Professional Documents
Culture Documents
Hinkley 1969
Hinkley 1969
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp
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.
Biometrika Trust and Oxford University Press are collaborating with JSTOR to digitize, preserve and extend access to
Biometrika.
http://www.jstor.org
This content downloaded from 139.184.14.159 on Thu, 01 Oct 2015 11:07:04 UTC
All use subject to JSTOR Terms and Conditions
Biometrika(1969), 56, 3, p. 635 635
Printed in GreatBritain
normalrandomvariables
On theratioof twocorrelated
By D. V. HINKLEY
ImperialCollege
SUMMARY
The distribution oftheratiooftwocorrelated normalrandomvariablesis discussed.The
exact distribution and an approximationare compared.The comparisonis illustrated
numerically in the linearmodel
forthe case of the normalleast squaresestimateof cc/,8
E(yj) = ac+,i (i = 1, ..., n) withuncorrelated error
normal terms.
1. INTRODUCTION
In a regression ofinterestto estimatetheratio
analysisofbivariatedata it is sometimes
oftwopopulationparameters. Two examplesare:
(i) The analysisofthe simplelinearmodelyi = ac+,/ui+ ei (i = 1,..., n), wheree], *,en
are independently normallydistributed withzeromeanand variance0.2; then - c/,8is the
interceptoftheregression linewith theu-axis.
Y
{
(ii) The analysisofthetwo-linelinearmodel
cc+/lbiu+i (i = 1,...,n1),
2+fl2Ui+6i (t=n,+ 1, ....,nL+n2),
withnotationas in (i); theratio(c1 - c2)/(,82 -1) is the abscissa of the intersectionof the
two regressionlines.
In each example the maximum likelihood estimate of the ratio is the ratio of two cor-
related normally distributedvariables, themselves estimates. In this paper the general
distributionofratios ofthis type is derivedand comparedwiththe approximationobtained
by assuming the denominatorrandom variable to be of constant sign. This has particular
relevance to the examples above, and a numerical comparisonis given for (i) above.
2. THEORETICAL RESULTS
Let X1 and X2 be normally distributedrandom variables with means Ot,variances ort
(i p, and let W = X1/X2. The exact distributionof W
= 1, 2) and correlationcoefficient
and the standard approximationbased on assuming X2 > 0 are examined in some detail.
This content downloaded from 139.184.14.159 on Thu, 01 Oct 2015 11:07:04 UTC
All use subject to JSTOR Terms and Conditions
636 D. V. HINKLEY
scale will provide the general ratio W.' This assumes there are fourparameters,whereas
thereare five: 0t, 02, o2t,o2 and p. In fact Z has no great advantage over W, since the distri-
butions of both involve the bivariate normal distribution.
If the joint densityof (X1, X2) is g(x,y) and the p.d.f. of W is f(w), then
f(w)
_ b(w)d(w)
f)V(27r) oi2a3(w)
(D b(w)
[ (1 -p2) a(w)f
(D b(w)
(1 -p2)a(w)JJ
H
2
p
+ V,(1 expt } (1
7Tf1, f2a2(W) Xp(2(1 _ p2>),1
021 1+02 2~
_= 2PI02J+
02 010 2
+ 2
1 2
- ca2(W)1
d(w) = exp ~b2(W)
2(w p2)a2(w)f (2)
will be a useful approximationto F(w). It is then importantto compare F(w) and F*(w) in
some detail.
This content downloaded from 139.184.14.159 on Thu, 01 Oct 2015 11:07:04 UTC
All use subject to JSTOR Terms and Conditions
Ratio of correlatednormal random variables 637
Now F(w) = pr (X1-wX2 < 0, X2 > O) + pr (X1-wX2 > 0, X2 < 0)
= F*(w) + pr(X2 < 0) {1-2 pr (X1-wX2 < 0 X2 < 0)}, (6)
so that IF(w)-F*(w)I < pr(X2 < 0) = 'D(-62/o2). (7)
This bound on the difference
is attained at w = + oo, for,by the definitionofF*(w) or by (6),
F*(-oo) = (D((-02/o2) and F*(oo) = 1-F(-02/o2), (8)
whereas F(w) is proper.Note that (7), a familiarresult,impliesF* -> F uniformlyas O-2 -> 0.
