Professional Documents
Culture Documents
CHI SQUARE TESTS FOR LINEAR TRENDS IN PROPORTIONS
CHI SQUARE TESTS FOR LINEAR TRENDS IN PROPORTIONS
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.
International Biometric Society is collaborating with JSTOR to digitize, preserve and extend access to
Biometrics.
http://www.jstor.org
SUMMARY
In many biological problems, the experimentermay prefer to use a chi-square
test for linear trends in proportions rather than the standard chi-square test of
homogeneity. In this paper the non-centralityparameterof this test for linear trend
in such a model is calculated. This permits the experimenterto make power com-
parisonsbetween the two tests and some tables are presented along such lines. The
power of the test for linearity in the same model is also studied. Such power studies
permit determinationof sample size and an exampleis discussedusing published data
on the possible effect of maternal smoking on the sex ratio at birth.
1. INTRODUCTION
The general results of Mitra [1958] and Diamond [1963] on the
asymptotic power of chi square tests permit experimenters to make
explicit power calculations and to design experiments or plan samnple
sizes with specific goals in mind, at least for reasonably large samples.
These general results however still require the experimenter to calculate
a rather complex expression usually involving matrix inversion and
while this could be done, if necessary, numerically for any given problem,
absence of simple formulae as well as long years of custom otherwise
have prevented biologists from making use of the available tools.
Recently however Meng and Chapman [1966] and Nam [1968] have
provided some of the explicit formulae required to calculate the non-
centrality parameter for several standard tests of the chi square family.
The aim of this paper is to provide some further formulae along these
lines, in particular giving consideration to the chi square test of trend
which has found nmuchuse in biological studies and which was discussed
in this journal by Cochran [1954] and by Armitage [1955]. Finally
an example is reviewed of use of such power calculations to determine
sample size.
2. THE BASIC THEORY
Consider k sequences of Nj trials which may result in E or E; it
is assumed that Pr [E occurs on jth sequence] = . Let v; be the
number of occurrences of E in the jth sequence.
315
E VI V2
... ... Vk V (1)
conditions on the functions pii and set R, under the alternative sequence
of hypotheses H,: p%i+- (ci/-/A/N) with i; c = 0, the c
statistic
f2 f
t=1i=l
[vi -
Nipii(a',,
Nipij(6, ...
*
,
at)
t)]2
where
= { ;V IX (3)
p= ZvN,
i=1
The limiting power of the test of size e based on the statistic X1 for
the specified sequence of alternatives is
Pr[xl _1 (X) > Xk,-l, ],
where X is given by (5). This formula for X was given by Meng and
Chapman [1966].
against
HA2: Pli a + /3xi+ (cl/-VN)
P21 - a -fXj - (cj/lVN),
where the matrix
...
X1 X2 Xkl
.
LC C2
:: Ck
has rank 2, i.e. not all c's are proportional to the x's.
The unknown parameters a, ,B can be estimated from the sample
(1) by the rnodified minimum chi-square estimation method.
Define co; = Nj/plp2j and co; = pp2, where Pi =v,/N;
Also let
kc / k ck k
H= ZXiXi/ Wi HZW-EX
j=l j=l j=l j=
k ic
=
XS cZA ((X -
) / E (X_ - .)2 (7)
jlo1 i=l
and
k - k / k
A AS 'i E Ai*8
CZE Aijpli E
x2
_ E[Vi-
Nipi(a,
d)]
i- Npjd S1pj(a )]
is distributed in the limit as chi-square with k - 2 D.F. under "02
and as non-central chi-square with k -2 D.F. anidnon-centrality param-
eter X under HA2 . We turn now to the evaluation of X.
For convenience write xi -,,, -= and as usual denote the true
values of pii with a superscript zero. Also the usual reparametrization
of the linear regression model reduces computations slightly. That is,
we write
Pi = a' + Oxj where a' = a - flj,
and work with the parameters a', d.
Fromn(3) and (4)
Q k
= -d, C L, .. ,Ck X Ck
= 1, =
pjj/,=a' &pud/a1 Xi,
and
PiA1x- )
B'B = p(-
-?E 0 -
= ~(~y~) k )2 k
coi- (E +/= CiZ)
(k )2 / k 6i?
The power of the test is Pr [Xk2(X) > Xk-2,f], where X is given by (9).
This will serve as an approximation to the actual power which is more
complicated wheii estimated c,3,fi, are used rather than the true co,A.
p= 1a + x = * k),
P2j = 1 Ppl
and test
H03: = 0 against HA3 : C/VN
= C (c z 0).
Suppose the wi's were klnown anid were used in the estimate of: of
formula (7); denote this estimate by ,. . It is easily seen that
k k
E(A.) =
j=l
E Zi? i=l
oixE i + o.T i t2
_ ( E @ X9
because
wijZi = 0,
aiid
/ c
j=1
=c2(Z
p1 p2j) (10)
pii = a + a +
H02 :x - OXj
p2j = I -Pli
against
tPlA a,
o + 0.19i + c j/l/
A 1 - A3j
E
j =1
wijpi -5.1786N1 so ca' = 0.341.
