Anderson

Hello,

I use the Anderson statistic to compare empirical distributions (the
equivalent of Cramer Von Mises with two samples).

More precisely, I have to divide my population in J subsets to make
sense, and my problem is to integrate the distances obtained by
applying the Anderson statistic separately.
Let Sj = sum (F1j(x) - F2j(x))^2, and Wj = n1jn2j / (n1j+n2j)^2

I hesitate between three ways of integration :
1) W = average (Wj)
2 ) W = sum (n1j+n2j) Wj / sum (n1j+n2j)
3) W = sum Sj / sum (n1j+n2j).

Which could be the more relevant ? (or maybe other ways are possible)
Thanks for you help.

Stephan

.

Quantcast