Simple Regression Analysis: April 2011
Simple Regression Analysis: April 2011
Simple Regression Analysis: April 2011
The researcher realizes that GDP has been underestimated by $100 in every country and that N
should have been regressed on G*, where G* = G + 100. Explain, with mathematical proofs,
how the following components of the output would have differed:
the coefficient of GDP
the intercept
RSS
R
2
A1.6 A researcher with the same model and data as in Exercise A1.6 believes that GDP in each
country has been underestimated by 50 percent and that N should have been regressed on G*,
where G* = 2G. Explain, with mathematical proofs, how the following components of the
output would have differed:
the coefficient of GDP
the intercept
RSS
R
2
A1.7 A variable Y
i
is generated as
Y
i
= |
1
+ u
i
(1.1)
where |
1
is a fixed parameter and u
i
is a disturbance term that is independently and identically distributed
with expected value 0 and population variance . The least squares estimator of |
1
is
2
u
o Y , the sample mean
of Y. Give a mathematical demonstration that the value of R
2
in such a regression is zero.
4
Answers to the starred exercises in the text
1.8 The output below shows the result of regressing the weight of the respondent in 1985, measured in
pounds, on his or her height, measured in inches, using EAEF Data Set 21. Provide an interpretation of
the coefficients.
. reg WEIGHT85 HEIGHT
Source | SS df MS Number of obs = 540
-------------+------------------------------ F( 1, 538) = 355.97
Model | 261111.383 1 261111.383 Prob > F = 0.0000
Residual | 394632.365 538 733.517407 R-squared = 0.3982
-------------+------------------------------ Adj R-squared = 0.3971
Total | 655743.748 539 1216.59322 Root MSE = 27.084
------------------------------------------------------------------------------
WEIGHT85 | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
HEIGHT | 5.192973 .275238 18.87 0.000 4.6523 5.733646
_cons | -194.6815 18.6629 -10.43 0.000 -231.3426 -158.0204
------------------------------------------------------------------------------
Answer: Literally the regression implies that, for every extra inch of height, an individual tends to weigh
an extra 5.2 pounds. The intercept, which literally suggests that an individual with no height would
weigh 195 pounds, has no meaning. The figure shows the observations and the fitted regression line.
0
50
100
150
200
250
300
350
40 50 60 70 80
height (inches)
w
e
i
g
h
t
i
n
1
9
8
5
(
p
o
u
n
d
s
)
1.10 A researcher has international cross-sectional data on aggregate wages, W, aggregate profits, P, and
aggregate income, Y, for a sample of n countries. By definition,
Y
i
= W
i
+ P
i
.
The regressions
i i
Y a a W
2 1
+ =
i i
Y b b P
2 1
+ =
are fitted using OLS regression analysis. Show that the regression coefficients will automatically satisfy
the following equations:
a
2
+ b
2
= 1
a
1
+ b
1
= 0.
5
Explain intuitively why this should be so.
Answer:
( )( )
( )
( )( )
( )
( )( )
( )
( )( )
( )
1
2 2
2 2
2 2
=
+
=
= +
Y Y
Y Y Y Y
Y Y
P W P W Y Y
Y Y
P P Y Y
Y Y
W W Y Y
b a
i
i i
i
i i i
i
i i
i
i i
( ) ( ) ( ) ( ) 0
2 2 2 2 1 1
= = + + = + = + Y Y Y b a P W Y b P Y a W b a
The intuitive explanation is that the regressions break down income into predicted wages and profits and
one would expect the sum of the predicted components of income to be equal to its actual level. The sum
of the predicted components is [(a
1
+ a
2
Y) + (b
1
+ b
2
Y)], and in general this will be equal to Y only if the
two conditions are satisfied.
1.12 Suppose that the units of measurement of X are changed so that the new measure,
*
X , is related to the
original one by . Show that the new estimate of the slope coefficient is
i i
X X
2 1
*
+ =
2 2
b , where b
2
is the slope coefficient in the original regression.
