Bandwidth Selection in Bai and Perron (2003)

[alexecon]

In computing the break ranges with robust covariances, the code in @BaiPerron assumes that the bandwidth for the quadratic spectral kernel grows at a rate of 1/3; it also finds the optimal bandwidth assuming a constant parameter of 0.75:

Code: Select all

if robust {
   mcov(lags=.75*(endr-startr+1)^.333,lwindow=quadratic) startr endr %resids
   # %rlfromtable(%eqntable(eqnshift))
   compute omega=%cmom/%nobs
   compute omega1=%cmom/%nobs
}

Why are these numbers assumed?

[TomDoan]

That's a standard "rule of thumb" for QS. Bai-Perron use a data-determined method, which seems like a really bad idea when those calculations will sometimes be done with very small numbers of observations. I would also note that if the residuals have strong "HAC" properties, then it's not really clear that the Bai-Perron method of finding breaks (which relies upon OLS) is a good choice.