Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

On Tests concerning Lines and Planes in Hyperspace

By B. M. BENNETT, Seattle 1)
Summary: As an application of tests of linear hypotheses methods are presented for determining confidence regions for intersections of several hyperplanes. These results constitute one
generalization of Tocher's method for point and interval estimates of a concurrent set of
regression lines in the plane.

1. Introduction
Let {y~} (0~ = 1, ..., N) be independent normal variables each with a common
variance and with means
I-I: G = E ( y ~ I x ~ ) = f 0 + f l x ~ +

... +flpxp~

(1)

where x~ = (Xl~, "", x~,) are N sequences of fixed numbers. It is known that tests of
hypotheses which specify q additional independent linear restrictions of the form:
Z13' = 0 where 13 = (/30, fix,'",f~) and Z is a given q x (p + 1) matrix may be determined as tests of linear hypotheses (KOLODZmJCZYK, 1935). The geometric interpretation of (1) is that of an unspecified hyperplane/7 with (p + 1) regression coefficients which are to be estimated on the basis of N sample points in the y - - x space.
Suppose now that there are q independent series of normal observations {Yr, ~,}
(r = 1, ".', q; o~r = 1, " " , Nr) for each of which the regression is assumed to be
linear
I-I,: rt,, ~, = fl,o + fl, l x,1. ~, + ... + f , p x,p, ~,

(2)

for r = 1 , - " , q on the basis of distinct sequences of known x's. The intersection So
or common region of the q hyperplanes H r in the y - - x space is of dimension _> p - q + 1 (e.g. McMAHON, 1923, Sommerville, ch. 6, 1929). It is the purpose of this
paper to consider various hypotheses further specifying the nature and dimensionality
of S Oand in particular to determine confidence region for So.
~) B. M. BENNETT,Universityof Aberdeen, Aberdeen, and Universityof Washington, Seattle.

42

B.M. BENNETT

2. Intersection of Regression Lines in a Plane


Thus for the case: p = 1 TOCHER (1952) discussed the problem of of testing the
hypothesis of a common intersection point ( = x0) of q regression lines with the x
( = xl) axis, i.e. the hypothesis: Xrl = --flro/flr~ = xo = constant for r = 1, ---, q,
and also showed how a confidence interval for x 0 could be constructed. The maximum
likelihood estimate may be obtained by minimizing the appropriate variance-ratio
test of this linear hypothesis with respect to x 0. WILLIAMS(1953, 1959, p. 137) has also
discussed TOCHER'S problem.
The method discussed by TocI-mR may in fact be readily extended to a somewhat
more general problem of determining a confidence region for the assumed point
(x0, Yo) of concurrence of a set or pencil of regression lines in the (x, y)-plane. Thus
if we are considering sets of linear growth data as a function of time ( = x) it may be
speculated that at a given time point ( = x0) the growth ( = Y0) is constant for each
set. Thus in extension of TOCHER'S problem if (xrj, YrJ) represent nr pairs of observations (r = 1, " ' , c; j = 1, " " , nr) such that E ( Y r j ) = oct + flrxra for the normally
and independently distributed y's, then a 100 (1 - - e) per cent. confidence region for
(xo, Yo) consists of pairs satisfying the inequality
F (Xo, Yo)

1 x;'@t
c {(y'"
( --_Yo) - b, xr.
Xo)_.__~}
2

c7
\n,
1

=c-~s , = ~ 1
\ nr

Srxx

"

(Y"--Y)2
~rxx

_<F~ (c, N - 2 c)

(3)

where Yr. (r = 1, " " , c) is the estimated ordinate of the r-th line with abscissa equal
to x0, and
( N - - 2c) sZ= ~ ~ ( y a - a r - brxrj) 2, Srxx= ~.~(xri- Xr)2
r /
j
at, br being the least squares estimates of o~r, fir. In (3) F, (c, N - - 2c) is the critical
value of the F test at the e - - level of significance with c and ( N - - 2 c) degrees of
freedom, if N = 2Jnr.

