Você está na página 1de 8

16/05/2016

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

Yi b1 b2 X i

b2

X X Y Y
X X
i

Suppose that the units of measurement of Y or X are changed. How will this affect the
regression results? Intuitively, we would anticipate that nothing of substance will be
changed, and this is correct.

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

Yi b1 b2 X i

b2

X X Y Y
X X
i

We will demonstrate this for the estimates of the regression coefficients in this section, and
we will trace the implications for the rest of the regression output in due course. We begin
by supposing that the true and fitted models are as shown above.

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

Yi b1 b2 X i

b2

Yi* 1 2Yi

X X Y Y
X X
i

Yi* b1* b2* X i

We now suppose that the units of measurement of Y are changed, with the new measure, Y*,
being a linear function of the old one.
3

16/05/2016

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

b2

Yi b1 b2 X i

Yi* 1 2Yi

X X Y Y
X X
i

Yi* b1* b2* X i

Typically, a change of measurement involves a simple multiplicative scaling, such as when


we convert pounds into grams.
4

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

b2

Yi b1 b2 X i

Yi* 1 2Yi

X X Y Y
X X
i

Yi* b1* b2* X i

However, one occasionally encounters a full linear transformation. Conversion of


temperatures from degrees Celsius to degrees Fahrenheit is an example.
5

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

b2

Yi b1 b2 X i

Yi* 1 2Yi

X X Y Y
X X
i

Yi* b1* b2* X i

X X Y Y X X Y Y
X X
X X
X X Y Y X X Y Y b

X X
X X

b2*

2 i

2 i

i
2

2 2

Regressing Y* on X, the new slope coefficient b2* is as shown.


6

16/05/2016

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

b2

Yi b1 b2 X i

Yi* 1 2Yi

X X Y Y
X X
i

Yi* b1* b2* X i

X X Y Y X X Y Y
X X
X X
X X Y Y X X Y Y b

X X
X X

b2*

2 i

2 i

i
2

2 2

We substitute for Y* from the linear relationship defining it.


7

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

b2

Yi b1 b2 X i

Yi* 1 2Yi

X X Y Y
X X
i

Yi* b1* b2* X i

X X Y Y X X Y Y
X X
X X
X X Y Y X X Y Y b

X X
X X

b2*

2 i

2 i

i
2

2 2

The 1 terms cancel. 2 is a common factor in the numerator.


8

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

b2

Yi b1 b2 X i

Yi* 1 2Yi

X X Y Y
X X
i

Yi* b1* b2* X i

X X Y Y X X Y Y
X X
X X
X X Y Y X X Y Y b

X X
X X

b2*

2 i

2 i

i
2

2 2

We find that the new slope coefficient is equal to the original one, multiplied by 2.
9

16/05/2016

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

b2

Yi b1 b2 X i

Yi* 1 2Yi

X X Y Y
X X
i

Yi* b1* b2* X i

X X Y Y X X Y Y
X X
X X
X X Y Y X X Y Y b

X X
X X

b2*

2 i

2 i

i
2

2 2

This is logical. A unit change in Y is the same as a change of 2 units in Y*. According to
the regression equation, a unit change in X leads to a change of b2 units in Y, so it should
lead to a change of 2b2 units in Y*.

10

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

b2

Yi b1 b2 X i

Yi* 1 2Yi

X X Y Y
X X
i

Yi* b1* b2* X i

X X Y Y X X Y Y
X X
X X
X X Y Y X X Y Y b

X X
X X

b2*

2 i

2 i

i
2

2 2

The effect on the intercept will be left as an exercise. The effect of a change in the units of
measurement of X will also be left as an exercise.
11

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

Yi b1 b2 X i

b2

X X Y Y
X X
i

However, we will consider a special case of a change in the measurement of X. Often the
intercept in a regression equation has no sensible interpretation because X = 0 is distant
from the data range.

12

16/05/2016

CHANGES IN THE UNITS OF MEASUREMENT


^
EARNINGS
= 13.93 + 2.46 S

120

Hourly earnings ($)

100
80
60
40
20
0
-20

10 11 12 13 14 15 16 17 18 19 20

Years of schooling (highest grade completed)

The earnings function discussed in the previous slideshow is an example, with the intercept
actually being negative.
13

CHANGES IN THE UNITS OF MEASUREMENT


. reg EARNINGS S
Source |
SS
df
MS
-------------+-----------------------------Model | 19321.5589
1 19321.5589
Residual | 92688.6722
538 172.283777
-------------+-----------------------------Total | 112010.231
539 207.811189

Number of obs
F( 1,
538)
Prob > F
R-squared
Adj R-squared
Root MSE

=
=
=
=
=
=

540
112.15
0.0000
0.1725
0.1710
13.126

-----------------------------------------------------------------------------EARNINGS |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
-------------+---------------------------------------------------------------S |
2.455321
.2318512
10.59
0.000
1.999876
2.910765
_cons | -13.93347
3.219851
-4.33
0.000
-20.25849
-7.608444
------------------------------------------------------------------------------

Here is the output for the regression.


14

CHANGES IN THE UNITS OF MEASUREMENT


. reg EARNINGS S
Source |
SS
df
MS
-------------+-----------------------------Model | 19321.5589
1 19321.5589
Residual | 92688.6722
538 172.283777
-------------+-----------------------------Total | 112010.231
539 207.811189

Number of obs
F( 1,
538)
Prob > F
R-squared
Adj R-squared
Root MSE

=
=
=
=
=
=

540
112.15
0.0000
0.1725
0.1710
13.126

-----------------------------------------------------------------------------EARNINGS |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
-------------+---------------------------------------------------------------S |
2.455321
.2318512
10.59
0.000
1.999876
2.910765
_cons | -13.93347
3.219851
-4.33
0.000
-20.25849
-7.608444
------------------------------------------------------------------------------

The intercept is a meaningless negative number. Literally, it is the predicted level of hourly
earnings for those with zero years of schooling, but nobody in the sample had fewer than
seven years of schooling and only 31 had fewer than twelve years.

