Am 12.05.2023 um 13:26 schrieb GIULIO PALOMBA:

Dear all,

I would like to report a problem with the "StrucBreak" package.

If I insert a dummy variable as a regressor, collinearity arises in (at least) one regime, and the estimation fails.

Thanks for the report, Giulio. First of all, this case needs to be handled instead of running into problems internally, of course.

THANK YOU, SVEN!

I WILL ANSWER POINT BY POINT.

My spontaneous reaction would be to disallow dummy variables such as this for the set of variables with breaking coefficients (Z), but I'm open to counter-arguments.

JACK ANDI HAVE TALKED ABOUT THIS AND WE THINK THE BEST IDEA IS TO USE THE dropcall FUNCTION IN THE DEFINITION OF THE REGRESSORS AT EACH REGIME (OR SUBSAMPLE).

Also, it's not entirely clear where to draw the line and how to define the criterion which variables are acceptable. Do you remember what was happening in your original Ox code?

IN MY OPINION THE USE OF DUMMY VARIABLES, AND IN GENERAL OF POTENTIALLY COLLINEAR EXPLANATORY VARIABLESIN AT LEAST ONE REGIME, SHOULD BE ALLOWED. SO I DON'T THINK THERE IS A NEED TO DRAW A LINE, BUT I TOO AM OPEN TO DISCUSS VIABLE ALTERNATIVE IDEAS.

UNFORTUNATELY, THIS WAS NOT FIXED IN OUR ORIGINAL OX CODE.

Apart from that, I also tried to put the dummy into the non-breaking set X, where it would make more sense I think. But not surprisingly, I got the same error, so again, this needs fixing.

AH, OK! THIS IS NOT SURPRISING.

thanks

sven