3. Intersections of Several Hyperplanes for Fixed y


Suppose first that two regression hyperplanes

1--[ : n = / ~ o + / h x l + "'" +/~pxp


I I ' : rl'=fl'o+fl'ix~ + "'" +fl~x'p

are independently estimated on the basis of the samples

(4)

On Tests concerning Lines and Planes in Hyperspace

43

(y~, xl~, "", xp~) (a = 1,..., N), (y'~., x ~ , , ..-, x'p,.)


(' --- 1, "", N') respectively, and it is required to determine a confidence region
So (X ) --- So (X1, "", X~ ) for X in the x-space as determined by the intersection
o f / 7 , / 7 ' with the x-space, i.e. ~? = 0, ~ / = 0.
We denote the usual two sample regression values

r ~ = ~ + b~ ( x ~ - x l ) +-.. + b, (x,~-x~)
, ,
-,

Y',, = y + bl (xi~, - xl) + " " + b, (x,~, - xp)

(5)

in terms of N y = ~ y~, N ' y ' = ~ y'~,, the least squares estimates of the regression
gt'

coefficients bi, bi' (i = 1, " ' , p ) and also denote [[cij[[ = [[aij[[-1, ][ct~'tl = I[aij't[ -x
by setting for i, j ---- 1, ..., p

au= 2 (Xi~-- Xi) (Xj~---Xj)


at

t
--!
r
--t
% = E (x,,,-~,)
(xj,,-xj)

(N+N,

2p_2) s 2 = X ( y _ y ~ ) 2 + ~ , , ( y , y,,)2
a~

at"

if N xi= X x,~,,N' x~= X x;~,,.


As an appropriate test for the linear hypothesis

~o + P~ x + ... + P~x =o
+ -. + p p, X po= 0

no:p;+p;xO
we

use the variance ratio

l [{y-b,(X-x~) .....

b, ( X - x ~ ) } 2

t'
4 ( ; ' - - b~ (Xo -- x l ) - - "'" - - b ; (X ---~;)}2]

(6)

{~-~-+22 cb(X--~',)(x--~9}l
and the inequality: F(X ) ~ F. (2, N + N ' - - 2p ~ 2) defines a 100 (1 - - e) per
cent. confidence region for X o, i.e. the region S o (X o).
If it is of interest to test the hypothesis that the regression planes 17, H' assume a
common (though usually unspecified) value say r/0, then the F-test

44

B.M. BENNETT

1 [ { y - - qo -- bt ( X - x l ) . . . . .

F=f(X"o)= ~

Ty--~-/~.=o

bp (Xp - x p ) } 2

_-T--~--T-y

[ {--~--+ZZ~,,(x,-~,~(X~-x~)}
+

{7'-,o - b~ (x -~',) . . . . .
~

b. (X- ~ ) y ]

~~

(7)

j
provides now an augmented (p + D-dimensional confidence region for (~o, X ) from
the inequality F (X o, ~0) ~ Fs-

Example 1. If in particular we consider the case: p = q =- 2 in (6), the region $o (X1 ,


X2) in the (xl, xz)-plane is defined by

<xo,xo =

[ (;- bl <x--/,).~ {-;' - b_.A(',x - -~,.___~


- b__i('~X - -~'~) } ~[~ _
{_~+ZZc;,(xO__~;)(xO_~,)}
<r.

(8~

I f we set G (X~ o, Xg.) = F ( X x , )(2 ) - - F v then G is an algebraic form of degree four


in X1 , X~, the coefficients of which will not be reproduced herein because of limitations of space. The b o u n d a r y points of the region: G (2(i , )(2 ) = 0 constitute the
100 (1 - - e) per cent. confidence region So (/(1 , 2(2). The m a x i m u m likelihood estimates (,~'1, ,~o) are obtained by an iterative process of minimization of F ( X t , X~).
The complete set of confidence regions S O(X o) are given by the contours (2(1, )(2 ,
F = F~) determined on each of the planes: F = F~ for various assigned levels of
significance = e. These regions So must be further distinguished into situations where
one or both coordinates m a y be i) inclusive, ii) exclusive or iii) non-existent according
to the classification of FIELLER (1954) depending u p o n the significance level used.

