I have the model $y_i=\beta_1+\beta_2 X_i+ u_i$ where $u_i\sim\text{iid } N(0,\sigma^2)$. I estimate $\beta_1$ and $\beta_2$ by drawing a straight line between the first $(x_1,y_1)$ and last dot $(x_n,y_n)$. So, $\hat{\beta}_2$ will be the slope of this straight line.

What is the variance of $\hat{\beta}_2$ (the estimate of $\beta_2$)? Make a t-test of $H_0: \beta_2=0$. What is the 95 % confidence interval for $\beta_2$?

When I calculate the variance of $\hat{\beta}_2$, I get that it equals to 0. Can it really be 0? What happens with the t-test and confidence interval if the variance, and thus the standard error, is zero?

  • $\begingroup$ Since this seems to be routine bookwork please add the self-study tag (remove one of your 5) and read its tag wiki info ... stats.stackexchange.com/tags/self-study/info ... modifying your question as needed. $\endgroup$ – Glen_b Aug 27 '14 at 20:58

The estimator for $\beta_2$ here is

$$\hat{\beta}_2 = \frac{y_n-y_1}{x_n-x_1}$$

The denominator is a constant. What is the distribution of $y_n$ and $y_1$? Use the following properties of variance to find $\text{Var}(\hat{\beta}_2)$:

If $c$ is a constant, $\text{Var}(cZ) =c^2\text{Var}(Z)$.

$\text{Var}(Z_1\pm Z_2)=\text{Var}(Z_1)+\text{Var}(Z_2) \pm 2\text{Cov}(Z_1,Z_2)$

($Z$, $Z_1$ and $Z_2$ are random variables)

Applying these rules, you will find that the variance is not zero.

| cite | improve this answer | |
  • $\begingroup$ Thanks! @comp_warrior The distribution of yn and y1 is not known. So when I calculate the variance I get: Var(β̂ 2)=1/(xn-x1)^2* Var(yn-y1)= 1/(xn-x1)^2* Var(β1+β2Xn+un-(β1+β2X1+u1)= 1/(xn-x1)^2* Var(β2Xn+un-β2X1-u1) Which I get equals to 0 since Var(β2Xn)=Var(β2X1)=0 since both β2 and Xi are assumed to be constants. And since Var(un)=Var(u1) these last terms cancel each other out. Where do I get the calculations wrong? $\endgroup$ – Matilda Aug 27 '14 at 17:41
  • $\begingroup$ @Matilda Are you sure the distributions of $y_n$ and $y_1$ are unknown? If $Z\sim N(\mu,\sigma^2)$, then $Z+a\sim N(a+\mu,\sigma^2)$. You can apply this property to both $y_n$ and $y_1$. $\endgroup$ – Comp_Warrior Aug 27 '14 at 18:06
  • $\begingroup$ @Matilda In your calculation, you cannot cancel out $u_n$ and $u_1$ since they are random variable and not constants. Instead, use the identities I have posted above. $\endgroup$ – Comp_Warrior Aug 27 '14 at 18:08
  • $\begingroup$ Yes, the information above is all information I have about yi. Aren't the variation of both un and u1 equal to σ2 according to ui∼iid N(0,σ2)? Wouldn't this imply that the terms Var(un)-Var(u1)=σ2-σ2=0 $\endgroup$ – Matilda Aug 27 '14 at 18:30
  • $\begingroup$ @Matilda Consider the equation for $y_1$: $\beta_1+\beta_2x_1$ is a constant, and we know $u_i\sim N(0,\sigma^2)$. So $y_i\sim N(\beta_1+\beta_2 x_1,\sigma^2)$. You are right that $\text{Var}(u_1)=\text{Var}(u_2)=\sigma^2$. But $\text{Var}(u_n-u_1)\ne \text{Var}(u_n)-\text{Var}(u_1)$. See my answer for the correct expression. $\endgroup$ – Comp_Warrior Aug 27 '14 at 18:52

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.