I got this assignment from Generalized Linear Model class. At first glace it looked like it is an easy task, but there are a lot of subtle (at least in my opinion) things, which I would like to clarify with somebody. So here is the problem:

Assume that we are interested in random variable $Y=e^Z$. Variable $Z$ is obtained from simple linear regression i.e. $Z=a+bX+\varepsilon$, where $\varepsilon \sim N(0,4)$. Given 100 observations of $(Z,X)$ we have following empirical characteristics $\bar x=2$, $\sigma^2_x=1$, $\bar z = -1$ and $\operatorname {cov} (z,y)=-1 $.

  1. How do we construct approximate Wald's 95% confidence interval for parameter $\theta = E(Y|X=1/2)$, using the fact $E e^{\lambda U} = e^{a \lambda+\sigma^2\lambda^2/2}$, when $U \sim N(a, \sigma^2)$?

  2. How do we construct exact Wald's 95% confidence interval for individual prediction of $Y$ when $X=1/2$?

  3. Now assume that we do not know variance of $\varepsilon$ and from those 100 observations we know that the empirical variance of $Z$, $\sigma^2_z=4$.

I started by plugging in the empirical statistics to specify simple regression part and I got $Z=1-X+\varepsilon$, so the $$\operatorname{\log} Y =1-X+\varepsilon. $$ As the $\operatorname {log} Y$ is normaly distributed then the distribution of the mean of $Y$ should be lognormal? If $Y$ is log normal then I can't write write it in canonical exponential family form $$\operatorname{exp}\left\{ \frac{y \theta-b(\theta)}{a(\phi)}-c(y,\phi) \right\},$$ because I have $\operatorname{log}y$ instead of $y$ multiplied by $\theta$. That means that I do not know what $b(\theta)$ and other parameters are, which I need for standard error estimation.

So any tips on getting at least first question answered would be great!



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.