Consider the following regression equation * \begin{equation} \label{eq:1} C_{t} = \beta_{1} + \lambda C_{t-1} + \epsilon_{t} \end{equation} and let \begin{equation} \label{eq:3} \epsilon_{t} = \rho\epsilon_{t-1} + u_{t} \end{equation} where the error component \begin{equation} u_{t}\end{equation} is iid with mean 0 and constant variance, and \begin{equation} \label{eq:4} E(u_{t}\mid C_{t-1},\epsilon_{t-1}) = 0 \end{equation} Is the OLS estimator of the coefficients in (*) unbiased and consistent under this error specification? Why?


1 Answer 1


By the definition of OLS estimator we can rewrite the $\hat\lambda$ as
$\hat\lambda=argmin \sum_{i=1}^{n}\epsilon_t^2$
From first order condition, we get $\hat\lambda=\frac{\sum_{i}^{n}(C_{t-1}-\bar{C}_{t-1})(C_t-\bar{C_t} )}{\sum_{i}^{n}(C_{t-1}-\bar{C}_{t-1})^2} $
You can get more simple expression as $\hat\lambda=\lambda+\frac{\sum_{i}^{n}(C_{t-1}-\bar{C}_{t-1})(\epsilon_t-\bar{\epsilon_t})}{\sum_{i}^{n}(C_{t-1}-\bar{C}_{t-1})^2}$
Since $E[(C_{t-1}-\bar{C}_{t-1})(\epsilon_t-\bar{\epsilon_t})]=0$, $E(\hat\lambda)=\lambda$.(unbiased)

Now apply Law of Large Numbers to show consistency.

$\hat\lambda =\frac{\sum_{i}^{n}(C_{t-1}-\bar{C}_{t-1})(C_t-\bar{C_t} )}{\sum_{i}^{n}(C_{t-1}-\bar{C}_{t-1})^2}=\frac{1/n\sum_{i}^{n}(C_{t-1}-\bar{C}_{t-1})(C_t-\bar{C_t} )}{1/n\sum_{i}^{n}(C_{t-1}-\bar{C}_{t-1})^2}$
By LLN, $\hat\lambda =\frac{Cov(C_t, C_t-1)}{Var(C_t-1)}=\frac{Cov(\beta_1 +\lambda C_t-1 +\epsilon_t, C_t-1)}{Var(C_t-1)}=\frac{\lambda Var(C_t-1)+Cov(\epsilon_t, C_t-1)}{Var(C_t-1)}=\lambda $.(consistent)

  • $\begingroup$ what does lln mean? $\endgroup$ Commented Apr 22, 2022 at 4:29
  • $\begingroup$ @JohnWilliams LLN means Law of Large Numbers $\endgroup$
    – guest
    Commented Apr 22, 2022 at 4:37
  • $\begingroup$ could you explain why its unbiased and consistent with more words? $\endgroup$ Commented Apr 22, 2022 at 4:47
  • $\begingroup$ @JohnWilliams I edited my solution. You need to know the concept of LLN to prove consistency. $\endgroup$
    – guest
    Commented Apr 22, 2022 at 5:14
  • $\begingroup$ The AR(1) is not unbiased for the lagged coefficient even in the case where the error term is purely iid. See this link for a couple of different and insightful explanations. The link doesn't cover whether it's consistent for the AR(1) error term case but I don't think it is. stats.stackexchange.com/questions/240383/… $\endgroup$
    – mark leeds
    Commented Apr 25, 2022 at 15:53

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.