Tests of Equality Between Sets of Coefficients in Two Linear Regressions

Author(s): Gregory C. Chow

Source: Econometrica, Vol. 28, No. 3 (Jul., 1960), pp. 591-605
Published by: The Econometric Society
Stable URL: https://www.jstor.org/stable/1910133
Accessed: 07-03-2019 07:20 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access
to Econometrica

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
 Econometrica, Vol. 28, 3 (July 1960)

TESTS OF EQUALITY BETWEEN SETS OF COEFFICIENTS IN TWO

LINEAR REGRESSIONS1

BY GREGORY C. CHOW

Having estimated a linear regression with p coefficients, one may wish to

test whether m additional observations belong to the same regression. This
paper presents systematically the tests involved, relates the prediction in-
terval (for m = 1) and the analysis of covariance (for m > p) within the
framework of general linear hypothesis (for any m), and extends the results
to testing the equality between subsets of coefficients.

1. INTRODUCTION

THE MODEL of normal linear regression has often been widely applied to the
measurement of economic relationships. In studies of the consumption
function, the mean of consumption is assumed to be a linear function of
income and other variables. In studies of consumer demand, the quantity
of a commodity is regressed linearly on its price, income, and perhaps the
price of an important complement or substitute. In studies of business
investment, linear regressions on profits, sales, liquid asset holdings, and
the interest rate, have been estimated. Other notable examples include
empirical studies of dividend policy, of prices of corporate stocks, and
of cost and supply functions.
When a linear regression is used to represent an economic relationship,
the question often arises as to whether the relationship remains stable in
two periods of time, or whether the same relationship holds for two different
groups of economic units. Is the consumption pattern of the American
people today the same as it was before World War II? Do the firms in the
steel industry and the firms in the chemical industry have similar dividend
policies? Statistically these questions can be answered by testing whether
two sets of observations can be regarded as belonging to the same regression
model.
Often there is no economic rationale in assuming that two relationships
are completely the same. It may be more reasonable to suppose that only
parts of the relationships are identical in two periods, or for two groups.
Maybe the price elasticity of demand for a certain food product has not
changed since World War II, while the income elasticity has changed.
Maybe the investments of two groups of firms are affected in the same
manner by profits, but not by liquid assets. Statistically, we are asking
whether subsets of coefficients in two regressions are equal.

1 An early draft of this paper has been revised after helpful comments from William
Kruskal, Edwin Kuh, and David L. Wallace, to all of whom I am grateful.
591

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
592 GREGORY C. CHOW

To state our problems more formally, let y be the dependent variable,

and xi, X2,. . ., xp be the explanatory variables. Assume that there is a
sample of n observations. These observations are governed by a model of
normal linear regression. In matrix notations, the model is:

Yi -XllXX2 ... Xlp #1 81

Y2 X21X22 ... X2p /32 82

_Yn_ XnlXn2 ... * Xnp _ _8n

Here the x's are p fixed variates. The fl's ar

is the intercept if xi is set identically equal to one. The 8's are independent
and normally distributed, each with mean zero and standard deviation C.
Assuming n > p and nonsingularity of the X matrix, we can estimate the
parameters 3i4, /2, . .., f3p and a. Our problems are the testing of wheth
additional observations are from the same regression as the first sample of n
observations, and the testing of whether subsets of coefficients in the two
regressions are identical. The present paper is devoted to a systematic and
unified treatment of these tests.
To test the hypothesis that both samples belong to the same regression,
the well-known prediction interval [8] can be used when the number m of
observations in the second sample equals one, and the analysis of covariance
[7] can be used when m > p. We will present two tests for the case 2 < m
< p. The first test, to be presented in Section 2, is based on a prediction
interval for the mean of m additional observations. The second test is an
F test, to be developed in Section 3. The relationship among this F test, the
prediction interval, and the analysis of covariance will be explained in
Section 4. In Section 5, our results will be extended to testing the equality
between subsets of regression coefficients in the two regressions. Two
examples of econometric applications are given in Section 6. These examples
are concerned with the temporal stability of a statistical demand function
for automobile ownership, and of a statistical demand function for new
automobiles. It was through these examples that I became interested
in the tests presented in this paper.