Answer:
( )( )
( )
| | | | ( )( )
| | | | ( )
( )( )
( )
( )( )
( )
2
2
1
2
2
2
1
2
1
2
2 2
1
2 2
1
2
2 1 2 1
1
2 1 2 1
1
2
* *
1
* *
*
2
b
X X
Y Y X X
X X
Y Y X X
X X
Y Y X X
X X
Y Y X X
b
n
i
i
n
i
i i
n
i
i
n
i
i i
n
i
i
n
i
i i
n
i
i
n
i
i i
=
=
+ +
+ +
=
=
=
=
=
=
=
=
=
1.13 Demonstrate that if X is demeaned but Y is left in its original units, the intercept in a regression of Y on
demeaned X will be equal to Y .
Answer: Let X X X
i i
=
*
and and be the intercept and slope coefficient in a regression of Y on
*
1
b
*
2
b
*
X . Note that 0
*
= X . Then
Y X b Y b = =
*
2
*
1
The slope coefficient is not affected by demeaning:
( )( )
( )
| | ( )( )
| | ( )
2
1
2
1
1
2
* *
1
* *
*
2
0
0
b
X X
Y Y X X
X X
Y Y X X
b
n
i
i
n
i
i i
n
i
i
n
i
i i
=
=
=
=
=
=
1.14 Derive, with a proof, the slope coefficient that would have been obtained in Exercise 1.5 if weight and
height had been measured in metric units. (Note: one pound is 454 grams and one inch is 2.54 cm.)
Answer: Let the weight and height be W and H in imperial units (pounds and inches) and WM and HM in
metric units (kilos and centimetres). Then WM = 0.454W and HM = 2.54H. The slope coefficient for the
regression in metric units, , is given by
M
b
2
6
( )( )
( )
( ) ( )
( )
( )( )
( )
929 . 0 179 . 0 179 . 0
54 . 2
454 . 0 54 . 2
2
2
2
2
2
2
= =
b
H H
W W H H
H H
W W H H
HM HM
WM WM HM HM
b
i
i i
i
i i
i
i i M
In other words, weight increases at the rate of almost one kilo per centimetre. The regression output
below confirms that the calculations are correct (subject to rounding error in the last digit).
. g WM = 0.454*WEIGHT85
. g HM = 2.54*HEIGHT
. reg WM HM
Source | SS df MS Number of obs = 540
-------------+------------------------------ F( 1, 538) = 355.97
Model | 53819.2324 1 53819.2324 Prob > F = 0.0000
Residual | 81340.044 538 151.189673 R-squared = 0.3982
-------------+------------------------------ Adj R-squared = 0.3971
Total | 135159.276 539 250.759325 Root MSE = 12.296
------------------------------------------------------------------------------
WM | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
HM | .9281928 .0491961 18.87 0.000 .8315528 1.024833
_cons | -88.38539 8.472958 -10.43 0.000 -105.0295 -71.74125
------------------------------------------------------------------------------
1.15 Consider the regression model
i i i
u X Y + + =
2 1
| |
It implies
u X Y + + =
2 1
| |
and hence that
i i i
v X Y + =
*
2
*
|
where Y Y Y
i i
=
*
, X X X
i i
=
*
, and u u v
i i
= .
Demonstrate that a regression of on
*
Y
*
X using (1.40) will yield the same estimate of the slope
coefficient as a regression of Y on X. Note: (1.40) should be used instead of (1.28) because there is no
intercept in this model.
Evaluate the outcome if the slope coefficient were estimated using (1.28), despite the fact that there is
no intercept in the model.
Determine the estimate of the intercept if
*
Y were regressed on
*
X with an intercept included in the
regression specification.
Answer: Let be the slope coefficient in a regression of on
*
2
b
*
Y
*
X using (1.40). Then
( )( )
( )
2
2 2
*
* *
*
2
b
X X
Y Y X X
X
Y X
b
i
i i
i
i i
=
= =
Let be the slope coefficient in a regression of
* *
2
b
*
Y on
*
X using (1.28). Note that
*
Y and
*
X are
both zero. Then
7
( )( )
( )
2
2
*
* *
2
* *
* * * *
* *
2
b
X
Y X
X X
Y Y X X
b
i
i i
i
i i
= =
Let be the intercept in a regression of on
* *
1
b
*
Y
*
X using (1.28). Then
0
* * *
2
* * *
1
= = X b Y b
1.17 Demonstrate that the fitted values of the dependent variable are uncorrelated with the residuals in a simple
regression model. (This result generalizes to the multiple regression case.)