3 3
2
Zi=l~~~~~~~~~
(cj/VATX)co j= 0.02172293N
TABLE 1
POWER OF CHI-SQUARE TEST OF H02 FOR SPECIFIED ALTERNATIVES
Casel:x = 0, 1,2 Case2:x = 0, 1,4
Non-centrality Non-centrality
Alternatives parameter Altern-atives parameter
A A A A A A
Pul P12 P13 'I
X/N1 P1l P12 P13 x/N1
/v
.2 .3 .6 .0324 .3 .3 .8 .0481
.2 .2 .6 .1463 .3 .5 .8 .0149
.2 .1 .6 .4595 .3 .6 .8 .0808
.3 .4 .6 .0071 .3 .8 .8 .4881
.3 .3 .6 .0745 .3 .2 .8 .7287
.3 .2 .6 .2384
.4 .45 .6 .0068 Case 3: x -, 1, 2, 3
.4 .4 .6 .0278 Non-centrality,
.4 .3 .6 .1212 Alternatives parameter
.3 .3 .8 .2050 Pil P12 P13 P14 /N1
.3 .5 .8 .0073
.3 .6 .8 .0078 .2 .2 .2 .6 .2609
.3 .8 .8 .2475 .2 .2 .4 .6 .0702
.2 .4 .2 .6 .3332
.2 .4 .4 .6 .0340
.3 .2 .1 .6 .1757
.3 .4 .3 .6 .3432
(A) Test (1) where alternative pia's are not restricted: For illustration
calculations are shown for the case HA : p1 = 0.2, p1A 0.5, pA =4 0.8,
x = 0, 1, 2, and N1_- N2 = N3; since P= 0.5, cl/VIN -0.3,
c2/V\N = 0, c3/ VN = 0.3; thus
P1(1 Pi)
P (f c1Qi)1 ] = 0.72N1
TABLE 2
POWER OF CHI-SQUARE TEST OF HOMOGENEITY. TEST (1) WITH UNRESTRICTED
ALTERNATIVES; TEST (3) WITH ALTERNATIVE RESTRICTED TO pl = a + fX .
Power of Power of
X for test (1) X for test (3)
test (1) N1 = 30 test (3) N1 = 30
Alterniative d
1. x = 0, 1, 2 (2 D. F.) (1 D. F.)
A A 4
Pul P12 P13
TABLE 3
SEX OF INFANTS ACCORDING TO SMOKING HISTORY OF MOTHERS
and
X = 0.000582N (test 3, 1 D.F.).
Since for test 1 X must be equal to 14.17 to achieve power equal to
0.9, N = 24389 while with test 3 where X need only equal 10.51 to
achieve this power N = 18058. Thus while the observed sample is
ACKNOWLEDGMENTS
The authors wish to express their thanks to Dr. Blair Bennett for
calling their attention to the data cited in section 8 and the problems
associated therewith. It should also be noted that we have considered
only one aspect of the data of Ravenholt and Levinski-these authors
also considered birth weight and previous smoking history of the mothers
among other factors.
cas sont pr6sent6s dans cette perspective. La puissance du test de lin6arit6 dans le
meme modele est egalement 6tudiee. De telles 6tudes de puissance permettent la
d6termination de la taille de l'6chantillon, et un exemple est discut6, utilisant des
donn6es publiees sur l'effet possible du tabagisme maternel sur le sex ratio a la
naissance.
REFERENCES
Armitage, P. [1955]. Tests for linear trends in proportions and frequencies.
Biometrics11, 375-85.
Cochran, W. G. [1954]. Some methods for strengthening the common x2 tests.
Biometrics10, 417-51.
Diamond, E. L. [1963]. The limiting power of categorical data chi-square tests
analogous to normal analysis of variance. Ann. Math. Statist. 34, 1432-40.
Fix, E. [1949]. Tables of the non-central chi-square. Univ. Calif. Pub. Statist. 1,
University of California Press, Berkeley, 15-9.
Lowe, C. R. [1965]. Smoking during pregnancy. Letter to the Editor, Lancet,
May 15, 1965, p. 1071.
Meng, Rosa C. and Chapman, Douglas G. [1966]. The power of chi-square tests
for contingency tables. J. Amer. Statist. Ass. 61, 965-75.
Mitra, S. K. [1958]. On the limiting power function of the frequency x2 test. Ann.
Math. Statist. 29, 1221-33.
Nam, Jun-mo [1968]. Comparison of dispersion and chi-square goodness-of-fit
tests of Poisson distribution. Submitted to J. Amer. Statist. Ass.
Ravenholt, R. T. [1965]. Chi-square test and smoking during pregnancy. Letter
to the Editor, Lancet, July 31, 1965, p. 239.
Ravenholt, R. T. and Levinski, M. J. [1965]. Smoking during pregnancy. Letter
to the Editor, Lancet, May 1, 1965, p. 961.
Sankaran, M. [1959]. On the non-central chi-square distribution. Biometrika 46,
235-7.