2. PREDICTION INTERVAL FOR THE MEAN OF M ADDITIONAL OBSERVATIONS

It is straightforward to extend the prediction interval idea from one

observation to the arithmetic mean of m observations.
First, let us rewrite the model in Section 1 briefly as

(1) y Xli + l

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
 TESTS OF EQUALITY 593

where both yi and 81 are column vectors with n elements, X1 is a nonsingular

n by p matrix, and f3i is the column vector of the p regression coefficients.
The subscript 1 denotes the first sample of n observations. The least-
squares estimator of f3i from this first sample is given by:

(2) l= (XiXi)-1 Xi yi = 1 + (XXl)- Xi 81

XjXl is the cross-product matrix of the p x's from the first sample.
Let the m additional observations Y2 of the dependent variable be speci-
fied by the model

(3) y2 =X22 + -822

X2 is a nonsingular m by p matrix, with its m rows representing the m

new observations on the p explanatory variables. 82 is normally distributed
with the covariance matrix 12. If we form the difference between the
vector Y2 and the vector of predictions based on the regression estimated
by the first n observations, we have, incorporating the relations (2) and (3),

(4) d = Y2 - X2b X2-2 X241 + 82 - X2(XlXl)-1 Xj8i1.

The expectation of d is

(5) E(d) X22 -X2#1

Because of the independence of 82 and 81, the covariance matrix of d be-

comes

Cov (d) CoV (82) + CoV [X2(XlXl)-1 X18i]

(6) -= J2 + X2(XiXl)-1 Xl'(Cov 81) Xi(XiXl)-1 X2

= [I +X2(XiX )1X2]a2.

In the special case when m = 1, both Y2 and d become scalars, and X2

becomes a row vector. From (6), the variance of d in this special case will be

(7) Var (d) = [1 + X2(XlXl)-1 X]a2

a2 can be estimated by s2, the (unbiased) square of t

the first n observations. Under the null hypot
expectation of d given in (5) will be zero, and the ratio

(8) ~~~~~~~d2
(8) [1 ? X2(X'Xl)-1 X2] s1

will be distributed as F(1, n - p). This test, which is based on the prediction
interval for one new observation, can be found in [8].

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
594 GREGORY C. CHOW

When we have m new observations and thus m differences dl, d2, ..., dm,
we may consider the average.
- I m
(9) d - d
i=1

Given the covariance matrix of d in (6) above, the variance of d is

(10) Var(d) 62 Var [Edi]= 2 ([I... 1] [I+ X2(XiX1)- 2X] []}

Similarly, under the null hypothesis that P2 =1,

K]dI22
{ 1 1] [I + X2(Xl"Xl)-' X2"] []}

will be distributed as F(1, n -p).

The use of a d test can be found in reference [4], although the formula
given in the Appendix of this reference is incorrect.2 There is little rationale
in using d for the purpose of testing. The test is obviously weak against a
number of alternative hypothesis. One can envisage many situations in
which d is small, not because the new' m observations have come from the
same regression, but because their deviations cancel out. The usefulness of
deriving the distribution of d lies probably more in the construction of pre-
diction intervals for the mean of additional observations-in so far as the
mean is of interest.

3. USE OF F RATIO FOR TESTING THAT E(d) IS A ZERO VECTOR

Instead of changing the null hypothesis P2 = 1 = # to

E(d) = 0, consider the quadratic form d' (Cov d)1d. It follows from(4) and (6)
that

(1 2) d'(Cov d) -1 d= [2X2 - 4X2] [I+X2(XiXi)-1 X2]-1 [X2#2 - X241] + ?

[eiE2] 1-X(XiXl) -1 X.] [I +X2(XlX) -1X2] -1[-X2(XiXl) -1X I] 8 ] -2

The last term is a quadratic form in with rank n-note that [I + X2

(X'X,)-1 X' -1 is m by m. It will be equal to d'(Cov d)'- d under the null
hypothesis that P2 = Pi, as can easily be seen from (12). Therefore under
the null hypothesis, d'(Cov d)-ld will follow X2(m) distribution; whereas
under the alternative hypothesis P2 = P2, d'(Cov d) -ld will follow a non-
central x2 distribution.

2 I am indebted to Robert Solow for pointing out this reference and the errors
therein.

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
 TESTS OF EQUALITY 595

It is well known that the square of the standard error from the first
2
regression, SI, times (n - p)/a2, follows X2(- p). This X2(n p) is inde-
2

pendent of d = Y2- X2b1. si is independent of bi and is certainly indepen-

dent of Y2. Therefore, under the null hypothesis, the ratio

d'(Cov d)-' d - d'[I + X2(XjXi)-l X2Q-1 d

(13) s (-? 1 sm
~2 S(pM
0r2 '(n -f

will follow F(m, n -). Since the numerator of (13) will have a non-central
x2 distribution when P2 # fl, the upper-tail F test can appropriately be
used. Clearly the test (13) reduces to the prediction interval (8) when m = 1.

