I would like to use the Cramer-Rao lower bound to help me estimate the variance of my maximum likelihood estimator, across a range of different samples of data.

My question is, how do I do this correctly? My model has two parameters, $\mu$ and $\psi$.

Thus far I have been evaluating the following at the maximum likelihood estimates of my two parameters $\hat{\mu}$ and $\hat{\psi}$:

$I_{11}=\frac{\partial^2 ln L(\mu,\psi)}{\partial \mu^2}$,$I_{12}=I_{21}=\frac{\partial^2 ln L(\mu,\psi)}{\partial \mu \partial \psi}$,$I_{22}=\frac{\partial^2 ln L(\mu,\psi)}{\partial \psi^2}$

I then put these into a matrix, making a 2x2 information matrix of the form: $I_{estimate}=[[I_{11} ,I_{12}],[I_{21} ,I_{22}]]$.

I then invert this matrix, to get the Cramer-Rao lower bound; picking out the upper left component if I want to get an estimate of the asymptotic variance of the MLE estimator for $\mu$ for example.

My particular questions regarding my implementation are as follows:

  1. Is it correct to replace $\mathbb{E} (I)$ in the CRLB formula by the matrix $I_{estimate}$ where I have evaluated the various components at the MLE estimated values?

  2. Do I need to take account of the sample size in my ultimate estimate of the variance of the MLE estimators? For example, if we call the inverse of the information matrix, $J$, would the corresponding estimator of (asymptotic) variance just be: $J_{11}$ or should it be: $J_{11}/N$, where $N$ is the sample size?




Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.