Newey-West HAC standard errors smaller than regress, robust ones

[Michele Scotto]

Hi all,

I am running an ARDL model in the form:

 reg y l.y x1 l.x1 x2 l.x2

After the estimation I check for serial correlation with both the 'estat durbinalt' and 'estat bgodgrey' postestimation commands for lags from 1 to 3. I do not find serial correlation at lag 1 and 2, but I do at lag 3. So I re-estimate the model using:

 newey y l.y x1 l.x1 x2 l.x2, lag(3)

My question is: why I get smaller standar errors (and coefficient breaches the significance level) with the HAC s.e.? Should not they be more conservative in order to correct for serial correlation and heteroskedasticity?

On a different line, assuming there is no serial correlation (so for instance at lag 1 or 2 and I do not run a test for lag 3), why would 'newey' lead to different s.e. than 'regress, robust'?

Thank you very much for your help. Best,

Michele

[Phil Bromiley]

The newy documentation says "If lag(0) is specified, the variance estimates produced by newey are simply the Huber/
White/sandwich robust variances estimates calculated by regress, vce(robust);". I haven't checked, but I assume this is correct. newy isn't just more conservative - it is a different formula.

If you really want to understand this, look at the Methods and Formulas for the two estimators. By the way, if you have serial correlation, then you may need to worry about the lagged dv.