Professional Documents
Culture Documents
Box’s M statistic is used to test for homogeneity of covariance matrices. The jth set
4
of r dependent variables in the ith cell are yij′ = xij′ B + eij′ where eij ~ Nr 0, wij−1Σ i 9
for i = 1,K, g and j = 1,K, ni . The null hypothesis of the test for homogeneity of
covariance matrices is Ho :Σ1 = L = Σ g . Box (1949) derived a test statistic based
on the likelihood-ratio test. The test statistic is called Box’s M statistic. For
moderate to small sample sizes, an F approximation is used to compute its
significance.
1
Box’s M statistic is not designed to be used in a linear model context; therefore
the observed cell means are used in computing the statistic.
Notation
The following notation is used throughout this chapter, unless otherwise stated:
1
Although Anderson (1958, Section 10.2) mentioned that the population cell
means can be expressed as linear combinations of parameters, he assumed that the
combination coefficients are different for different cells, which is not the model
assumed for GLM .
1
2 Appendix 14
Statistics
Means
∑
ni
yi = y ij ni
j =1
%K 3 83 8′ 1n − 16
∑
ni
wij y ij − y i y ij − y i if ni > 1
=&
i
j =1
K'0
Si
if ni ≤ 1
%K∑ 1n − 16S 0n − g5
g
if n > g
S=&
i i
i =1
K'0 if n ≤ g
Box’s M Statistic
%K g
M=&
0n − g5 log S − ∑ 1n − 16 log S i i if S > 0
K'SYSMIS i =1
if S ≤ 0
Appendix 14 3
Significance
1
1 − CDF. F γM, f1, f2 6
where CDF.F is the SPSS function for the cumulative F distribution and
0 50 5
f1 = g − 1 r r + 1 2
1 g
1
0 50 5 1n − 16 0n − g5
2 r 2 + 3r − 1
ρ = 1−
6 r +1 g −1 ∑ −
i
i =1
0r − 150r + 25 g
1
60 g − 15 ∑ 1n − 16 0n − g5
1
τ− − 2 2
i =1 i
f1 + 2
f2 =
1 6
τ − 1− ρ
2
γ =
1ρ − f f 6
1 2
f1
References
Anderson, T. W. 1958. Introduction to multivariate statistical analysis. New York:
John Wiley & Sons, Inc.
Box, G. E. P., 1949. A general distribution theory for a class of likelihood criteria.
Biometrika, 36: 317–346.
Seber, G. A. F. 1984. Multivariate observations. New York: John Wiley & Sons,
Inc. (Section 9.2.6).