4. RELATIONSHIPS OF PREDICTION INTERVAL AND ANALYSIS OF COVARIANCE

TO THEORY OF LINEAR HYPOTHESES

This section shows the relationships among the F test of (1

tion interval for one additional observation, and the analysi
(for m > p). All three methods are special applications of the theory of
testing general linear hypotheses. It will therefore be convenient to summa-
rize first the theory of linear hypotheses as applied to testing the homogeneity
of the (entire) sets of coefficients in two regressions. The size m of the second
sample will first be assumed to be larger than j, and then reduced to one.
In our context, the model of general linear hypotheses takes the form3
(14) yi - XluPl + 0 l2 + 81
y2 - ? O 1 +X242 + 82
or

[y2] [? 0] [0] + 82]

Under the-null hypothesis (Ho: ,u = P2 = fi), the model becomes

(15) [+] [X] +

The sum of squares of the residuals under Ho will be shown to equal the
sum of squares of residuals under the alternative hypothesis (Ha: Pi # f2)
plus the sum of squares of the deviations between the two sets of estimates
of y under these two hypotheses. The ratio between the latter two sums,
adjusted for their numbers of degrees of freedom, will be shown to follow
an F distribution if the null hypothesis is true.
3 The developments here follow, and are special applications of, Kempthorne [6].

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
596 GREGORY C. CHOW

If the null hypothesis is true, the least-squares (also maximum likelihood)

estimator of fi, denoted by bo, is

(16) bo [(X' X2) X)]- [X X2] [Yi

' ' -1 ' t rYll, '2-12 1 2 Y2
= [XXIl + X2X2]l [X X'] [Y] - + [X/X + XX2

The residuals from this regression are:

(17) [Yi]- [Xi] bo - [X2X] ? [81] [X2] A

[X2] [XiXl + X2X2]-1 [X X2] [ 81]

[I -(Xi) (XIXi + XIX2)l (X' X')] [81]

The sum of squares of the residuals under Ho can be written as

(18) - (Xl) bo 2b=o (X2) b] [Q'l) -(X) b]

[8j82] [I (X1) (XXIl + X'X2) -1 (X'X')] [81]

Since these residuals are from a regression of n + m observations on p

explanatory variables, the quadratic form (18) in the 8's has rank n + M- p.4
If the alternative hypothesis (Ha: l =A # 2) is true, we are back to the
model (14), and the least-squares estimators of flu and fl2 are

(1)_l xX - x 1 rYll r(XiX1) -1 XI

(b21] 0 X2X21 [o X2l [y2j L(X2X2)-1 X2 Y2]

The residuals under Ha will be

(20) rYl -Xb

Ly2 -[IX2b2]
- XL[I(XiX1) -1 Xi]
- X2 (x2X2)-1 X'2] 1
82]

Similarly, the sum of squares of these residuals will be

y,- X1b1 2
(21) y2-X2b2 = jJyu-Xubujj2 + IjY2-X2b2j12

= -I X1(XlX1)-1 Xl]81 + 8' [I - X2(X2X2)-1 X'] 82

Since the last two quadratic forms have ranks n - p and m - p respectively
and since 81 and 82 are independent, the rank of the quadratic form (21)
willben +m- 2p.

4 For a proof of this, see Kempthorne [6].

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
 TESTS OF EQUALITY 597

Now the sum of squares (18) under Ho will be decomposed into the sum
of squares (21) under Ha plus the sum of squares of the differences

[Xibi - X1bo] and [X2b2 - X2bo]

First start from the identity

(22) [yi- XiNo] y1-Xib2 + [X2b -X2bo]

Summing the squares of the elements on both sides of (22) gives

yi- Xibo02 yy -Xibi 2 X1b1- X1bo2

(23) Y2 - X2bo Y2 - X2b2 ? X2b2 - X2bo
because the cross-product term on the right side of (23) can easily be seen
to be zero. To economize space, (23) will also be written as

(24) Ql =Q2 +Q3 .