Further comparison is possible by examination of the derivative of F*(w), which by
(5) and (2) is (w) b )(w)
f*Mw) V2rooa3(w)
= (9)
where 3/b1= PO1Oi02-020i and = When b(w) > 0, it is easy to see that
?I2 P620102-l0i2.
f(w) > f*(w). For by (1) and (9)
The inequality (D(- u) < u-1s(u) (u > 0) gives the desired result. Hence f(w) > f*(w) for
all w. That is, F(w)-F*(w) increases monotonically from - D(- 02/o2) at w = -0o to
FD(-02/o2) at w = +oo.
The next step is to determinethe point at which F(w) = F*(w). Now by (6)
(__2) (11)
0 2 _2f_-?? -??
g((x,y)dxdy,
F(w) = F*(w)+D
whereg(x,y) is the bivariate normal density.The double integralin (11) is easily reduced to
~(u)du 2 u ( _p2) )
=
f:2 1(l-
When = 0, the solution is wo =
/fr2 P01/02, and in general the solution wo satisfies
wo > P if pb2> 0;
02
in factIWO
- increasesas h2l increases.
(PC1/02)1f
This content downloaded from 139.184.14.159 on Thu, 01 Oct 2015 11:07:04 UTC
All use subject to JSTOR Terms and Conditions
638 D. V. HINKLEY
Finally, moreprecisebounds than (7) can be obtained forfinitew. Startingwith (12) it is
possible to put bounds on the double integral,firstintegratingby parts and then using the
inequality J(D(u)< 5(u)/(- u) fornegative u. The resultingbounds on F(w) - F*(w) are
J < F(w)-F*(w) < J+K(w) (w > PO10C2),
where J= (D )
[12' (a J2V(lp2)}]
3. A NUMERICAL ILLUSTRATION
To illustrate the results in the previous section, in particular the new bounds (13) on
F(w) - F*(w), several calculations were carried out for the simple linear regressioncase
mentioned in ? 1. The variables X1 and X2 are respectively& and and specificvaluesAl,
a = 0, ,8 = 02, o-= 1 are used. The covariance matrixof (&(, ) gives the formulaefor o-,
J2 and p. Lastly, O, = acand O2 = fi.
The shorttable given (Table 1) is forthe case n = 10. Values of F(w), F*(w), bounds (13)
and uniformbound (7), are given forseven values of w. The values of F(w) were obtained
by a two-dimensionaliterative Simpson's Rule, and are correctto the number of figures
given.
The value of- 3fr1/If2 is 7 0 and ?f2> 0, SO that by (10) F*(w) is decreasingforw > 7 0,
albeit veryslowly.The precisionof the bounds (13) is quite striking,and it is interestingto
see that the uniformbound is itselfnearly attained over a large part of the range. The
bounds do become slacker in the tails. This and other examples suggest as an improved
approximation to F(w) F**(w) = F*(w) - @((- 02/02), (14)
when F*(w) > F(w) over most of the range; the correctiontermin (14) would be added if
F*(w) < F(w) over most of the range.
The relative simplicityofthe approximationsF*(w) and F**(w) facilitatesextensiveuse
of F(w), particularlyin the two regressionexamples of ? 1, where approximation is very
This content downloaded from 139.184.14.159 on Thu, 01 Oct 2015 11:07:04 UTC
All use subject to JSTOR Terms and Conditions
Ratio of correlatednormal random variable8 639
accurateformoderatesamplesizes.Foran applicationto thetwo-phaseregressionsituation,
see Hinkley(1969).Approximation maybe usefulevenforcomputing a singlevalueofF(w)
sincecomputation of(3) can ofteninvolvetrivariate fromtablesofL(h,k;y).
interpolation
to MrsE. A. Chambersforthenumericalcomputations.
I am grateful
REFERENCES
This content downloaded from 139.184.14.159 on Thu, 01 Oct 2015 11:07:04 UTC
All use subject to JSTOR Terms and Conditions