Pitman estimator for location paratemer

440 Views Asked by At

Given $Y_1,\ldots, Y_n$ are i.i.d random variables with pdf $\ f(y|\lambda) = e^{\lambda - y}$ for $y\geq \lambda$, and $\ f(y|\lambda) = 0$ otherwis, and $\lambda\in (-\infty, \infty)$. Find the Pitman estimator for a location-parameter $\lambda$ and compute the $MSE$ of that estimator.

My attempt: First, we see that the likelihood function $L(y|\lambda) = e^{-n(\overline{y} - \lambda)}$ for $y_{(1)}\geq \lambda$ ($y_{(1)} = min_{i=1,\ldots,n} y_i)$, and $L(y|\lambda) = 0$ otherwise. Now, using the formula to compute Pitman estimator of location, we have:

${\hat{\lambda}}(y) = \frac{\int_{-\infty}^{y_{(1)}} \lambda e^{-n(\overline{y} - \lambda)}d\lambda}{\int_{-\infty}^{y_{(1)}} e^{-n(\overline{y} - \lambda)}d\lambda} = \frac{\frac{y_{(1)}}{n}e^{-n(\overline{y} - y_{(1)})} - \frac{e^{-n(\overline{y} - y_{(1)})}}{n^2}}{\frac{1}{n}e^{-n(\overline{y} - y_{(1)})}} = y_{(1)} - \frac{e^{-n}}{n}$

Now, we have: $$MSE({\hat{\lambda}}(y)) = E(\hat{\lambda}(y)-\lambda)^2 = E(y_{(1)}^2) - 2(\lambda + \frac{e^{-n}}{n})E(y_{(1)}) + (\lambda + \frac{e^{-n}}{n})^2. $$

Now, since $E(y_{(1)}^2) = \frac{2}{n^2}$ and $E(y_{(1)}) = \lambda + \frac{1}{n}$ (are these results correct?), we plug them into the formula above to obtain

$$MSE({\hat{\lambda}}(y)) = \frac{1}{n^2} (2-2e^{-n} + e^{-2n}) -\frac{1}{n}(\lambda^2+2\lambda)$$ (I am skeptical about this result, because it's too ugly;p)

My question: Could someone kindly confirm if my Pitman estimator and particularly the MSE above are correct? If it is not, could you help point out the mistake?

1

There are 1 best solutions below

1
On BEST ANSWER

You messed up in calculating the Pitman Estimator. Since every term has exp(-n(y-bar-y(1)) you can just factor it out. Which leaves the Pitman Estimator as y(1) - 1/n. This will make the MSE much nicer looking.