Answer: The numerator of the sample correlation coefficient for Y
\
|
=
= =
n
i
i i
n
i
i i
n
i
i i
e X X b
n
e X b b X b b
n
e e Y Y
n
by (1.53). Hence the correlation is zero.
1.22 Demonstrate that, in a regression with an intercept, a regression of Y on
*
X must have the same R
2
as a
regression of Y on X, where . X X
2 1
*
+ =
Answer: Let the fitted regression of Y on
*
X be written . (Exercise 1.12).
* *
2
*
1
*
i i
X b b Y + =
2 2
*
2
/ b b =
2
2 1
1
2
b
b =
2
1
2
* *
2
*
1
b
X b Y X b Y b = = .
Hence
( )
i i i
Y X
b b
b Y
2 1
2
2
2
2 1
1
*
= + + =
The fitted and actual values of Y are not affected by the transformation and so R
2
is unaffected.
1.24 The output shows the result of regressing weight in 2002 on height, using EAEF Data Set 21. In 2002 the
respondents were aged 3744. Explain why R
2
is lower than in the regression reported in Exercise 1.5.
. reg WEIGHT02 HEIGHT
Source | SS df MS Number of obs = 540
-------------+------------------------------ F( 1, 538) = 216.95
Model | 311260.383 1 311260.383 Prob > F = 0.0000
Residual | 771880.527 538 1434.72217 R-squared = 0.2874
-------------+------------------------------ Adj R-squared = 0.2860
Total | 1083140.91 539 2009.53787 Root MSE = 37.878
------------------------------------------------------------------------------
WEIGHT02 | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
HEIGHT | 5.669766 .3849347 14.73 0.000 4.913606 6.425925
_cons | -199.6832 26.10105 -7.65 0.000 -250.9556 -148.4107
------------------------------------------------------------------------------
8
Answer: The explained sum of squares (described as the model sum of squares in the Stata output) is
actually higher than that in Exercise 1.5. The reason for the fall in R
2
is the huge increase in the total sum
of squares, no doubt caused by the cumulative effect of diversity in eating habits.
Answers to the additional exercises
A1.1 Expenditure on food consumed at home increases by 5.3 cents for each dollar of total household
expenditure. Literally the intercept implies that $1,923 would be spent on food consumed at home if total
household expenditure were zero. Obviously, such an interpretation does not make sense. If the
explanatory variable were income, and household income were zero, positive expenditure on food at
home would still be possible if the household received food stamps or other transfers. But here the
explanatory variable is total household expenditure.
A1.2 Housing has the largest coefficient, followed perhaps surprisingly by food consumed away from home,
and then clothing. All the slope coefficients are highly significant, with the exception of local public
transportation. Its slope coefficient is 0.0008, with t statistic 0.40, indicating that this category of
expenditure is on the verge of being an inferior good.
EXP
n b
2
s.e.(b
2
) R
2
F
FDHO 868 0.0527 0.0027 0.3052 380.4
FDAW 827 0.0440 0.0021 0.3530 450.0
HOUS 867 0.1935 0.0063 0.5239 951.9
TELE 858 0.0101 0.0009 0.1270 124.6
DOM 454 0.0225 0.0043 0.0581 27.9
TEXT 482 0.0049 0.0006 0.1119 60.5
FURN 329 0.0128 0.0023 0.0844 30.1
MAPP 244 0.0089 0.0018 0.0914 24.3
SAPP 467 0.0013 0.0003 0.0493 24.1
CLOT 847 0.0395 0.0018 0.3523 459.5
FOOT 686 0.0034 0.0003 0.1575 127.9
GASO 797 0.0230 0.0014 0.2528 269.0
TRIP 309 0.0240 0.0038 0.1128 39.0
LOCT 172 0.0008 0.0019 0.0009 0.2
HEAL 821 0.0226 0.0029 0.0672 59.0
ENT 824 0.0700 0.0040 0.2742 310.6
FEES 676 0.0306 0.0026 0.1667 134.8
TOYS 592 0.0090 0.0010 0.1143 76.1
READ 764 0.0039 0.0003 0.1799 167.2
EDUC 288 0.0265 0.0054 0.0776 24.1
TOB 368 0.0071 0.0014 0.0706 27.8
A1.3 The summary data indicate that, on average, the respondents put on 25.7 pounds over the period 1985
