286 Some Elementary Statistical Inferences
ofQ 3 =
∑ 4
1 (Xi−npi^0 )
(^2) /(np
i 0 )is
(6−5)^2
5
- (18−15)^2
15
(20−25)^2
25
(36−35)^2
35
64
35
=1. 83.
The following R segment calculates the test andp-value:
x=c(6,18,20,36); ps=c(1,3,5,7)/16; chisq.test(x,p=ps)
X-squared = 1.8286, df = 3, p-value = 0.6087
Hence, we fail to rejectH 0 at level 0.0250.
Thus far we have used the chi-square test when the hypothesisH 0 is a simple
hypothesis. More often we encounter hypothesesH 0 in which the multinomial prob-
abilitiesp 1 ,p 2 ,...,pkare not completely specified by the hypothesisH 0 .Thatis,
underH 0 , these probabilities are functions of unknown parameters. For an illustra-
tion, suppose that a certain random variableYcan take on any real value. Let us
partition the space{y:−∞<y<∞}intokmutually disjoint setsA 1 ,A 2 ,...,Ak
so that the eventsA 1 ,A 2 ,...,Akare mutually exclusive and exhaustive. LetH 0 be
the hypothesis thatY isN(μ, σ^2 )withμandσ^2 unspecified. Then each
pi=
∫
Ai
1
√
2 πσ
exp[−(y−μ)^2 / 2 σ^2 ]dy, i=1, 2 ,...,k,
is a function of the unknown parametersμandσ^2. Suppose that we take a random
sampleY 1 ,...,Ynof sizenfrom this distribution. If we letXidenote the frequency
ofAi,i=1, 2 ,...,k,sothatX 1 +X 2 +···+Xk=n, the random variable
Qk− 1 =
∑k
i=1
(Xi−npi)^2
npi
cannot be computed onceX 1 ,...,Xkhave been observed, since eachpi, and hence
Qk− 1 , is a function ofμandσ^2. Accordingly, choose the values ofμandσ^2 that
minimizeQk− 1. These values depend upon the observedX 1 =x 1 ,...,Xk=xkand
are calledminimum chi-square estimatesofμandσ^2. These point estimates of
μandσ^2 enable us to compute numerically the estimates of eachpi. Accordingly,
if these values are used, Qk− 1 can be computed onceY 1 ,Y 2 ,...,Yn, and hence
X 1 ,X 2 ,...,Xk, are observed. However, a very important aspect of the fact, which
we accept without proof, is that nowQk− 1 is approximatelyχ^2 (k−3). That is, the
number of degrees of freedom of the approximate chi-square distribution ofQk− 1 is
reduced by one for each parameter estimated by the observed data. This statement
applies not only to the problem at hand but also to more general situations. Two
examples are now be given. The first of these examples deals with the test of the
hypothesis that two multinomial distributions are the same.
Remark 4.7.1.In many cases, such as that involving the meanμand the variance
σ^2 of a normal distribution, minimum chi-square estimates are difficult to com-
pute. Other estimates, such as the maximum likelihood estimates of Example 4.1.3,
μˆ=Y andσˆ^2 =(n−1)S^2 /n, are used to evaluatepiandQk− 1. In general,Qk− 1
is not minimized by maximum likelihood estimates, and thus its computed value