Take the usual linear model: $$ Y=\alpha+\beta X + e$$

In good hypothesis (IID sample, normal error assumption $e \sim N(0,\sigma_e^2) $ and homoschedasticity) the distribution of the sample slope $\beta_n=\frac{S_{xy}}{S_x^2}$ in a simple linear regression is known to be: $$ \hat\beta_n \sim N\left(\beta,\frac{1-\rho^2}{n-3}\frac{\sigma_Y^2}{\sigma_X^2} \right) $$

Knowing $\sigma_Y $ and $\sigma_X $ it's straightforward to setup a CI on $\beta, $ due to normality of $\hat\beta_n$. $$ $$

If $(1-\rho^2)\sigma_Y^2$ is not known, due to it's identification with $\sigma_e^2$

(in fact $(1-\rho^2)\sigma_Y^2 = V[Y|X]=E[e^2|X]=\sigma_e^2)$,we can studentize (providing non trivial consideration) by knowing that the residual sum of squares is $rss \sim\sigma_e^2 \chi^2_{n-2}$ distributed and by it's consistency for $\sigma_e^2$.

From this we can find the pivotal quantity $t_{n-2}=\frac{Z}{\sqrt{\frac{K}{n-2}}}$ which clearly does not depends on $\sigma_e$ and allows us to set a t-student based CI on $\beta$.

My question is:

If instead $\sigma_e^2$ is known and $\sigma_X^2 $ is not, with a similar procedure I can't get rid of the parameter due to the fact that it appears in the denominator of the variance of $\hat\beta_n$. Which pivotal quantity should I use in order to put a proper CI in this case?

  • $\begingroup$ regression conditions on the X's $\endgroup$ – Glen_b Jul 11 '18 at 5:52
  • $\begingroup$ Only one regressor and normally jointed with Y $\endgroup$ – omega Jul 11 '18 at 8:25

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.