We will proceed to show that the rank of the quadratic form Q3 can at
most be p. From (16) and (19), it follows that

(25) [X/X1l + X'X2]bo = Xiyl + Xy2 XXbXib + X2X2b2

which implies

(26) b2 -b(X2X2)-1 XiX1 (b1 - bo)

Substituting (26) into Q3, we have

(27)QX(bb)2
(27) Q3 X2(X'2X2) -1 XX

- [bi - bo] [Xi -XXl (XX2)

(27) is a quadratic form in b1 - b

than p. But b1 - bo is a linear transformation of the 8's, as can be shown
from (2) and (16):

(28) b1-bo = - f-F +{[(X'X1) -1 Xi 0]-[X'X1 + X'X2]-1 [Xi X2] } [e]

Under the null hypothesis Pl = P2 = fi, Q3 will thus be a quadratic form

in the 8's with a maximum rank of p. From (28), we also see that Q3 will tend
to be larger when the null hypothesis is not true.
It has already been observed that the rank of Q2 is m + n - 2p. Since
the rank of Qi is smaller than or equal to the rank of Q2 plus the rank of Q3,
the rank of Q3 must be p. Under the null hypothesis Q2 and Q3 will be
distributed independently as X2(M - n - 2p)a2 and X2(p)02. While the

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
598 GREGORY C. CHOW

distribution of Q3 is affected if Ho does not hold, Q2 will have the same

distribution regardless. We thus can test Ho by the F ratio

(29) F(p, m + n 2p)- Q31p

_ jjX1bj - X1boJ2 + IjIX2b2 - X2bo0j2 (m + n -2p)
1I yi- Xibi 112 + 11 y2-X2b2 J12 p

(29) is the standard analysis-of-covariance test when mi > p.5

A few remarks will suffice to indicate the application of the theory of
linear hypotheses to the case m < p. Let us rewrite (23) as

y1-Xbo 2
(30) Yl-X = IIy-XibiII2 +? IXibi -XiboII2 + IIY2
Y2 -X2b50
+ IIX2b2 X2boII2

Our models for Ho and Ha are (15) and (14), as before. The sum of squares
under HO is clearly Qi whether m > p or m <? p. The sum of squares under
Ha will become yi - X1b1 l2 when m < p-this can be seen either by
evaluating the sum of squares of the residuals from regression (14) or by
noting that the residuals from the second sample will simply be zero.

Regardless of the size m, ly - X1b]1 2 will be distributed as X2

and will be independent of the sum of the other three terms on the right side
of (30). The sum of these three terms equals

IIXibi-XiboII2 + IIY2-X2boII2

even if b2 is undefined. When m < p, we can test Ho by the ratio

(31) F(m, n - p) - I IX1bi-X1boI 12 + I1Y2-X2boII2 (n-P)

Ily" - b,X 1 12 m
When m > p, (31) remains valid. However, using (31) instead of (29) in
this situation would amount to taking a part of Ql, i.e., ||Y2 - X2b22,
which is not affected by the inequality between f1 and P2, and placing it
in the numerator of the F ratio. This would reduce the power of the test.
The theory of linear hypotheses has now been applied to testing the homo-
geneity of two regressions. To provide a link between the analysis of covarian-
ce (29) and the prediction interval (8), we will point out that the test (13)
in Section 3, including its special case (8), is identical with the test (31). The
proof of this identity requires only the proof that

(32) d'[I + X2(X1Xl)-1 X2]-1 d = IIXib, - Xbo 112 + I1Y2 - X2boII2.

5 Additional references on the analysis of covariance include [1], [5], [9], and [10].
[1] is a special issue devoted to the analysis of covariance mainly for the design of
experiments.

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
 TESTS OF EQUALITY 599

From (25), we deduce

(33) b= [I + (XiXXl)-' X2X2]bo (X'X1)-1 XIy2.

Substitute (33) into d:

(34) d Y2 -X2b= [I + X2(X'X1)- X2] [Y2- X2bo].

Given (34), we evaluate the quadratic form

(35) d'[I + X2(X'X1) -' X']-' d [y2- boX2] [I + X2(X'Xl) -' X2] [y2- X2

-[Y2 b0X2] [Y2 -X2b] ? [y2 - b'X'X2] (X'Xl)-' [X'y2 - XX2b0].

