Questions on Chow Break Point and Quandt Andrews

[Mila]

Hi I'm currently using the Chow Break Point and Quandt Andrews tests to check for breaks in my time series regressions.
Some of my regressions are based on Newey West HAC, and some on default standard errors. Some contain dummy variables.
I'd just like to seek some clarification:
i. Both tests are not applicable if I use Newey West standard errors? I'm not "allowed" to performed the QA test, but I can still perform the Chow test, albeit I can't specify specific coefficients to be tested.
ii. For default standard errors, I can still use both tests, even if I have dummy variables? as long as I'm only interested in a subset of coefficients.
I'm essentially listing out what I've done, but hoping for some clarification on whether my understanding is right.

Many thanks.

[Mila]

Correction:
If I have impulse dummy variables in my equations, the quandt andrews test may or many not work (singular matrix in some instances but not in others). Could someone explain why this happens? I had assumed that both the Chow Break Point and Quandt Andrews tests would drop the impulse dummy variables that are not in the subsamples. Is this not the case? How DO the tests handle impulse dummy variables in Eviews?

Say I have ls y y(-1) x z d. x,z exogenous variables and d impulse dummy (takes 1 in one period, 0 otherwise). I'm only interest in whether x varies over time, and only test for this variable. I'm using Eviews 7.2.

If anyone could clarify, I'd be really grateful. Many thanks!

[EViews Gareth]

Dummy variables are treated the same as any other variables. That's why you can often get singular matrix errors (since in a particular sub-sample you get perfect colinearity between the dummies and/or the constant).

I'll let someone else go into the details on HAC, but I'll point out that the issues surrounding HAC are based on how you calculate the kernel - do you use a different kernel estimate for each sub-sample (in which case you might run out of observations etc...), or do you fix the kernel for each sub-sample based upon the full sample.