Let $X$ be a random variable with probability density function $f(x,\theta, \beta)=\beta e^{-\beta(x-\theta)} \mathbb{1}_{(\theta,\infty)}$ with $\beta>0, \theta \in \mathbb{R}$ (a two parameter exponential distribution) from which a random sample is taken. If $\beta$ is known and $\theta$ unknown, find an optimal confidence interval for $\theta$.

So I need help finding the pivotal quantity for this example. I thought using $U_{i}=X_{i}-\theta \sim \operatorname{Exp}(\beta)$ then $\overline{U}=\overline{X}-\theta \sim \operatorname{Gamma}(n,\beta)$ to homologate the procedure to find a confidence interval for $\lambda$ from $\operatorname{Exp} \sim (\lambda)$.

  • $\begingroup$ You did find a pivotal quantity $U$ for $\theta$. Alternatively, since $X_i-\theta$ is i.i.d $\mathsf{Exp}$ with mean $1/\beta$, we have $2\beta(X_i-\theta)$ i.i.d $\chi^2_2$, thus giving the pivot $2\beta\sum (X_i-\theta)\sim \chi^2_{2n}$. Now the confidence interval can be obtained using chi-square fractiles. $\endgroup$ – StubbornAtom May 6 at 16:20

Indeed $\overline X-\theta$ is a valid pivot for $\theta$ when $\beta$ is known. All you have to do now is find $a,b$ such that $a<\overline X-\theta<b\iff \theta\in(\overline X-b,\overline X-a)$ with the desired confidence coefficient $P_{\theta}\left[\theta\in(\overline X-b,\overline X-a)\right]$ (for a two-sided interval). This is easily done using software.

Note that you have $X_i-\theta\stackrel{\text{i.i.d}}\sim\mathsf{Exp}$ with mean $1/\beta$, which implies $X_{(1)}-\theta\sim \mathsf{Exp}$ with mean $1/(n\beta)$ where $X_{(1)}=\min\limits_{1\le j\le n}X_j$. So yet another pivotal quantity is $$T(\mathbf X,\theta)=2n\beta(X_{(1)}-\theta)\sim \chi^2_2$$

We expect a confidence interval based on this pivot to be 'better' (in the sense of shorter length, at least for large $n$) than the one based on $\sum\limits_{i=1}^n X_i$ as $X_{(1)}$ is a sufficient statistic for $\theta$.

Now you can derive a two-sided confidence interval with confidence level $1-\alpha$ starting from

$$P_{\theta}(\chi^2_{1-\alpha/2,2}< T< \chi^2_{\alpha/2,2})=1-\alpha\quad\forall\,\theta$$

Here $\chi^2_{\alpha,2}$ is of course the $(1-\alpha)$th fractile of $\chi^2_2$, i.e. $P(\chi^2_2>\chi^2_{\alpha,2})=\alpha$. Notably if you have an observed sample at hand, then calculations involving chi-square fractiles for the confidence interval can be done by hand since printed chi-square tables are readily available.


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.