2
$\begingroup$

Looking at the Fisher information matrix for a simple linear model, such as here, I do not understand how to use the matrix to compute confidence intervals. There are multiple examples on the internet showing how to obtain the matrix, but I suppose that since it is thought to be trivial, it is nowhere shown how to use it.

My understanding after some reading, is that confidence intervals for a parameter under a normal distribution should be computed from $CI_{\theta} = \theta \pm 1.96 * \sqrt{\frac{1}{I(\theta)}}$.

For example, if I wanted to compute the 95% confidence interval for $\beta_1$ in the model $yi=\beta_0+\beta_1x_i+\epsilon_i$, how should I proceed given the matrix $\left[ \begin{matrix} \tfrac{n}{\sigma^2} & \tfrac{\sum_i^n x_i}{\sigma^2} & 0\\ \tfrac{\sum_i^n x_i}{\sigma^2} & \tfrac{\sum_i^n x_i^2}{\sigma^2} & 0\\ 0 & 0 & \tfrac{n}{2\sigma^4} \end{matrix} \right]$?

$\endgroup$
2
$\begingroup$

You have three parameters in your problem, so $\theta = (\beta_0, \beta_1, \sigma^2)$. $I(\theta)$ is a matrix and you cannot "divide by" $I(\theta)$, as in the formula in your second paragraph.

What you need instead is to take the inverse of $I(\theta)$. If you are looking for the confidence interval e.g. of $\beta_1$, you would take the element $[2,2]$ of $I(\theta)^{-1}$. and plug it in place of $I(\theta)$ in your formula. You would have then

$$CI_{\beta_1} = \hat\beta_1 \pm 1.96\sqrt{\frac{1}{I(\theta)^{22}}}$$

where $I(\theta)^{22}$ denotes the element $[2,2]$ of $I(\theta)^{-1}$.

$\endgroup$
  • $\begingroup$ So If I was trying to determine the intervals for beta0, I would use element [1,1] ? $\endgroup$ – Dom Jo Nov 28 at 17:27
  • $\begingroup$ Is it taking the square root of element [2,2] or taking square root of the inverse matrix and then taking element [2,2] $\endgroup$ – Dom Jo Nov 28 at 18:59

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.