I've got a linear model: $y_i=β_1x_{i1}+β_2x_{i2}+ε_i$ where E($ε_i$)=0 and Var($ε_i$)= $σ^2I_n$ for i=1,...,n

Supposed we don't have the data for $x_{i2}$ and we estimate: $y_i=β_1x_{i1}+ε_i$ for i=1,...,n

So far I've shown that $\hat{\beta}_1$ is biased and that $s^2= {ε_i'ε_i}/(n-1)$ is also biased and that such bias is positive.

I now have to explain why I'd expect that the OLS estimator of $\hat{\beta}_1$ is inconsistent and then suggest a suitable estimator giving conditions where it'd be consistent.

Please could someone give me a helping hand, I'm not sure why $\hat{\beta}_1$ is inconsistent?


  • $\begingroup$ What you hav writtn os not mak any snse! $\beta_1$ is a number, not a random variable, so what does it mean for it to b inosistnt? something is surely missing hre. $\endgroup$ – kjetil b halvorsen May 31 '15 at 10:48
  • $\begingroup$ @kjetilbhalvorsen oops most of those β should be β hat. I'm not sure how to amend that! Also added the other equation which I missed out! $\endgroup$ – Emma May 31 '15 at 11:01
  • $\begingroup$ it your post! to get $\hat{\bta}$ write (within dollar signs) \hat{\beta} $\endgroup$ – kjetil b halvorsen May 31 '15 at 13:09
  • $\begingroup$ There we go, should be fully correct now! $\endgroup$ – Emma May 31 '15 at 13:44
  • $\begingroup$ @kjetilbhalvorsen sorry forgot to tag you! $\endgroup$ – Emma May 31 '15 at 16:28

$\hat{\beta} = (x_1'x_1)^{-1}x_1'y$

Now substitute the true model $y=X\beta + \epsilon$ into $y$

$\hat{\beta} = (x_1'x_1)^{-1}x_1'y = \beta_1 + (x_1'x_1)^{-1}x_1'x_2\beta_2+ (x_1'x_1)^{-1}x_1'\epsilon_i$

Taking the limit sends the last term to $0$, which leaves

$\hat{\beta} \rightarrow \beta_1 + \frac{cov(x_1,x_2)}{var(x_1)}\beta_2$

Which means all you need to do if figure out when $\frac{cov(x_1,x_2)}{var(x_1)}\beta_2$=0

  • $\begingroup$ why is it $cov(x_1,x_2)$ and $var(x_1)$? I get E($x_1x_2'$) and E($x_1x_1'$)? $\endgroup$ – Emma Jun 2 '15 at 7:50
  • $\begingroup$ or are they the same? $\endgroup$ – Emma Jun 2 '15 at 7:50
  • $\begingroup$ Well, the issue is you have no constant in your regression, so I'm not sure they would be exactly the same. With a constant, then we know that $\hat{\beta}=\frac{cov(x_1,y)}{var(x_1)}$. Therefore, by the same logic as above we get $\hat{\beta} \rightarrow \beta_1 + \beta_2\frac{cov(x_1,x_2)}{var(x_1)}$. I chose to use covariances in my answer because I think it shows more clearly the conditions you need to satisfy. But, the middle line in the solution is the more general answer. $\endgroup$ – Greg Jun 2 '15 at 12:48

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.