Consider a random sample $\{X_1,X_2,X_3\}$ where $X_i$ are i.i.d. $Bernoulli(p)$ random variables where $p\in(0,1)$. Check if $T(X)=X_1+2X_2+X_3$ is a sufficient statistic for $p$.

Firstly, how can we find the distribution for $(X_1+2X_2+X_3)$? Or should it be broken down to $X_1+X_2+X_2+X_3$ and then will this follow $Bin(4,p)$? I think not because note that all the variables are not independent here.

Alternately, if I employ the factorization condition by just considering the joint p.m.f. of $(X_1,X_2,X_3)$ then $f(X_1,X_2,X_3)=p^{x_1+x_2+x_3}(1-p)^{3-(x_1+x_2+x_3)}=[p^{t(x)}(1-p)^{3-t(x)}]p^{-x_2}(1-p)^{x_2}$ where $t(x)=x_1+2x_2+x_3$.

This shows that $T$ is not sufficient.

But what if I want to follow the definition and want to apply $\dfrac{f(X|p)}{g(T(X)|p)}$ to check if this ratio is independent of $p$? Then I need to know the distribution of $g$. What then, is the distribution of $T(X)=X_1+2X_2+X_3$?

  • 1
    $\begingroup$ Hint: You don't need to know the full distribution of $T(X)$. Consider, for instance, the case $T(X)=2$: what is the conditional probability distribution of $(X|T(X)=2)$? $\endgroup$ – whuber Dec 31 '14 at 18:19
  • $\begingroup$ If $T(X)=2$ then $(X_1,X_2,X_3)\in\{(1,0,1),(0,1,0)\}$. So $P(X|T(X)=2)=p^2(1-p)+p(1-p)^2=p(1-p)$ which is dependent on $p$, correct? $\endgroup$ – Landon Carter Dec 31 '14 at 18:28
  • 1
    $\begingroup$ That's the right idea--but I don't see why you are adding the two probabilities. Isn't $X$ a vector? (If you like, you can use the same kind of calculations to find the full distribution of $T(X)$ (it can only attain the values $0,1,2,3,4$), but that's no longer necessary, is it?)) $\endgroup$ – whuber Dec 31 '14 at 18:29
  • $\begingroup$ Yeah, right. Thanks! So once we show that this ratio is not independent of $p$ for at least once sample, then we are done! Thank you. And HAPPY NEW YEAR :) $\endgroup$ – Landon Carter Dec 31 '14 at 18:31
  • $\begingroup$ Yes $X$ is a vector, but more importantly $X=(X_1,X_2,X_3)$ and the probability $P(X|T(X)=2)=P(T(X)=2)=P(X=(1,0,1))+P(X=(0,1,0))$. Please correct me if I am wrong. $\endgroup$ – Landon Carter Dec 31 '14 at 18:35

I had a discussion with "whuber" and maybe I got a (correct?) hint to look at any sample point: evaluate $\dfrac{P(X=x)}{P(T(X)=T(x))}$ at that sample point $x$ and check if this ratio is independent of the parameter, in this case $p$.

So take $x=(1,0,1)$ then $T(1,0,1)=2$. So we evaluate $\dfrac{P(X=(1,0,1))}{P(T(X)=2)}$.Now, $$T(X)=2 \text{ iff } X\in\{(1,0,1),(0,1,0)\}.$$ Due to the i.i.d. property, $$P(X=(1,0,1))=p^2(1-p)\text{ and }P(X=(0,1,0))=p(1-p)^2.$$ Also $$P(T(X)=2)=P(X=(1,0,1))+P(X=(0,1,0))=p(1-p).$$

Hence $$\dfrac{P(X=(1,0,1))}{P(T(X)=2)}=\dfrac{p^2(1-p)}{p(1-p)}=p$$ which is clearly dependent on $p$, and therefore $T$ is not a sufficient statistic.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.