Scale family ancillary statistic

88 Views Asked by At

This is the example 6.2.19 of Statistical Inference Book by Casella and Berger.

Scale parameter families also have certain kinds of ancillary statistics. Let $X_1,...,X_n$ be iid observations from a scale parameter family with cdf $F(x/\delta), \delta >0$. Then any statistic that depends on the sample only through the $n-1$ values $X_1/X_n,..., X_{n-1}/X_n$ is an ancillary statistic. For example,

$$\frac{X_1+...+X_n}{X_n}=\frac{X_1}{X_n}+...+\frac{X_{n-1}}{X_n}+1$$ is an ancillary statistic.

To see this fact, let $Z_1,...,Z_n$ be iid observations from $F(x)$ (corresponding to $\delta =1$) with $X_i=\delta Z_i$. The joint cdf of $X_1/X_n,..., X_{n-1}/X_n$

$$F(y_1,...,y_{n-1}|\delta)=P_\delta (X_1/X_n \leq y_1,..., X_{n-1}/X_n \leq y_{n-1})=P\delta (\delta Z_1 / (\delta Z_n) \leq y_1,..., \delta Z_{n-1} / (\delta Z_n) \leq y_{n-1})=P_\delta (Z_1/Z_n \leq y_1,....,Z_{n-1} / Z_n \leq y_{n-1})$$

The last probability does not depend on $\delta$ because the distribution of $Z_1,...,Z_n$ does not depend on $\delta$.

I can understand the above, which is clear. But I am confused that the statement "Then any statistic that depends on the sample only through the $n-1$ values $X_1/X_n,..., X_{n-1}/X_n$ is an ancillary statistic." Does this mean these $n-1$ values need to show up at the same time?

For example, if $n=4$, is $X_1/X_4, X_2/X_4$ an ancillary statistic?