Let $Y_1,...,Y_n$ be described by the relationship $Y_i=\theta x_i^2+\epsilon_i$, where $x_1,...,x_n$ are fixed constants and $\epsilon_1,...,\epsilon_n$ are iid $N(0,\sigma^2)$. How can I find the best unbiased estimator?

my work:

I am considering using the Rao-Blackwell Theorem, where $\phi(T)=E(W|T)$ is the BUE/MVUE of $\theta$. Here, $T$ is a complete, sufficient statistic for $\theta$ and $W$ is an unbiased estimator for $\theta$.

Since $f(y|\theta)=\frac{1}{\sqrt{2\pi\sigma^2}}exp[-y_i^2/(2\sigma^2)]exp[y_ix_i\theta/(\sigma^2)]exp[-\theta^2x_i^2/(2\sigma^2)]$, we have that $T=\sum^n_{i=1}y_i$ is a complete, sufficient statistic for $\theta$ since $\{\theta y_i: y_i \in R^1\}$. However, I am struggling finding a good unbiased estimator such that I can obtain $\phi(T)$.

  • 1
    $\begingroup$ Complete sufficient for $\theta$ is $\sum x_i^2y_i$. If you take the expectation you will get the answer as expected in a regression without intercept. Must have been answered here before. $\endgroup$ May 4 '20 at 5:34
  • 1
    $\begingroup$ Related: stats.stackexchange.com/q/412388/119261. $\endgroup$ May 4 '20 at 5:52
  • $\begingroup$ @StubbornAtom Thank you for linking that related post. I have posted my work as an answer below. I appreciate it! $\endgroup$
    – Jen Snow
    May 4 '20 at 23:54

As @StubbornAtom stated, the correct complete sufficient statistic for $\theta$ is $T(\mathbf{Y})=\sum_iy_ix_i^2$, since we can form a full-rank exponential family from the random sample $Y_i \sim N(\theta x_i^2,\sigma^2)$, where $i=1,...,n$.

$ET(\mathbf{Y})=\sum \theta x_i^4$. So, if we take $\phi(\mathbf{Y})=\frac{\sum y_i x_i^2}{\sum x_i^4}$, then $E\phi(\mathbf{Y})=\theta$.

So $\phi(\mathbf{Y})$ is an unbiased estimator for $\theta$ based on the complete sufficient statistic $T(\mathbf{Y})$ for $\theta$. By Rao-Blackwell Theorem, $\phi(\mathbf{Y})$ is our best unbiased estiamtor for $\theta$.


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.