3

Let $X_1, X_2...X_n$ be iid with $f(x,\theta)=\dfrac{2x}{\theta^2}$ and $0<x\leq\theta$. Find $c$ such that $\mathbb{E}(c\hat{\theta})=\theta$ where $\hat{\theta}$ denotes MLE of $\theta$.

What I have tried: I found the MLE of $f(x;\theta)$ to be $\max\{X_1,X_2\cdots X_n\}$ (which aligns with the answer at the back) but now I am stuck at this question. The answer given is $\dfrac{2n+1}{2n}$.

I would have proceeded as: $$\begin{align*} \mathbb{E}(c\hat{\theta})&=\theta\\ \int_0^\theta c \dfrac{2x}{y^2} &=\theta \quad (y = \max\{x_1,x_2,\cdots,x_n\})\\ \dfrac{1}{y^2}\int_0^\theta c .{2x}{} &=\theta \end{align*}$$ But continuing this way gives me an answer far off from the one in the book (I don't have a term of n to begin with).

Ferdi
  • 5,179
Kplee
  • 33
  • 2
  • You can't find the expected value of the MLE (hence, you can't correct the bias) without knowing the distribution of the MLE. It is related to order statistics as suggested below, but you don't need to be familiar with order statistics to solve this problem. If $Y=\textrm{max}\left{X_1,X_2,\ldots,X_n\right}$, what is the CDF of $Y$, $F_Y(y)=P(Y\leq y)$? Hint: If $Y\leq y$, what does that say about each of the $X_i$? – assumednormal Oct 30 '12 at 03:48
  • 1
    @Max Working on your hint $Y \leq y$ implies that each $X_i\leq y$ but how do I use this in my problem? I understand that my expectation is taken over $\hat{\theta}$ and thus, we must use its distribution. Now that brings up 2 questions in my head: 1. Is it different from the distribution of $X_i$? The answer seems to be no (although I am not completely sure why). 2. If it's different, then what is it? – Kplee Oct 30 '12 at 03:55
  • The distribution of $Y$ is different from the distribution of each of the $X_i$. I'll take the hint a step further. I hope that's alright. If $Y\leq y$, then each of the $X_i$ are also at most $y$, as you've stated. Turn that into a probability statement: $P(Y\leq y)=P(X_1\leq y, X_2\leq y,\ldots, X_n\leq y)$. Are you able to explicitly compute that probability in terms of $\theta$? – assumednormal Oct 30 '12 at 03:58
  • 1
    So,$$P(\max {X_i}<y)=P(X_1<y,X_2<y,...,X_n<y) = \prod_{i=1}^n P(X_i<y)$$ $$=(\int_0^y\dfrac{2x}{\theta^2})^n=\dfrac{y^{2n}}{\theta^{2n}}$$ – Kplee Oct 30 '12 at 04:08
  • Exactly. Now you have the CDF of the maximum value of the sample. How can you use that to compute the expected value of the maximum value of the sample? Hint: The derivative, with respect to $y$, of the CDF will produce the PDF. – assumednormal Oct 30 '12 at 04:10
  • 1
    $$f(y) = \dfrac{2n.y^{2n-1}}{\theta^{2n}}$$ $$E(c\hat{\theta}) = \int_0^\theta\dfrac{c\cdot2n.y^{2n}dy}{\theta^{2n}}=\dfrac{c\cdot 2n}{\theta^{2n}}\times\dfrac{\theta^{2n+1}}{2n+1}=\theta$$ $$c = \dfrac{2n+1}{2n}$$ Awesome! Thanks! I'll mark the existing answer as the solution. – Kplee Oct 30 '12 at 04:17
  • 1
    @Max Actually, an easier calculation would be $$E[\hat{\theta}] = E[\max X_i] = \int_0^{\theta} 1 - \left(\frac{y}{\theta}\right)^{2n},\mathrm dy = \theta\left(1-\frac{1}{2n+1}\right)$$ giving $c = \frac{2n+1}{2n}$ – Dilip Sarwate Oct 30 '12 at 04:25
  • @DilipSarwate That is a nice, simple solution. Unfortunately, my approach to problems like these has become so burnt into my memory that I don't even consider other options. I found a way that worked for me and stuck with it throughout my coursework. You're integrating the survival function because $E[Z]=\int_{0}^{\infty}P(Z\geq z)dz$ so long as $Z$ is a positive random variable. – assumednormal Oct 30 '12 at 04:52

1 Answers1

3

What's the distribution of the largest observation from a sample of size $n$ with distribution F?

You want a special case of the formula for $f_{X_k}(X)$ where $k=n$ (which formula should be obvious in any case):

http://en.wikipedia.org/wiki/Order_statistic#The_joint_distribution_of_the_order_statistics_of_an_absolutely_continuous_distribution

Glen_b
  • 282,281
  • 1
    We haven't covered Order Statistics yet, any way to do without that? – Kplee Oct 30 '12 at 03:22
  • 1
    Yes, as the answer pretty much implies ("should be obvious in any case"), you can work it out yourself from first principles, using simple probability arguments. The probability that the largest value of a sample of $n$ is $\leq x$ is the probability that all $n$ observations are $\leq x$. Write that probability down (a power of a cdf at $x$), take derivates w.r.t $x$ to find the density. Then take expectations. to get $c$. (Though there are other ways to go about it, they're all going to involve some effort.) – Glen_b Oct 30 '12 at 20:20