Example 2. Suppose now for the case: q = 2 in (2) we are concerned with determining
a confidence region So = (Xl, "'" , X~ ) wherein ~ / = ~' = 0, i.e. we concern ourselves
with a test of the hypothesis:/3o + fllXt + " " + fl~,X~, = flo' + fit' X1 + " " + fl~' X~,.
The appropriate test is the Student-t with N + N ' - - 2 (p - - 1) degrees of freedom
t=

--t

[(Y - y') + ~" b, ( X - x , ) - ~. b, (X, - x,)[

(9)

1) + E ~,j(x,-x,)(x,
os ~(1.-g-+-~. -xj)+yc,j(x,-x,)(xj-xj)
,
o
-,
-,
,

i,j

and S o is defined as the set of all points such that t ~_ t~ ( N + N ' - - 2p - - 2), t~ being
the value of the t-test which is significant at the e per cent. level. It is seen that S o is a

On Tests concerning Lines and Planes in Hyperspace

45

"quadric variety" (SOMMERVILLE,1929, p. 59). The inequality (9) above is equivalent to


P

diiXX+~.eiX+A<_O

Q (X)= ~
i,j=l

(10)

where

dij = (bi- b;) (bj -

bj) - (s 2 t~) (cij + ci'j)

e~ = (b j - bj) { ( y - y ' ) - ~ b , x i + E b~x;}


t
S2 t 2

--

, --,

A = { ( y - - y ' ) - - Z b , x , + ~ b~x;} 2
2

' -'-'

Finally if we are concerned with determining a confidence region Sq for the intersection of q independently estimated regression planes ~/1 = 0, ~h -- 0, -. ", ~/q = 0 as
specified in (2), then the region Sa = Sq (Xl, "'', X~) is determined by the inequality

1 " {y,-b,.(X-x,.) .....

b,p(Xp-x-,o)}Z<_F.(q.U_pq_q)

(11)

as a variance-ratio test, where now


q

N~

{ N _ ( p + l ) q } s 2 = ~ y, (y,,_y,~,)2
t = l cL= l

if Yrs, represents the sample estimate based on the r-th regression, and Ilaa~l1-1 = ]lc~i~ll
for r = 1, "", q and
q

a,,l=E(xr~--x,,)(x~#-x,i ), N= ~ N,.
~r

r = 1

The above formulation of the regression problem for the determination of confidence regions S O(X), etc. has been in terms of unweighted y's. In case the y's
represent some transformation of a qualitative response (e.g. probits) the above
results may be suitably modified. The problem arose initially in fact in determining a
confidence region in the (xv x2)-Plane for the line: ~ / = fllxl +/3~x 2 : ~70,~' : f l l t X 1 '
+ flz'x~'= ~lo, where r/0 = probit = 5.0, i.e. the line determined by the common
50 % response rate in two independently determined probit planes (e. g. Finney, 1952,
ch. 7). It should be noted that the examples previously discussed are illustrative of
various types of linear hypotheses with geometric interpretation in hyperspaee.

46

B.M. BENNETT

Acknowledgement. It is a pleasure to note the assistance of Dr. D. J. FINNEYboth in


proposing the problem initially and in guidance towards its general formulation. This
research was conducted while the author was a Senior Fulbright Scholar at the
University of Aberdeen.
References
FIELLER, E. C., CREASY,M. A., and DAVID,S. T. (1954): Symposium on interval estimation.
J. Roy. Statist. Soc. B 16, 175.
FINNEY,D. J. (1952): Probit Analysis. 2nd edition, Cambridge University Press.
KOLODZm~CZYK, ST. (1935): On an important class of statistical hypotheses. Biometrika
27, 161.
McMAHoN, J. (1923): Hyperspherical goniometry; and its applications to correlation theory
for n variables. Biometrika 15, 173.

SOMMERVILLE,D. M. Y. (1929): An Introduction to Geometry of N Dimensions. London,


Methuen & Co.
Tocrma, K. D. (1952): On the concurrence of a set of regression lines. Biometrika 39, 109.
W r L ~ S , E.J. (1953): Tests of significance for concurrent regression lines. Biometrika
40, 297.
(1959): Regression Analysis. John Wiley and Sons.

You might also like