Choosing lag order in Newey west regression, panel data ( fixed effects)

[Alexandra TD]

HI, how do I choose the lag order in newey2 (newey west for panel data) ?

If I take into consideration the test i attached, this means i could consider as lag in the regression for panel data with newey west covariance matrix, lag 10 or 17 (18 )?

Thank you,
Alexandra

[Mark Schaffer]

There are methods for automatic bandwidth selection using the Newey-West (aka Bartlett kernel) approach, but they don't work the way you suggest. I don't recall seeing anything specific to panel data, but you could estimate each panel separately using automatic bandwidth selection (various Stata command support this), see what the recommended bandwidth is, and then use that. This assumes that you have a large T panel setting (T->infinity) since this is needed for the Newey-West approach to work.

Alternatively, if you have reason to believe that the autocorrelation is MA(q) and dies out after q lags, then the right kernel to use is the truncated kernel with bandwidth=q, as discussed in Chapter 6 of Hayashi's Econometrics.

[Alexandra TD]

Dear Mark, I have for the initial regressions about 70 * 5 ( -14 panels) data, for the last ones about 300* 5(-14 panels) data.
Thank you, I will see if I get any more ideas, considering your suggestions too.
Thank you,
Alexandra