Let $X_1 , X_2 , ... , X_n$ be a random sample from a distribution with probability density function $\frac{1}{\theta}x^{\frac{1- \theta}{\theta}}I_{[0,1]}, \theta>0$. I want to find the 100$\alpha \% $ confidence interval for $\theta$.

This is what I have done:

$F_X(x;\theta) = x^{\frac{1}{\theta}}I_{[0,1]} + I_{(1, \infty)}$

We can pick $0<q_1, q_2<1$ such that $P(q_1 < \prod_{i=1}^{n}F_{x}(X_i ; \theta) < q_2) = \alpha$

Now we have that

$P(q_1 < \prod_{i=1}^{n}F_{x}(X_i ; \theta) < q_2) = P(q_1 < \prod_{i=1}^{n}X_i^{\frac{1}{\theta}} < q_2) = P(log(q_1) < \frac{1}{\theta}log(\prod_{i=1}^{n}X_i) < log(q_2)) = P(\frac{log(\prod_{i=1}^{n}X_i)}{log(q_1)} < \theta < \frac{log(\prod_{i=1}^{n}X_i)}{log(q_2)})$

Therefore a confidence interval for $\theta$ is $[\frac{log(\prod_{i=1}^{n}X_i)}{log(q_1)},\frac{log(\prod_{i=1}^{n}X_i)}{log(q_2)}]$. I believe this is correct. What I find unusual is, the expected length of the confidence interval seems to tend to infinity as n tends to infinity, I find this very counterintuitive, can this really be correct:

The length of the interval is $log(\prod_{i=1}^{n}X_i)(\frac{1}{log(q_2)} - \frac{1}{log(q_1)})$, this has expected value $nE(log(X_1))(\frac{1}{log(q_2)} - \frac{1}{log(q_1)}) = -n\theta(\frac{1}{log(q_2)} - \frac{1}{log(q_1)})$ (because $E(log(X_1)) = \int_{0}^{1} \frac{1}{\theta}log(x)x^{\frac{1}{\theta} - 1} dx = -\theta$). We can see that as n tends to $\infty$ so does the length of the confidence interval.

Can this be correct?


1 Answer 1


Your method for finding a confidence interval is correct. But there is a small mistake:

$$q_1<\prod_{i=1}^n X_i^{1/\theta}<q_2\iff \frac{\sum\ln X_i}{\ln \color\red{q_2}}<\theta<\frac{\sum\ln X_i}{\ln \color\red{q_1}}$$

You might also notice that $X_i\stackrel{\text{i.i.d}}\sim \text{Beta}\left(\frac{1}{\theta},1\right)$, so that $-\frac{2}{\theta}\ln X_i\stackrel{\text{i.i.d}}\sim\chi^2_2$.

Hence a pivot for $\theta$ is $$T(\mathbf X,\theta)=-\frac{2}{\theta}\sum_{i=1}^n\ln X_i\sim\chi^2_{2n}$$

So, $$P_{\theta}\left(\chi^2_{1-(1-\alpha)/2,2n}\leqslant T\leqslant \chi^2_{(1-\alpha)/2,2n}\right)=\alpha\quad\forall\,\theta$$

Equivalently, $$P_{\theta}\left(\frac{-2\sum\ln X_i}{\chi^2_{(1-\alpha)/2,2n}}\leqslant\theta\leqslant\frac{-2\sum\ln X_i}{\chi^2_{1-(1-\alpha)/2,2n}}\right)=\alpha\quad\forall\,\theta$$

So a $100\alpha\%$ C.I. for $\theta$ is

$$\left[\frac{-2\sum\limits_{i=1}^n\ln X_i}{\chi^2_{(1-\alpha)/2,2n}},\frac{-2\sum\limits_{i=1}^n\ln X_i}{\chi^2_{1-(1-\alpha)/2,2n}}\right]$$

Here, $\chi^2_{\alpha,2n}$ is of course the $(1-\alpha)$th fractile of a chi-square distribution with $2n$ degrees of freedom.

Length of this C.I is $$L=-2\sum_{i=1}^n \ln X_i\left[\frac{1}{\chi^2_{1-(1-\alpha)/2,2n}}-\frac{1}{\chi^2_{(1-\alpha)/2,2n}}\right]$$

When you take the expected length, the quantity within parenthesis is surely a constant, but it is dependent on $n$ (say, $c(n)$). That is, expected length is of the form $$E(L)=-2c(n)E\left[\sum_{i=1}^n \ln X_i\right]=2nc(n)\theta$$

I think you are ignoring the limiting behaviour of $c(n)$ (your $q_1,q_2$ also depend on $n$) when you are taking the limit $n\to\infty$.


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.