˆ   MSE  

h
0
h
0
1
1
h
h
h
h
2
2
2
h
h
XX
2
2
h
h
 n
SS XX

 Analysis of Variance Approach to Regression
Analysis
SSTOT  SSR  SSE
Yi  Y  Ŷi  Y  Yi  Ŷi
 Y  Y    Ŷ  Y Y  Ŷ 
  Ŷ  Y    Y  Ŷ 
2
2
i
i
i
2
i
2
i
i
Basic Table
Source of
Variation
SS




Regression
SSR   Yi  Y
Error
SSE   Yi  Yi
Total
SSTOT    Yi  Y 
2
2
2
df
MS
E{MS}
1
MSR 
SSR  2   2   X  X  2
1
i
1
n-2
MSE 
SSE
n2
2
n-1
o Test Statistic
o F Distribution
o Numerator
Degrees of Freedom dfR=1
o Denominator
MSR
MSE
MSR
F* 
~ F 1, n  2 
MSE
Tail Probability :
F* 


*
Degrees of Freedom dfR=n-1 P F  F 1   ;1, n  2   
H 0 : 1  0
o Hypothesis:
H 1 : 1  0
For  level of significance
o Decision Rule:
If F *  F 1   ;1, n  2   H 0
If F *  F 1   ;1, n  2   H1
Fitting a regression in SAS
data Toluca;
infile 'C:\stat231B06\ch01ta01.txt';
input lotsize workhrs;
proc reg;
/*least square estimation of regression
coefficient*/
model workhrs=lotsize;
output out=results p=yhat r=residual;/*yhat denotes for fitted values
and residual denotes for residual values*/
run;
proc print data=results;
var yhat residual; /*print the fitted values and residual vales*/
run;
SAS Output
The REG Procedure
Model: MODEL1
Dependent Variable: workhrs
Number of Observations Read
Number of Observations Used
25
25
Analysis of Variance
Source
DF
Sum of
Squares
Mean
Square
Model
Error
Corrected Total
1
23
24
252378
54825
307203
252378
2383.71562
Root MSE
Dependent Mean
Coeff Var
48.82331
312.28000
15.63447
Variable
Intercept
lotsize
DF
1
1
R-Square
Adj R-Sq
Parameter Estimates
Parameter
Standard
Estimate
Error
62.36586
26.17743
3.57020
Obs
1
2
0.34697
yhat
residual
347.982
169.472
51.018
-48.472
F Value
Pr > F
105.88
<.0001
0.8215
0.8138
t Value
2.38
Pr > |t|
0.0259
10.29
<.0001
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
240.876
383.684
312.280
276.578
490.790
347.982
419.386
240.876
205.174
312.280
383.684
133.770
455.088
419.386
169.472
240.876
383.684
455.088
169.472
383.684
205.174
347.982
312.280
-19.876
-7.684
48.720
-52.578
55.210
4.018
-66.386
-83.876
-45.174
-60.280
5.316
-20.770
-20.088
0.614
42.528
27.124
-6.684
-34.088
103.528
84.316
38.826
-5.982
10.720
Extension to multiple regression model
Extension to multiple regression model:
Extension to multiple regression model
Extension to multiple regression model
Extension to multiple regression model
Degrees of freedom: the number of independent components that are needed to calculate the
respective sum of squares.
(1) SSTOT is the sum of n squared components. However, since  ( yi  y)  0 , only n1 components are needed for its calculation. The remaining one can always be
n 1
calculated from y n  y    ( yi  y) . Hence, SSTOT has n-1 degrees of freedom.
i 1
(2) SSE=  ei2 is the sum of n squared components. However, there are p restrictions
among the residuals, coming from the p normal equations
XY  XX b  X' (Y  Xb )  X' e  0 . Hence SSE has n-p degrees of freedom.
p1
p p p1
(3) SSR=  ( ŷi  y)2 is the sum of p squared components since all fitted values Ŷi are
calculated from the same estimated regression line. One degree of freedom is lost
because of one restriction
 (ŷi  y)   (ŷi  yi  yi  y)   (ŷi  yi )   (yi  y)  0 . Hence SSR has p-1
degrees of freedom.
Extension to multiple regression model
Extension to multiple regression model