4
$\begingroup$

Let $X_i$ be i.i.d $exp(\lambda)$ and take any $a > 0$.

I want to find the UMVUE of $P(X_i < a ) = 1-\exp(-\lambda a)$.

My attempt

By properties of the exponential family we know $\sum_i X_i$ is a complete sufficient statistic for $\lambda$.

Hence we only need to:

  1. Find an unbiased estimator of $P(X_i < a)$

  2. Condition it on $T(X) = \sum_i X_i$ and apply Lehmann-Scheff Theorem

An obvious candidate for an unbiased estimator is using indicator functions. For example

$$\delta (X) = 1(X_1 < a)$$

Is an unbiased estimator. Hence the estimator

$$T(X) = E[\delta(X) | T(X) = t]$$

Is the UMVUE of $P(X < a)$. However, I am having trouble deriving this. The problem I am having is the fact I have conditioned a continuous random variable on a discrete condition $T(X) = t$. How can I proceed?

$\endgroup$
4
  • $\begingroup$ That you have conditioned on a continuous random variable taking a particular value is not a problem. We are considering the event $\{t-h\le T\le t+h\}$ as $h\to 0$ as is usually the sense for conditional distributions in the continuous case. $\endgroup$ Oct 27 '18 at 18:34
  • $\begingroup$ I reckon this post should answer your question. I had found the conditional expectation directly (conditional probability) using well-known results involving the Beta and Gamma distributions in my answer. $\endgroup$ Oct 27 '18 at 18:58
  • $\begingroup$ Wow! That's a really clever work around to avoid the integration. Thanks @StubbornAtom $\endgroup$
    – Xiaomi
    Oct 28 '18 at 1:07
  • $\begingroup$ @StubbornAtom apologies for asking the same question twice! I am reworking through Lehmann-Casella and forgot I had asked it. I have given an alternative answer (hence bumping this) as I think it will be useful to people who might stumble across this. It is similar to the method used for normal probabilities. $\endgroup$
    – Xiaomi
    May 19 '19 at 7:22
4
$\begingroup$

Going to post a late answer as it's quite simple but useful trick. The UMVUE is indeed given by

$$P(X_i < a|T)$$

The trick is to use Renyi's representation which says that

$$\frac{X_1}{\sum_{i=1}^n X_i} =_d U_{(1)}$$

Where $U_{(1)}$ is the 1st order statistic from a uniform(0,1) random sample

$$U_1, U_2, \dots U_{n-1}$$

$$P(X_1 < a | \sum_i X_i = t) = P \left(\frac{X_1}{\sum_i X_i} < \frac{a}{t}|\sum_i X_i=t \right) = P\left(\frac{X_1}{\sum_i X_i} < \frac{a}{t}\right)$$

The last inequality follows from Basu's theorem which says every complete sufficient statistic is independent of every ancillary statistic. Since $X_i$ come from a scale family, it follows $X_i/\sum_i X_i$ is ancillary. Hence the probability is independent of the complete sufficient statistic $T(X) = \sum_i X_i$.

Hence by the above Renyi's representation, using the fact that the minimum of a Uniform(0,1) sample (of size $n-1$) has distribution $P(U_{(1)} \leq u ) = 1- (1-u)^{n-1}$, it follows that the UMVUE is given by

$$P(X_1 < a | \sum_i X_i = t) = P\left(U_{(1)} < \frac{a}{t} \right) = 1 - \left(1-\frac{a}{t} \right)^{n-1}$$

ie.

$\delta (X) = 1 - (1 - \frac{a}{\sum_i X_i})^{n-1}$ is UMVUE

$\endgroup$

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.