Suppose I have a sequence of iid random variables $X_1, \ldots, X_n$ following the pdf:

$$ f_\theta (x) = \theta x^{\theta-1} $$

for $\theta >0$ and $0 <x<1$.

I would like to find a sufficient statistic $T(X)$, such that the family $f_\theta (x)$ has a monotone likelihood ratio (MLR) in $T(X)$.

I do this by having:

$$ \frac{f(x|\theta_1)}{f(x|\theta_2)} = \frac{\prod_{i=1}^{n}\theta_1x_i^{\theta_1-1}}{\prod_{i=1}^{n}\theta_2x_i^{\theta_2-1}} = \left(\frac{\theta_1}{\theta_2}\right)^n \prod_{i=1}^n\left(x_i\right)^{\theta_1-\theta_2} = \left(\frac{\theta_1}{\theta_2}\right)^n \left(\prod_{i=1}^nx_i\right)^{\theta_1-\theta_2} $$

At this point, is the statistic corresponding to the MLR $\prod_{i=1}^nx_i$? If this is the case, what is the distribution of $\prod_{i=1}^n$ which is HARD to find?

Or would it be:

$$ \frac{f(x|\theta_1)}{f(x|\theta_2)} = \left(\frac{\theta_1}{\theta_2}\right)^n \left(\prod_{i=1}^nx_i\right)^{\theta_1-\theta_2} = \left(\frac{\theta_1}{\theta_2}\right)^n \left(e^{\sum_{i=1}^n \log(x_i)}\right)^{\theta_1-\theta_2} $$

and hence the sufficient statistic is $\sum_{i=1}^n \log(x_i)$ which has Gamma distribution?

  • $\begingroup$ One problem you don't explain where y comes from. $\endgroup$ – Michael R. Chernick Dec 25 '16 at 22:20
  • $\begingroup$ I don't think the function is a density for all theta >0 without further constraint on the range of x. Maybe that is where y comes in. $\endgroup$ – Michael R. Chernick Dec 25 '16 at 22:23
  • $\begingroup$ @MichaelChernick Sorry, $y$ should be $x$ here, I fixed it above $\endgroup$ – user321627 Dec 25 '16 at 22:26
  • $\begingroup$ I think that will help a lot. $\endgroup$ – Michael R. Chernick Dec 25 '16 at 22:34

You are correct that$$\prod_{i=1}^n x_i$$is sufficient. But since any bijective transform of a sufficient statistic is also sufficient$$\sum_{i=1}^n \log(x_i)=\log\left(\prod_{i=1}^n x_i\right)$$is also sufficient.

You are also correct that$$y=-\log(x)\sim\mathcal{G}a(1,\theta)$$and thus that the sufficient statistic$$-\sum_{i=1}^n \log(x_i)=\sum_{i=1}^n y_i\sim\mathcal{G}a(n,\theta)$$

| cite | improve this answer | |
  • $\begingroup$ What is confusing me is that if we instead used the first statistic, $\prod_{i=1}^n x_i$, its distribution will change and no longer be Gamma. How can a monotone likelihood property admit multiple sufficient statistics up to a bijection but also yield different distributions of that sufficient statistic? In other words, why isn't there a unique distribution when asked to find the sufficient statistic for a monotone likelihood property? Thanks!! $\endgroup$ – user321627 Dec 27 '16 at 16:40
  • $\begingroup$ This is clearly stated in my answer: any bijective transform like $\log$ is also sufficient, hence there is no unique distribution for a sufficient statistic. $\endgroup$ – Xi'an Dec 27 '16 at 21:33

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.