As a slight modification of my previous problem:

Let $X_1, X_2, . . . , X_n$ be iid random variables having pdf

$$f_X(x\mid\theta) =\theta(1 +x)^{−(1+\theta)}I_{(0,\infty)}(x)$$

where $\theta >0$. Give the UMVUE of $\theta$, the Cramer-Rao Lower Bound (CRLB) for unbiased estimators of $\theta$ and compute the variance of the UMVUE of $\theta$.

I have that $f_X(x\mid\theta)$ is a full one-parameter exponential family with

$h(x)=I_{(0,\infty)}$, $c(\theta)=\theta$, $w(\theta)=-(1+\theta)$, $t(x)=\text{log}(1+x)$

Since $w'(\theta)=1$ is nonzero on $\Theta$, the CRLB result applies. We have

$$\text{log }f_X(x\mid\theta)=\text{log}(\theta)-(1+\theta)\cdot\text{log}(1+x)$$

$$\frac{\partial}{\partial \theta}\text{log }f_X(x\mid\theta)=\frac{1}{\theta}-\text{log}(1+x)$$

$$\frac{\partial^2}{\partial \theta^2}\text{log }f_X(x\mid\theta)=-\frac{1}{\theta^2}$$

so $$I_1(\theta)=-\mathsf E\left(-\frac{1}{\theta^2}\right)=\frac{1}{\theta^2}$$

and the CRLB for unbiased estimators of $\tau(\theta)$ is

$$\frac{[\tau'(\theta)]^2}{n\cdot I _1(\theta)} = \frac{\theta^2}{n}[\tau'(\theta)]^2=\boxed{\frac{\theta^2}{n}}$$

As for finding the UMVUE of $\theta$, since $\frac{1}{n}\sum \text{log}(1+X_i)$ is unbiased for $\frac{1}{\theta}$ then perhaps something similar to $\frac{n}{\sum\text{log}(1+X_i)}$ will be unbiased for $\theta$. After finding the expected value, I can hopefully make a slight adjustment to get an unbiased estimator. Let $T=\sum \text{log}(1+X_i)$

$$\mathsf E\left(\frac{n}{T}\right)=n\cdot\mathsf E\left(\frac{1}{T}\right)=n\int_0^{\infty}\frac{1}{t}f_T(t)dx$$

We must next find the distribution of $T$, but first let's find the distribution of $t=\text{log}(1+X)$. Let $Y=\text{log}(1+X)$. Then

$$\begin{align*} F_Y(y) &=\mathsf P(Y\leq y)\\\\ &=\mathsf P(\text{log}(1+X)\leq y)\\\\ &=\mathsf P(1+X\leq e^y)\\\\ &=\mathsf P(X\leq e^y -1)\\\\ &=F_X\left(-(1+e^y-1)^{-\theta}+1\right)\\\\ &=1-e^{-\theta y} \end{align*}$$

So $Y\sim \text{exp}\left(\frac{1}{\theta}\right)$ and hence $T\sim\text{Gamma}\left(\alpha=n,\beta=\frac{1}{\theta}\right)$


$$\begin{align*} \mathsf E\left(\frac{n}{T}\right) &=n\int_0^{\infty} \frac{1}{t} \frac{\theta^n}{\Gamma(n)}t^{n-1}e^{-\theta t}dt\\\\ &=\frac{n\theta}{n-1}\underbrace{\int_0^{\infty}\frac{\theta^{n-1}}{\Gamma(n-1)}t^{n-2}e^{-\theta t}dt}_{=1}\\\\ &=\frac{n}{n-1}\theta \end{align*}$$

It follows that $$\frac{n-1}{n}\cdot\frac{n}{\sum\text{log}(1+X_i)}=\boxed{\frac{n-1}{\sum\text{log}(1+X_i)}}$$ is an unbiased estimator of $\theta$ which is a function of the complete sufficient statistic $T$, and so by the Lehmann-Scheffe Theorem, it's the unique UMVUE of $\theta$.

As $\hat{\theta}\sim(n-1)\cdot\text{Inv-Gamma}(n,\theta)$


$$\mathsf{Var}\left(\frac{n-1}{T}\right)=(n-1)^2\cdot\mathsf{Var}\left(\frac{1}{T}\right)=(n-1)^2 \cdot \frac{\theta^2}{(n-1)^2\cdot(n-2)}=\boxed{\frac{\theta^2}{n-2}}$$

Are these valid solutions?

