SIMPLE LINEAR REGRESSION We consider the model for

advertisement
S IMPLE L INEAR R EGRESSION
We consider the model for response variable, Y , as a function of the predictor, X, observed to take the
value x. Specifically we consider the model
Y = β0 + β1 x + ²
where β0 and β1 are the intercept and slope parameters respectively, and ² is a random variable with
expectation zero and variance σ 2 . In this model
E[Y |X = x] = β0 + β1 x.
To estimate the parameters β0 and β1 from data (xi , yi ), i = 1, . . . , n, we use the least-squares criterion,
b and β
b to minimize the sum of squared errors
and choose the values β
0
1
SSE(β0 , β1 ) =
n
X
e2i
=
n
X
i=1
(yi − (β0 + β1 xi ))2
i=1
It can be shown that the parameter estimates depend on the following sample summary statistics:
• Sample mean of x values:
n
1X
xi
x=
n
i=1
• Sample mean of y values:
n
1X
y=
yi
n
i=1
• Sum of Squares SSxx :
SSxx
n
X
=
(xi − x)2
i=1
• Sum of Squares SSxy :
SSxy
n
X
=
(xi − x)(yi − y)
i=1
The least-squares estimates are:
b = SSxy
β
1
SSxx
b =y−β
b x
β
0
1
yielding fitted-values
b +β
b xi
ybi = β
0
1
and residual errors (or residuals)
ebi = yi − ybi .
An estimate of the residual error variance is given by
σ
b2 =
b ,β
b )
SSE(β
0
1
n−2
1
E XAMPLE : B LOOD V ISCOSITY AND PACKED C ELL V OLUME
The following data are measurements of packed cell volume (PCV) and blood viscosity in samples taken
from 32 hospital patients. We wish to model viscosity (y) as a function of PCV (x).
Reference: Begg, C. B. and Hearns, J. B. (1966) Components of Blood Viscosity. The relative contributions
of haematocrit, plasma fibrinogen and other proteins, Clinical Science, 31, 87-92.
Unit
1
2
3
4
5
6
7
8
PCV
x
40.00
40.00
42.50
42.00
45.00
42.00
42.50
47.00
Viscosity
y
3.71
3.78
3.85
3.88
3.98
4.03
4.05
4.14
Unit
9
10
11
12
13
14
15
16
PCV
x
46.75
48.00
46.00
47.00
43.25
45.00
50.00
45.00
Viscosity
y
4.14
4.20
4.20
4.27
4.27
4.37
4.41
4.64
Unit
PCV
x
51.25
50.25
49.00
50.00
50.00
49.00
50.50
51.25
17
18
19
20
21
22
23
24
Viscosity
y
4.68
4.73
4.87
4.94
4.95
4.96
5.02
5.02
Unit
25
26
27
28
29
30
31
32
PCV
x
49.50
56.00
50.00
47.00
53.25
57.00
54.00
54.00
• Sample mean of x values: x = 47.938; sample mean of y values: y = 4.646
• Sums of Squares
SSxx
n
X
=
(xi − x)2 = 615.75
SSxy
i=1
n
X
=
(xi − x)(yi − y) = 75.386
i=1
Thus
b = SSxy = 0.122
β
1
SSxx
The estimate of the residual error variance is
σ
b2 =
b =y−β
b x = −1.223
β
0
1
b ,β
b )
SSE(β
2.721
0
1
=
= 0.091
n−2
30
Blood Viscosity vs PCV: Least−squares fit
^
β1 = 0.122
4.0
4.5
Viscosity
5.0
5.5
^
β0 = − 1.223
40
45
50
PCV
2
55
Viscosity
y
5.12
5.15
5.17
5.18
5.38
5.77
5.90
5.90
Download