Re: [Gretl-devel] restricted coeff (in system)
On Mon, 10 Jun 2013, Sven Schreiber wrote:
I'm noticing a slight inconsistency (without practical
implications)
when I restrict coefficients in a system. Here's a partial output:
In the first equation the restricted term (b[1,4] = 0) shows up like this:
ECT2sub2_1 0.00000 2.76876e-09 2.716e-07 1.0000
while to give an example for the expected behavior I get in other
equations (b[2,2] = 0):
ECT1sub2_1 0.00000 0.00000 NA NA
I suspect there may be something buggy going on, since 1e-9 should be
bigger than machine precision (while for the t-ratio a value of 1e-7
just seems to indicate a side effect of a coeff around 1e-16 in
combination with that standard error value).
I don't think this betokens a bug. 2.77e-09 may be smaller than
machine precision on a single calculation, but it's well within
bounds for the chain of calculations that results in a standard
error.
Gretl performs "sanitization" of very small standard errors for
printout purposes (i.e. setting them to nominal zero), but we're
deliberately conservative in doing so. We don't want to pretend that
a standard error is zero if it might not be; the downside is that
we'll sometimes print a non-zero standard error when in principle it
really should be zero.
Allin