Can I use the maximum likelihood to show a statistic is complete?

28 Views Asked by At

Problem: Consider a random sample of size $n$ that follows a density probability function given by:

$$f(x,\theta)=\frac{1}{\theta} x^{-\frac{\theta+1}{\theta}}\mathbb{1}_{(1,+\infty)},\:\:\theta>0$$

where $\theta$ is unknown.

Provide an sufficient statistic for this model.

In another thread I was able with the help of mathstckaexchenge members to derive the following maximum likelihood estimator $\hat{\theta} = \left( \frac{1}{n} \sum\limits_{i=1}^n \ln x_i \right) + 1$

I want to use the statistic obtained to answer the question of sufficiency. For that I am going to use the factorization theorem, which states that a statistic is sufficient for $\theta$ iff

$$f(x|\theta)=g(T(x)|\theta)h(x)$$

However to compute the joint distribution I would use the same procedure as to compute the likelihood function and I would obtain:

$f(x|\theta)=\prod_\limits{i=1}^{n}f(x_i,\theta)=\frac{1}{{\theta}^n}\prod_\limits{i=1}^{n} x_i^{-\frac{{(\theta+1)}}{{\theta}}}\mathbb{1}_{(1,+\infty)}$

However I am not able to separate the the x_{is} from $\theta$ if I do not use the logarithm.

Questions:

How do I solve the problem? Is it possible to use the maximum likelihood function for sufficiency? Can I use the logarithm? If so why?

Thanks in advance!

1

There are 1 best solutions below

0
On BEST ANSWER

I hope I understood your question correctly after the clarification in the comments.

Indeed you can use logarithms if you want to, and you don’t need to consider $\log f$ instead of $f$ in order to do so. Since

$$ f(x\mid\theta)=\prod_{i=1}^nf(x_i,\theta)=\frac1{\theta^n}\prod_{i=1}^n x_i^{-\frac{\theta+1}\theta}=\frac1{\theta^n}\exp\left(-\frac{\theta+1}\theta\sum_{i=1}^n\ln x_i\right)\;, $$

we have

$$ f(x\mid\theta)=g(T(x)\mid\theta)h(x)\;, $$

with

$$ T(x)=\sum_{i=1}^n\ln x\;. $$

(Note that I didn’t include the indicator function, which isn’t needed here (nor in your maximum likelihood estimator) because the samples lie in $[1,\infty$] anyway.)

Thus $T(x)$ is a sufficient statistic.

But you don’t have to use logarithms. You could also just write

$$ f(x\mid\theta)=\prod_{i=1}^nf(x_i,\theta)=\frac1{\theta^n}\prod_{i=1}^n x_i^{-\frac{\theta+1}\theta}=\frac1{\theta^n}\left(\prod_{i=1}^n x_i\right)^{-\frac{\theta+1}\theta}\;, $$

which factorizes with $S(x)=\prod_{i=1}^nx_i$. That $S(x)$ and $T(x)$ are likewise sufficient statistics is not surprising, since one is an injective function of the other ($S(x)=\exp(T(x))$ and $T(x)=\log(S(x))$), and an injective function of a sufficient statistic is again a sufficient statistic.