Our proof of (32) will be complete by observing the relationship, based on

(25), that

(36) XIy2 - X'X2bo=- [XX1b1 - X`Xibo]

5. TESTS OF EQUALIT.Y BETWEEN SUBSETS OF COEFFICIENTS IN TWO REGRESSIONS

The results given so far, as summarized- by (29) and (31), will now be
extended to testing the equality between subsets of coefficients in two
regressions. As before, we will first examine the case m > /.
Under the alternative hypothesis, our model is

Y= X1fl + 81 Zlyl + Wi3i + 81,

(37)
Y2 = X242 + 82 Z2y2 + W262 + 82,
or

[][2]
= ?[?Wi 0 Y2 +r:8i
Z2 0 W2 [i 82
(32

where the coefficients flu are divided into y, and b(i, the matrix X1 is cor-
respondingly divided into Z, and W1, and similarly for f2 and X2. Let yi
and y2 be column vectors of q elements each; and &1 and 62 be column
vectors of q q elements each. The subsets of coefficients to be tested
are y, and y2-
The null hypothesis is y, = y2 = y, implying the model

(38) [y] [ZiW1 W2] [0 ] + r:ii

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
 600 GREGORY C. CHOW

Under the null hypothesis, the least-squares estimators of the coefficients

are

-CO z;z1 + Z2Z2 Z'w1 zW 2] -1 z1 ' 2z

(39) dio WiZi WiWi 0 Wi l [
rio] I WI y2]
Ld20 - WZ2 0 W'W2J O WLO
The sum of squares of the residuals under H0, analogous to (18), will be

[ _ z1 Wi 0 (Z1 + Z2 Z1W1 Z2W2\iI ZI Z2 1 E

(40) [s'e'] [I -( 0 W2Zi W'Wi )0 K ' W fI [02]
W2'Z2 0 W'2W2W
with m + n - 2- + q degrees of freedom.
Under the alternative hypothesis yl #A y2, the least-squares estimators
are

-l Zi'zizl 0 ZWi 0 -i Z, -
(41) C2 _ Z2Z2 ? Z2'W2 ? Z2
( ) di Wi'Zi ? W1iF ? Wi 0 L21
_d2_ L O W2Z2 0 W2'W2_ _LO W2'_

The sum of squares of the residuals under Ha, which is identical with (21),
will have m + n - 2p degrees of freedom.
As before, the sum of squares under H, can be broken up into the sum
of squares under Ha plus the sum of squares of the differences between the
two sets of estimates of y, namely,

_ll\ Zi Wi 0 C6o\0112 ,// Ci 2

(42) / i )d~io yi\ 0 Wi C
kY) \Z2 ?to0 W2 Kd
? Yi2 W 0 (ZY i W2(2 d2
11Zi 0 W, 0 ) C20
+ 2 10~~
+ Z1? ? W2/ dly Z 0 W2)dJ

or

Q1 Q2 + Q3*

We will omit the proof that each of the cross-products on

(42) is zero, but will indicate following identity which may be used in the
proof:

(43) [Z Wi w Lo 0 Wi 1 I
( ) 22 ? W2] 0 Z2 ? W2] 0 I O
LO O Ii

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
 TESTS OF EQUALITY 601

Given that the ranks of Q* and Q

m + n - 2p, the rank of Q3* must
most q. To show the maximum rank o
of the regression of y, on Wl, Zl. as
for Y2. and Z2.. Then it can be proved

(44) [Cl (ZVI.Zl.)-lZloylj

LC2] (Zt-Z2-)- Z2Yy]
and

45) rdl [(WiWi)-lWiyil 1(WWi1)

Id2] [(W2W2)-1W2y2 J [(TW'W2)-1WWZ2c2

Rather than digressing to complete the proofs of (44) and (45), we simply
indicate that essentially they involve partitioning the matrix of the cross-
products of the explanatory variables in (41) into four blocks and then
inverting the partitioned matrix. The same method will also prove

(46) CO = [Vi. Zl. + Z2. Z2-1-1 [Z1. Z2.] [Y2]

and

(47) -dl i (Wiwi)-iW{y 1 ] (W'Wi)-AWiZico]

( ) Ld20 L(W2W2)-1W2Y21 L(W2'W2)-1W2,f2CO -
Using (45) and (47), we can rewrite the vector of the differences between
the estimates of y under Ha and under H,:
_Cj

0 W, 0 pC2 1 Wi 0 CO
(48) [dZ 0 Wi ] [il] - 0 wW2 CO
0 220 W2 di 0 W2 d20

I I-O W2('2W2)-1W2[
wff'wi)-l W 0 -Z
Z2(C2 -CO))
The developments from now on will correspond to the developments of (25),
(26), (27), and (28) in Section 4. From (44) and (46), it follows that

