I've noticed some things which may be quirks in relation to the 'mle' block.
- If I wrongly insert several "params" lines I get a weird data error
instead of a syntax error, so it took me a while to realize my mistake.
Could that be made clearer?
- In the guide in Listings 26.1 and 26.2 and throughout other places in
ch. 26 the "genr" keyword is used. I think we want to replace that with
series/scalar etc. - In that context there the rationale for using it is
not quite clear to me, since it is surrounded by the cleaner use of
series/scalar... Anything I'm missing?
- I've been struggling without success to tell the mle command what the
parameters should be called. The background being that I need to specify
a single composite vector to accommodate several specifications. In the
guide (ch.26) I saw a similar problem, and there the workaround was to
silence mle and instead semi-manually invoke the modprint apparatus. But
wouldn't it be more efficient if simply the wanted names could be
injected, replacing the generic stuff that is in the $parnames accessor?
Also because for modprint the additional stats then have to be put
- Also relating to 'modprint': no on-the-fly construction of the matrix
input seems possible, i.e. directly using $coeff~$stderr didn't work.
- Back to mle: its $model bundle has a "rho" member of value 1.#QNAN
which seems spurious in that context. (Fallout from the biprobit rho??)
Well, so much for that.
this happens on a Jan-30th snapshot with the new ARprobit package: From
the package's GUI output window I click on the bundle icon and select a
scalar (t2 for example) to be saved to the session. Everything seems to
go fine, but in the end in the session icon view under "scalars" nothing
In contrast, saving the "bic" member works. Could it be related to t2
(and others) being shown as "int", not "scalar"?
I'm getting a matrix non-pos-def error when using the ivpanel addon. It
concerns the function matrix_tsls, namely:
matrix nonzero = (abs(R[diag]) .> 1.0e-12)'
if minr(nonzero) == 1
V = invpd(XQ * XQ')
where R is from a QR decomposition. I've checked R, and indeed there is
one value just slightly above (in abs) 1e-12, it's -1.4099e-012. Then
the smallest eigenvalue of XQ*XQ' comes in slightly negative at
-4.8615e-011 which explains the error.
I guess there should be some error handling in the package -- I also
tried to make the numerical check stricter to reach the alternative code
branch, but this just lead to a collinearity problem down the road
So the solution is apparently: remove the same regressors that Jack's
felogit package removes automatically as (numerically) collinear. At
least it works in my case. Perhaps the relevant code could be moved to
ivpanel, because it is an addon, and then felogit could depend on it for
(BTW: matrix_tsls is another example of duplication between commands
--tsls-- and a wanted function-style interface. Or to put it
differently: maybe this would be something as a public thing for
'extra'? But that's a different topic.)
I've noticed that a couple of dialog windows for graphing only have
generic names. For example choosing Boxplot from the menu yields a
dialog which just says "gretl: define graph", and if you hadn't clicked
Boxplot (and knew that you didn't miss) you wouldn't have any idea that
this is going to be a boxplot. In contrast, the Q-Q plot window has a
descriptive title -- is this coming from inside gui/library.c ?
The relevant matches seem to live in gui/callbacks.c. OK to
differentiate these title strings and make them more specific? I get six
matches, but some of them are not quite clear to me in terms of the
context, I'm afraid.
BTW, let me mention also a different issue: when selecting two series in
the main variable list this selection seems to be remembered for the
graph dialogs, but not when a single series is marked; one then has to
search and select again (inside the dialog).
here's the recipe:
- open credscore.gdt
- mark "age" as discrete
- plot the frequency distribution
-> the result are relatively thick bars (in gnuplot "impulses")
- then edit something, e.g. the font size
- apply: and the bars become very thin, impossible in the GUI to make
them thick again (but feasible in the underlying gnuplot code in the
session icon view)
This is with the dark2 theme.
I've found quite unusual error which happens only when I call a package
which uses MPI in my script: we get SIGSEGV in purebin_write_data() from
'purebin.c' since pointer to 'list' is NULL (and gretl crashes on nv =
BTW I'd like to ask two guestions regarding C level:
1. Which error type gretl should return in such a case, E_DATA?
2. What is the most efficient compilation scheme (instead of make all
install) when I change something in, f.e., 'purebin.c'? I'm asking
because after 'make -B plugin lib cli' gdb claims that
cli/.libs/gretlcli is still older than a source file?