  • 2
    $\begingroup$ I think you're making a mistake with $E\left(\frac 1{X+Y}\right) \neq E(1/X) + E(1/Y)$ when you check your work $\endgroup$
    – jld
    Oct 25, 2018 at 19:07
  • 3
    $\begingroup$ With a change of variables $Y=\ln (1+X)$, I get an exponential density for $Y$ having mean $1/\theta$. This should help you to get a Gamma density for the complete sufficient statistic $T=\sum \ln(1+X_i)$, and hence an unbiased estimator of $\theta$ based on $T$. $\endgroup$ Oct 25, 2018 at 19:32
  • 2
    $\begingroup$ No. From the transformation formula, pdf of $Y=\ln(1+X)$ is $$f_Y(y)=f_X(e^y-1)\left|\frac{dx}{dy}\right|\mathbf1_{y>0}$$, where $f_X$ is the pdf of $X$. (Note that $x>0\implies y>0$) $\endgroup$ Oct 25, 2018 at 19:58
  • 1
    $\begingroup$ I used the CDF method, as I am more familiar with it, and came to the same conclusion. I will try to get get the unbiased estimator from here and update my answer. I appreciate your help. $\endgroup$
    – Remy
    Oct 25, 2018 at 20:09
  • 1
    $\begingroup$ The CRLB is correct. $\endgroup$ Oct 26, 2018 at 6:31

1 Answer 1


From your previous question, you already have the complete sufficient statistic:

$$T(\mathbf{X}) = \sum_{i=1}^n \ln(1+X_i).$$

The simplest way to find the UMVUE estimator for $\theta$ is to appeal to the Lehmann-Scheffé theorem, which says that any unbiased estimator of $\theta$ which is a function of $T$ is the unique UMVUE.

To find an estimator with these properties, let $T_i = \ln(1+X_i)$ and observe that $T_i \sim \text{Exp}(\theta)$ so that $T \sim \text{Gamma}(n,\theta)$. Hence, we can use the complete sufficient statistic to form the estimator:

$$\hat{\theta}(\mathbf{X}) = \frac{n-1}{T(\mathbf{X})} = \frac{n-1}{\sum_{i=1}^n \ln(1+X_i)} \sim (n-1) \cdot \text{Inv-Gamma}(n,\theta).$$

From the known moments of the inverse gamma distribution, we have $\mathbb{E}(\hat{\theta}) = \theta$ so we have found an unbiased estimator that is a function of the complete sufficient statistic. The Lehmann-Scheffé theorem ensures that our estimator is UMVUE for $\theta$. The variance of the UMVUE can easily be found by appeal to the moments of the inverse gamma distribution.

  • $\begingroup$ Thank you. I had not heard of the inverse gamma distribution until now. I have updated my answer attempting to solve for the variance of the UMVUE. Why is it not $\text{Inv-Gamma}(n,\frac{1}{\theta})$? $\endgroup$
    – Remy
    Oct 26, 2018 at 3:16
  • $\begingroup$ Never mind, my textbook has a different way of expressing the gamma distribution than wikipedia. My textbook has $f(x\mid\alpha,\beta)=\frac{1}{\Gamma(\alpha)\beta^{\alpha}}x^{\alpha-1}e^{-\frac{x}{\beta}}$ $\endgroup$
    – Remy
    Oct 26, 2018 at 3:40
  • $\begingroup$ @Remy If your UMVUE is $\hat\theta=\frac{n-1}{T}$, then \begin{align} \operatorname{Var}(\hat\theta)&=(n-1)^2\operatorname{Var}\left(\frac{1}{T}\right) \\&=(n-1)^2\left[E\left(\frac{1}{T^2}\right)-\left(E\left(\frac{1}{T}\right)\right)^2\right] \end{align} , which you can calculate from the distribution of $T$. $\endgroup$ Oct 26, 2018 at 5:15
  • $\begingroup$ @StubbornAtom Is what I have after using the method suggested by Ben incorrect or are you just suggesting an alternative approach? $\endgroup$
    – Remy
    Oct 26, 2018 at 5:25
  • 1
    $\begingroup$ @Remy I was suggesting a direct approach (without using Inverse Gamma distribution) since you only need to calculate a couple of moments. $\endgroup$ Oct 26, 2018 at 5:34

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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