(49)
(49 [Z1. 1. [Z'. Z= Z1
2 22.1 C0 +z yl. zl
4 Z2. 1
Y2= 'Z'ZC1
+ '?l Z1
Z 2 Z'
22. C2
which corresponds to (25). (49) can be used to express C2 - cO as a linear
tranformation of cl - cO, a transformation similar to the one in (26). Replacing
C2 - CO in (48) by this transformation of cl - cO, we will observe that Q3 is a
quadratic form in cl -cO, with maximum rank q. A further step analogous
to (28) will show that Q* tends to be larger when y, # y2. We thereforehav

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
602 GREGORY C. CHOW

(50) FF(q, m + n-2p) Q2/(m ? n-2p)

lZici ? Widi-Zico-Widiol 12 + I IZ2c2 + W2d2-Z2cO

Il1y- Z,c1 - WidiII2 + 1IY2- Z2c2- W2d2II2
x (m + n - 2p)
q

The remaining case for examination is (p - q) < m < p. As long as

m > (p - q), least-squares estimators under Ho can be obtained by (39),
but not when m < (p - q) because W2W2 will then be singular. The sum
of squares of the residuals under Ho will still be Qr. The sum of squares of
the residuals under Ha will be |Jyi - Z,c, - W,di|2 with n p degrees
of freedom. Corresponding to (31), we have

(51) F(m -+q, )

IZic, + Widi- Zico- Wid,o112 + IY2 -Z2co- W2d20o12 (%-p)
IIy,- Z,ct- Wid,II2 (m-p + q)
' The results of this paper, which are really contained
can be summarized briefly. To test the equality between se
in two linear regressions, we obtain the sum of squares
assuming the equality, and the sum of squares without assuming the
equality. The ratio of the difference between these two sums to the latter
sum, adjusted for the corresponding degrees of freedom, will be distributed as
the F ratio under the null hypothesis. This latter sum of squares will be com-
puted only from the first sample of n observations when the second sample
is not large enough for computing a separate regression. We have attempted
to show how the theory of general linear hypotheses is applied to our prob-
lem and how the prediction interval and the analysis of covariance are
related to each other and to the theory of general linear hypotheses. While
we have dealt with the comparison of coefficients in only two regressions,
the proofs of (29) and (50) can obviously be generalized to the case of many
regressions.

6. EXAMPLES

To illustrate how some of the tests given above are applied, one
example for each of (29) and (31) will now be provided. These
originated in a study of the demand for automobiles in the Un
[2]. We will not here go into the economic justifications of them
contained in the reference cited. The study utilizes annual ob
on the following variables:

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
 TESTS OF EQUALITY 603

Xt, ownership of automobiles measured in "new-car - equivalents" per

capita at the end of year t. The unit is per cent of a new-car equivalent per
capita.
Xt, purchase of new cars during year t, with the same unit of measurement
as above.
Pt, relative price index of automobile stock, with 1937 as 100.
Idt, real disposable income per capita in 1937 dollars.
Iet, real "expected" income per capita in 1937 dollars used by Milton
Friedman in his A Theory of the Consumption Fugnction (Princeton: Princeton
University Press, 1957).
A statistical demand function for automobile ownership computed
from observations of 33 years from 1921 to 1953 is

(XI=e) -.7247 - .048802 Pt + .025487 Iet, R2 = .895,

(.004201) (.001747) s =.618.
A statistical demand function for new purchase computed from observations
of 28 years, from 1921 to 1953 but excluding 1942 to 1946, is

=1 2 .07791 - .020127 Pt + .0 11699 Idt - .23104 Xt-1, R2.8

(.002648) (.001070) (.04719) s = .308.
Four years after the study had been made, four additional observations
were available for testing whether these demand functions remained
stable over time. Since four observations are sufficient for computing a
separate regression of the form (Xl e), test (29) was used. To determine
the stability of (4s), test (31) was used. The follow-up study is described
more thoroughly in [3]. Before presenting the analysis of covariance (29)
for the demand function (Xl e), we exhibit here the estimated values of the
dependent variable together with the deviations of the observed values
from the estimated values.
1954 1955 1956 1957
Xt estimated from (XI e) 12.665 12.993 13.328 13.025
Xt observed minus estimated -.613 .079 .102 .437

