2
$\begingroup$

I'm estimating $y_i= \beta_1 \times x_{1i} + \varepsilon_i$ on a time series on $y$ and $x$, so in presence of Heteroschedasticity and Autocorrelation. My model does not include any intercepts. Are the Newey West standard error calculated on such model appropriate (in other words, the NW correction requires a full model with intercept)? I am using R sandwich and lmtest packages.

$\endgroup$
3
$\begingroup$

Here are two screenshots (if that is not OK I hope moderators will let me know) of the papers by Newey and West (Review of Economic Studies 1994) and Andrews (Econometrica 1991), which are follow up papers to the more high-level Newey/West paper:

enter image description here

Andrews:

enter image description here

As you can see, having constant plays no role. What matters is that the coefficient estimates for which you want autocorrelation-robust standard errors are themselves ($\sqrt{T}$-)consistently estimated, so, more or less, that they are computed from a correctly specified model.

Now, if your model is such that it indeed does not need a constant (which may happen), that should be fine.

$\endgroup$
  • 2
    $\begingroup$ Thanks, Christoph, I agree. The input that the sandwich package needs are the scores for a model with the "usual" central limit theorem, i.e., , consistency and asymptotic normality at rate sqrt(n). The scores are just the observation-wise contributions to the gradient (first-order condition). All of this also applies to linear models without constant so I wouldn't see any potential problems. $\endgroup$ – Achim Zeileis Mar 18 '15 at 19:08

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.