This is not homework. I am a frequent user on math.stackexchange, but I am learning a bit about time series models and came across this example. Any ideas would be greatly appreciated.

A linear regression model was fit to some time-series data by ordinary least squares. The residuals from the fit were then used to create two new variables, namely $E$ with values $\hat{e}_2,...,\hat{e}_n$ and $E_1$ with values $\hat{e}_1,...,\hat{e}_{n-1}$. A linear “regression through the origin” was then run with E as the dependent variable and $E_1$ as the predictor. The slope estimate was 0.412 with a standard error of 0.133. Assume that the $e_t$ follow a standard AR(1) model.

  1. Estimate the first order autocorrelation $\rho$ of the AR model.

  2. Can the output be used to obtain a valid standard error for the estimate in 1?

Is the answer to #1 above just the slope of the regression E~E_1?


Yes to all three questions. But it is more efficient to estimate the AR error model at the same time as the linear regression model using GLS or MLE.

  • $\begingroup$ I agree in practice, which is why this exercise I found seemed rather strange. $\endgroup$
    – Justin
    Jul 23 '12 at 1:52

The $\rm AR(1)$ model is given as $X_n= \rho X_{n-1} + e_n.$ The parameter $\rho$ is normally estimated by conditional least squares. If the model is correct the eis have mean $0$ and variance $\sigma_e^2.$

The parameter $\rho$ is $$\rho=\frac{{\rm Cov}(X_n, X_{n-1})}{{\rm Var}(X_n)}.$$

When the estimate of $e_n$ is paired with the estimate of $e_{n-1}$ the slope is theoretically $0$ but will not be exactly zero because the estimates of residuals are based on the estimate of $\rho$.

The answer to 2 is yes.


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.