2002. Was this due to the relatively heavy becoming even heavier, or to a general increase in weight?
The regression output indicates that weight in 2002 was approximately equal to weight in 1985 plus 23.6
pounds, so the second explanation appears to be the correct one. Note that this is an instance where the
constant term can be given a meaningful interpretation and where it is as of much interest as the slope
coefficient. The R
2
indicates that 1985 weight accounts for 68 percent of the variance in 2002 weight, so
other factors are important.
9
A1.4 The slope coefficient indicates that hourly earnings increase by 80 cents for every extra inch of height.
The negative intercept has no possible interpretation. The interpretation of the slope coefficient is
obviously highly implausible, so we know that something must be wrong with the model. The
explanation is that this is a very poorly specified earnings function and that, in particular, we are failing to
control for the sex of the respondent. Later on, in Chapter 5, we will find that males earn more than
females, controlling for observable characteristics. Males also tend to be taller. Hence we find an
apparent positive association between earnings and height in a simple regression. Note that R
2
is very
low.
A1.5 The coefficient of GDP: Let the revised measure of GDP be denoted G*, where G* = G + 100. Since
for all i, 100
*
+ =
i i
G G 100
*
+ = G G and so G G G G
i i
=
* *
for all i. Hence the new slope
coefficient is
( )( )
( )
( )( )
( )
2
2 2
* *
* *
*
2
b
G G
N N G G
G G
N N G G
b
i
i i
i
i i
=
.
The coefficient is unchanged.
The intercept: The new intercept is ( ) 0 . 23 100 100
2 1 2
* *
2
*
1
= = + = = b b G b N G b N b .
RSS: The residual in observation i in the new regression, , is given by
*
i
e
( ) ( )
i i i i i i
e G b b b N G b b N e = + = = 100 100
2 2 1
* *
2
*
1
*
,
the residual in the original regression. Hence RSS is unchanged.
R
2
:
( )
=
2
2
1
N N
RSS
R
i
and is unchanged since RSS and ( )
N N
i
2
are unchanged. Note that this
makes sense intuitively. R
2
is unit-free and so it is not possible for the overall fit of a relationship to be
affected by the units of measurement.
A1.6 The coefficient of GDP: Let the revised measure of GDP be denoted G*, where G* = 2G. Since
for all i,
i i
G G 2
*
= G G 2
*
= and so ( ) G G G G
i i
= 2
* *
for all i. Hence the new slope coefficient is
( )( )
( )
( )( )
( )
( )( )
( )
010 . 0
2
4
2
4
2
2
2
2 2
* *
* *
*
2
= =
b
G G
N N G G
G G
N N G G
G G
N N G G
b
i
i i
i
i i
i
i i
where b
2
= 0.020 is the slope coefficient in the original regression.
The intercept: The new intercept is 0 . 25 2
2
1 2
2 * *
2
*
1
= = = = = b G b N G
b
N G b N b , the original
intercept.
RSS: The residual in observation i in the new regression, , is given by
*
i
e
i i i i i i
e G
b
b N G b b N e = = = 2
2
2
1
* *
2
*
1
*
the residual in the original regression. Hence RSS is unchanged.
10
R
2
:
( )
=
2
2
1
N N
RSS
R
i
and is unchanged since RSS and ( )
N N
i
2
are unchanged. As in Exercise
A1.6, this makes sense intuitively.
A1.7
( )
( )
=
i
i
i
i
Y Y
Y Y
R
2
2
2
and Y
i
=
Y for all i.