The residuals of the observed values from the estimated values are very
small, as compared with the standard error of .618. They do not indicate
any shifts in the pattern of demand for automobile ownership during the
four years 1954 to 1957.
We will now proceed with the analysis of covariance (29). The method
involved can be described very simply. Suppose that n observations are
used to estimate a regression with P parameters (p - 1 coefficients plus
one intercept). Suppose also that there are m additional observations, and
we are interested in deciding whether they are generated by the same
regression model as the first n observations. To perform the analysis of
covariance, we need the following sums of squares:

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
 604 GREGORY C. CHOW

A, sum of squares of n + m deviations of the dependent variable from the

regression estimated by n + m observations, with n + m - p degrees of
freedom.
B, sum of squares of n deviations of the dependent variable from the regres-
sion estimated by the first n observations, with n - p degrees of freedom.
C, sum of squares of m deviations of the dependent variable from the
regression estimated by the second m observations, with m - p degrees of
freedom.

From (29), the ratio of (A - B - C)/p to (B + C)/(n + m - 2p) will be

distributed as F(p, n + m - 2p) under the null hypothesis that both groups
of observations belong to the same regression model. For testing the demand
function (Xle), the sum of squares A is 10.1155, and B + C is 9.6130. The
ratio F(3,26) is therefore 0.45. In order to interpret the new observations
as coming from a different structure at the 5 per cent level of significance,
F would have to be at least 2.98. Our impression from examining the four
deviations that there was no change in structure is strongly confirmed.
The following is a comparison of the estimated and the observed values
of the dependent variable of demand function (4s).

1954 1955 1956 1957

Xl estimated from (4s) 3.452 3.730 3.630 3.270

Xl observed minus estimated -.044 .608 -.087 .226

Again, inspection of the residuals reveals that they are not large relative to
0.308, the standard error of estimate for (4s). The year 1955 is an exception,
where we find the residual to be twice as large as the standard error. To
apply test (31), we compute the sum of squares A of the 32 deviations from
the regression including the four new observations, with 32-4 or 28 degrees
of freedom. A turns out to be 2.6444 numerically. The sum of squares B
of the 28 deviations from the regression of the original set of observations
turns out to be 2.2818, with 24 degrees of freedom. The sum of squares C
vanishes as long as the number m of new observations does not exceed the
number of parameters p. According to (31), the F ratio is the ratio of
(A - B)/4 to B/24, or 0.95 numerically. Therefore, we accept the null
hypothesis that automobile purchases in the years 1954 to 1957 were govern-
ed by the same relationship as before.

Cornell University

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms
 TESTS OF EQUALITY 605

REFERENCES

[1] Biometrics, Vol. XIII, No. 3, September, 1957.

[2] Chow, Gregory C.: Demand for Automobiles in the United States, (Amsterdam:
North-Holland Publishing Co., 1957).
[3] : "Statistical Demand Functions for Automobiles and Their Use for
Forecasting," pp. 147-78, The Demand for Durable Goods, ed. by A. C. Harber
(Chicago: University of Chicago Press, 1960).
[4] DAVIS, TOM E.: "The Consumption Function as a Tool of Prediction," The
Review of Economics and Statistics, Vol. XXXIV, No. 3, August, 1952, p. 270.
[5] FRIEDMAN, MILTON.: "Testing the Significance Among a Group of Regression
Equations," Econometrica, Vol. V (1937), p. 194-5.
[6] KEMPTHORNE, OSCAR: The Design and Analysis of Experiments, (New Yor
John Wiley & Sons, Inc., 1952), pp. 54-66.
[7] KENDALL, MAURICE G.: The Advanced Theory of Statistics, (London: Charles
Griffin and Company, Limited, 1946), Vol. II, pp. 242 ff.
[8] MOOD, ALEXANDER M.: Introduction to the Theory of Statistics, (New York:
McGraw-Hill Book Company, Inc., 1950), pp. 304-5.
[9] WALLIS, W. A.: "The Temporal Stability of Consumption Patterns," The Review
of Economic Statistics, Vol. XXIV (1942), pp. 177-83.
[10] WELCH, B. L.: ,,Some Problems in the Analysis of Regression Among k Samples
of Two Variables," Biometrika, Vol. XXVII (1935), pp. 145-60.

This content downloaded from 213.219.153.168 on Thu, 07 Mar 2019 07:20:22 UTC
All use subject to https://about.jstor.org/terms