1
$\begingroup$

I have two scalars $\hat{\beta}_1$ and $\hat{\beta}_2$ which are unbiased estimators of $\beta$. Furthermore, the covariance matrix is given by $$var\begin{pmatrix}\hat{\beta_1}\\ \hat{\beta_2} \end{pmatrix}=\begin{pmatrix}\sigma^2_1&0\\0&\sigma^2_2\end{pmatrix}.$$ The ratio $\sigma_2/\sigma_1$ is known. Now, I have to construct a more efficient unbiased estimator of $\beta$. Could someone help me out here? I am fairly sure I should use Generalised least squares, but I don't know in what form. Thanks in advance!

$\endgroup$
1
$\begingroup$

Hi: divide the elements in the matrix by $\sigma^2_1$. Then, the covariance matrix becomes $\sigma^2_1 \times \left[ {\begin{array} {cc} 1 & 0 \\ 0 & \frac{\sigma^2_{2}}{\sigma^2_{1}} \\ \end{array} } \right] $

so that $\sigma^2_{1}$ is just a scale factor. Now, you can solve the problem:

minimize $w_1^2 \times 1 + w_2^2 \times \frac{\sigma^2_2}{\sigma^2_1}$

subject to $w_1 + w_2 = 1 $.

The resulting estimator, $w_1 \times \hat\beta_1 + w_2 \times \hat\beta_2$ is unbiased and minimum variance. You can use the lagrange multiplier approach to deal with the constraint or substitute $1-w_1$ for $w_2$ and take the derivative.

ADDENDUM A: Use the second approach.

This results in $w_1^2 + (1-w_1)^2 \times \frac{\sigma^2_2}{\sigma^2_1}$

Taking the derivative gives: $2 \times w_1 - 2 \times (1 - w_1) \times \frac{\sigma^2_2}{\sigma^2_1} = 0.$. This leads to

$2 \times w_1 + 2 \times w_1 \times \frac{\sigma^2_2}{\sigma^2_1} = 2 \times\frac{\sigma^2_2}{\sigma^2_1} $

$w_1 \times \left (2 + 2 \times \frac{\sigma^2_2}{\sigma^2_1} \right) = 2 \times \frac{\sigma^2_2}{\sigma^2_1}$

Please check the algebra because I expected something simpler for $w_1$.

ADDENDUM B: Actually, for $\sigma^2_1 = \sigma^2_2$, $w_1 = \frac{1}{2}$ which is correct (give same weight to both estimators) so the algebra is probably fine.

$\endgroup$

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.