15

16/05/2016

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

Yi b1 b2 X i

b2

X i* X i X

X X Y Y
X X
i

Yi b1* b2* X i*

Sometimes it is useful to deal with this problem by defining X* as the deviation of X about
its sample mean.
16

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

Yi b1 b2 X i

b2

X i* X i X

X X Y Y
X X
i

Yi b1* b2* X i*

*
i

X i X
X i nX

nX nX 0
1
X i* 0
n

X*

Note that, by definition, the sample sum of Xi* is 0, and hence the mean value of X* is zero..
17

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

Yi b1 b2 X i

b2

X i* X i X

Yi b1* b2* X i*

X X Y Y
X X
i

X X Y Y X Y
X X
X

X X Y Y b

X X

b2*

*
i

i
* 2

*
i

*
i

*2
i

X* 0

If we regress Y on X* instead of X, the slope coefficient is not affected.


18

16/05/2016

CHANGES IN THE UNITS OF MEASUREMENT

Yi 1 2 X i ui

Yi b1 b2 X i

b2

X i* X i X

Yi b1* b2* X i*

X X Y Y
X X
i

X X Y Y X Y
X X
X
X X Y Y b

X X

b2*

*
i

i
* 2

*
i

*
i

*2
i

X* 0

b Y b X Y
*
1

*
2

The intercept is now the fitted value of Y at the sample mean of X. This is the sample mean
of Y and it may be more informative for analytical purposes.
19

CHANGES IN THE UNITS OF MEASUREMENT


-----------------------------------------------------------------------------. sum S
Variable |
Obs
Mean
Std. Dev.
Min
Max
-------------+-------------------------------------------------------S |
540
13.67222
2.438476
7
20
. gen SDEV = S - 13.67

We will investigate the effect of demeaning schooling int he wage equation example. First,
we find the sample mean. (In Stata, we use the sum command.) We find that the mean is
13.67 years, and we define a new variable SDEV by subtracting 13.67 from S.

20

CHANGES IN THE UNITS OF MEASUREMENT


. reg EARNINGS SDEV
Source |
SS
df
MS
-------------+-----------------------------Model | 19321.5587
1 19321.5587
Residual | 92688.6723
538 172.283778
-------------+-----------------------------Total | 112010.231
539 207.811189

Number of obs
F( 1,
538)
Prob > F
R-squared
Adj R-squared
Root MSE

=
=
=
=
=
=

540
112.15
0.0000
0.1725
0.1710
13.126

-----------------------------------------------------------------------------EARNINGS |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
-------------+---------------------------------------------------------------SDEV |
2.455321
.2318512
10.59
0.000
1.999876
2.910765
_cons |
19.63077
.5648401
34.75
0.000
18.5212
20.74033
------------------------------------------------------------------------------

Here is the output using SDEV instead of S. The intercept, 19.63, now gives the predicted
earnings of those with mean schooling.
21

16/05/2016

CHANGES IN THE UNITS OF MEASUREMENT


. reg EARNINGS SDEV
Source |
SS
df
MS
Number of obs =
540
-------------+-----------------------------F( 1,
538) = 112.15
Model | 19321.5587
1 19321.5587
Prob > F
= 0.0000
Residual | 92688.6723
538 172.283778
R-squared
= 0.1725
-------------+-----------------------------Adj R-squared = 0.1710
Total | 112010.231
539 207.811189
Root MSE
= 13.126
-----------------------------------------------------------------------------EARNINGS |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
SDEV |
2.455321
.2318512
10.59
0.000
1.999876
2.910765
_cons |
19.63077
.5648401
34.75
0.000
18.5212
20.74033
-----------------------------------------------------------------------------. reg EARNINGS S
Source |
SS
df
MS
Number of obs =
540
-------------+-----------------------------F( 1,
538) = 112.15
Model | 19321.5589
1 19321.5589
Prob > F
= 0.0000
Residual | 92688.6722
538 172.283777
R-squared
= 0.1725
-------------+-----------------------------Adj R-squared = 0.1710
Total | 112010.231
539 207.811189
Root MSE
= 13.126
-----------------------------------------------------------------------------EARNINGS |
Coef.
Std. Err.
t
P>|t|
[95% Conf. Interval]
S |
2.455321
.2318512
10.59
0.000
1.999876
2.910765
_cons | -13.93347
3.219851
-4.33
0.000
-20.25849
-7.608444
------------------------------------------------------------------------------

Comparing the new output with the original, we see that, apart from the standard error and t

statistic of the intercept, nothing else has changed.

22

CHANGES IN THE UNITS OF MEASUREMENT


^
EARNINGS
= 19.63 + 2.46 S

120
100

Hourly earnings ($)

80
60
40
20

-14 -13 -12 -11 -10 -9

-8

-7

-6

-5

-4

-3

-2

-1

-20

Years of schooling (deviation from mean)

This is exactly as one would expect. The only effect of the demeaning of S is to move the
vertical axis to the point that was formerly 13.67. As a consequence, the intercept becomes
19.63. Otherwise, the regression line is unaffected.

23

Você também pode gostar