[Gretl-devel] Re: Stringified series cannot be over-written

On Thu, 5 Dec 2019, Artur Tarassow wrote: > Am 04.12.19 um 23:19 schrieb Allin Cottrell: >> On Wed, 4 Dec 2019, Artur Tarassow wrote: >> >>> I am bit puzzled why one cannot assign new values to a stringified >>> series (using current git). Why should be the behavior differently >>> compared to numerical series? >>> >>> <hansl> >>> set verbose off >>> nulldata 2 >>> >>> series foo = seq(1,2)' >>> strings s = defarray("A", "B") >>> stringify(foo, s) >>> print foo >>> >>> series foo = 1 # "Cannot overwrite entire string-valued series >> </hansl> >> >> Well, the intent is to preserve the integrity of such series. We >> obviously couldn't have anything like series foo = log(foo) and still >> have a string-valued series. >> > Ok, I understand this. Thanks. > >> In the case of assigning a scalar value (that's within the range of >> the existing encodings) to the entire series, as above, no harm would >> be done. >> >> The question then is whether we just issue a blanket ban on assigning >> to an entire string-valued series (the case at present), or take the >> trouble to code an observation-by-observation check on the validity >> of such an assignment. > > What about a compromise that for a string-valued series one cannot > change separate entries but assigning a complete new value, e.g. by > doing <series foo = 1>, is allowed?  I was think of something like: > > <> > series y = normal() > series foo = seq(1,2)' > strings s = defarray("A", "B") > stringify(foo, s) > > foo[1] = "C"       # not allowed > series foo = 99    # allowed > series foo = y     # allowed as y is of same length as foo > </> That's the opposite of what we have now: please take a look at chapter 15 of the Guide on string-valued series, especially section 15.3. The guiding ideas are that (a) a string-valued series (internally) is an integer-valued encoding of a set of strings, and (b) once a given series is string-valued it should stay that way. From that perspective "series foo = y", where foo is supposed to be string-valued and y = normal() cannot possibly work. It's fairly easy to check that assignment to an individual observation of a string-valued series works OK, not so easy to check assignment to an entire such series.

>> We could do the latter but it didn't seem worthwhile to me when we >> first introduced such series. It wasn't clear to me that there was >> any real use case -- do you have one? > > I have a really huge panel data set ( > 4GB) and run multiple > estimations by means of a loop-block. Each round I export some list of > series and want to construct a series "foo" holding some > round-specific string value(s). Currently, I need to generate each > round some new series "foo_$i$. > > As deleting a series within a loop-block is also not allowed, I could > not think of an alternative solution to this. I'm not sure I understand the case, but you can use the --force option if you're confident this isn't going to destroy the integrity of the loop: (Sorry, that should be documented!)