3
$\begingroup$

Suppose that I have a multivariate random walk:

$X_{t+1} = X_t + \epsilon_t$ where $\epsilon_t \sim N(0,\Sigma)$

Estimating the covariance matrix $\Sigma$ is straightforward from first differences $X_{t+1}-X_t$ using the sample covariance.

Instead if I would like to estimate the covariance directly from the levels $X_0,...,X_n$ again using the sample covariance but this time on the levels and then using the relation $Cov(X_n-X_0) = n \Sigma$ to find $\Sigma$ by simply dividing by $n$, I have a feeling that I will underestimate the variances hence overestimate the correlations. I also cannot decide whether to subtract the unconditional mean $X_0$ or sample mean of the levels during the calculation of covariance matrix.

Basically does $\hat{Cov}(X_n-X_0) = \frac{1}{N} \Sigma_1^N(X_n-X_0)(X_n-X_0)^T$ or the same formula where $X_0$ replaced with sample average $\hat{\mu} = \frac{1}{N} \Sigma_1^N X_i$ makes more sense (in this scenario we divide by $N-1$ because of degrees of freedom but that's not the issue)?

Would I be underestimating the variances and overestimating the correlations?

Obviously my intention is not to estimate the covariance of innovations in this manner where there is already a way to estimate them using the difference series. My main question is how this is affecting OLS estimates of multiple regression when regressors have unit roots, as

$\beta = \large{\frac{Cov(X,y)}{Cov(X)}}$

doesn't seem to be so well defined anymore? Yes they are super consistent in case of a cointegration, but that's a little after the fact in terms of estimation.

$\endgroup$
  • $\begingroup$ You may want to have a look at this paper: sciencedirect.com/science/article/pii/0304407686900011 $\endgroup$ – Matthias Schmidtblaicher May 11 '17 at 19:13
  • $\begingroup$ @MatthiasSchmidtblaicher I think that's the seminal paper of Phillips, but waaay to technical for me. $\endgroup$ – Cagdas Ozgenc May 11 '17 at 19:16
  • $\begingroup$ @CagdasOzgenc Can you give some explanation of why you want to do this, and why you are not comfortable with the standard approach? Also, you may consider the simpler univariate case before adding complexity with the multivariate processes. $\endgroup$ – John May 17 '17 at 17:51
  • $\begingroup$ What is the relation between $n$ and the sample size? Because if $n$ is the last observation in your sample, then you have just one (vector-valued) observation to calculate the empirical variance matrix from. Also, as $n$ goes to infinity, your estimator should be getting wilder, as it is the variance of a random walk which does not exist in population. $\endgroup$ – Richard Hardy May 17 '17 at 18:11
  • $\begingroup$ @RichardHardy I have n samples. Each sample is a vector. $\endgroup$ – Cagdas Ozgenc May 17 '17 at 19:45
1
+50
$\begingroup$

Consider the following single-equation model:

$$y_t = \rho y_{t-1} + \beta x_t + u_t,\;\;\; t=0,...,T,\;\;\;y_0=0 \tag{1}$$

with the following stochastic assumptions:
1) The $x$'s come from an i.i.d. sample
2) The $x$'s are strictly exogenous to the disturbance
3) The disturbance $u_t$ is white noise.
4) It follows but let's add it for clarity, that current $x$'s (sometimes called "forcing factors" in the ARMA-X literature) are independent of past dependent variables also.

Denote also $M_x = I - X(X'X)^{-1}X'$, with appropriate dimensions to fit the lagged structure, which is the residual maker matrix associated with the regressor $X$, a symmetric and idempotent matrix.

On the side, denote $S_t = \sum_{i=1}^t x_i,\;\;\; U_t = \sum_{i=1}^t u_i$. Then, if in reality $\rho =1$ (a unit root), we could re-write the model for $t-1$ as

$$y_{t-1} = \beta S_{t-1} + U_{t-1} \tag{2}$$

Keep that.

By partitioned-regression results (which are algebraic and do not depend on stochastic assumptions), we have that

$$\hat \rho_{OLS} = \left(\mathbf y_{-1}'M_x\mathbf y_{-1}\right)^{-1}\mathbf y_{-1}'M_x\mathbf y = \left((M_x\mathbf y_{-1})'(M_x\mathbf y_{-1})\right)^{-1}(M_x\mathbf y_{-1})'M_x\mathbf y \tag{3}$$

..the second equality due to the symmetry and idempotency of $M_x$.

Note that $\mathbf y_{-1}'M_x$ are the residuals (in row-vector form) from regressing $y_{t-1}$ on ${x_t}$, in an OLS regression $y_{t-1}= ax_t + v_{t-1}$. But given our stochastic assumptions, and as it should be obvious from $(2)%$, we have $a=0$, and the OLS regression will have no problem in detecting that. Which means that the residuals from this regression will tend to equal the dependent variable : $\mathbf y_{-1}'M_x \to \mathbf y_{-1}$.

It follows that

$$\hat \rho_{OLS} \to \left(\mathbf y_{-1}'\mathbf y_{-1}\right)^{-1}\mathbf y_{-1}'(M_x\mathbf y) \tag{4}$$

Moreover, $M_x\mathbf y$ are the residuals from regressing $y_t$ on $x_t$, and, given our stochastic assumptions it is easy to conclude that $M_x\mathbf y \to \rho \mathbf y_{-1} + \mathbf u$.

With this we obtain

$$\hat \rho \to \rho + \left(\mathbf y_{-1}'\mathbf y_{-1}\right)^{-1}\mathbf y_{-1}'\mathbf u \to \rho \tag{5}$$

...because given our stochastic assumptions, current distrubances are orthogonal to past dependent variables. So we see that, irrespective of whether $\rho=1$ or not, the OLS estimator will estimate it consistently.

What about the $\beta$ coefficient? Aplying again partitioned regression results, and analogous notation as before, we have that

$$\hat \beta_{OLS} = \left(X'M_{-1}X\right)^{-1}X'M_{-1}\mathbf y \tag{6}$$

Using previous reasoning $X'M_{-1} \to X'$. Using the specification, we have

$$\hat \beta_{OLS} \to \left(X'X\right)^{-1}X'\mathbf y = \rho\left(X'X\right)^{-1}X'\mathbf y_{-1} + \beta + \left(X'X\right)^{-1}X'\mathbf u$$

With the same steps and reasoning as before we get that the first and third term will go to zero (always given our stochastic assumptions), and so

$$\hat \beta_{OLS} \to \beta \tag{7}$$

I have not touched on the matter regarding the distribution of the estimator.

I will return as regards the rate of convergence and wether the estimator is superconsistent or not (and if it is for which coefficients).

$\endgroup$
  • $\begingroup$ It will take sometime for me to process all this. Can you summarize the main points? Also how will your answer change in case $x_t$ is an integrated process? In experiments I am running as long as $y_{t-1}$ is there it doesn't matter much if $x_t$ is IID, or integrated. $\endgroup$ – Cagdas Ozgenc May 19 '17 at 8:05
  • $\begingroup$ Basically in the setup you describe both $\beta$ and $\rho$ consistenly estimated irrespective of wheter $\rho$ is 1 or not. Is this the conclusion?. $\endgroup$ – Cagdas Ozgenc May 19 '17 at 8:07
  • $\begingroup$ @CagdasOzgenc Yes they are both consistently estimated. $\endgroup$ – Alecos Papadopoulos May 19 '17 at 14:10
  • $\begingroup$ @CagdasOzgenc Thanks for the heads-up. During the weekend. $\endgroup$ – Alecos Papadopoulos Sep 8 '17 at 11:10

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.