Find covariance of estimator and derivative of the log-likelihood function

122 Views Asked by At

Problem:
Given an estimator $\hat k$. The estimation method is either max likelihood or other method. We know that it's unbiased. Let $L$ be the likelihood function and $\ell = ln L$.
Find $\Bbb Cov( \frac{d \ell}{d k},\hat k)$

My attempt:
$$\Bbb Cov( \frac{d \ell}{d k},\hat k) = E( \frac{d \ell}{d k} \cdot \hat k) - E( \frac{d \ell}{d k}) \cdot E( \hat k)$$

As unbiased $E( \hat k) = k$. Also, $\frac{d \ell}{d k}$ is score function. We know that $E(s(k)) = 0$ Hence, we have

$$E( \frac{d \ell}{d k} \cdot \hat k) - E( \frac{d \ell}{d k}) \cdot E( \hat k) =E( \frac{d \ell}{d k} \cdot \hat k) $$

I decided to take $\hat k$ into the differential.

$$=E( \frac{d \ell \cdot \hat k}{d k})$$

Since the expectation is linear, I can take the derivative out of it.

$$= \frac{d E(\ell \cdot \hat k)}{d k}$$

However, I am still stuck. Am I doing something wrong?
Note. The friend of mine did manage to solve this, so the task has enough information to be solved.

1

There are 1 best solutions below

0
On BEST ANSWER

As all the commentators noted, the likelihood function makes a little sense if your estimator is obtained by, say, the method of moments. I would assume that, in this case, they don't correlate, so the covariance is zero.

Let $\hat k$ be MLE.
As unbiased, let replace the true parameter $k$ in log-likelihood function with $E[\hat k]$. Then, if you swap the derivative and expectation, you will have something like $\ell (\hat k)$. The MLE maximizes the log-likelihood function, consequently, the derivative is zero. The further steps are trivial.