its a reasonable, practical thing to do, tho of course its a hack.
if there are very few in a category being consolidated, then its not going
to matter much if you consolidate. if there are more AND the processes
driving those events are very different, then you have a composite
category that needs different specifications but gets only one. so you
can cause problems for yourself.
incidentally, altho there is no rare events multinomial logit, it is
completely feasiable to create one. most of the same math for the regular
logit applies. so even if you don't have time now, its something you
might want to consider doing. you could get others to use it.
Gary
On Wed, 2 May 2007, Anya Vodopyanov wrote:
hey guys,
does anyone know if treating a rare outcome event as a non-rare outcome event in
a multinomial logit bias coefficients? i have a 4-category logit (and a 5th for
the
baseline), and 2 of the categories have very few observations. i am
considering consolidating 1 of the problematic categories with another (which is
something the authors do for one of their tables as well so i
think it's kosher) and taking out the other problematic category altogether
(it's
the baseline in my model). substantively, i think the transformation would be
ok, but my question is whether it makes sense methodologically. specifically,
is it: A) worthwhile - ie are rare-event outcomes a problem (in terms of
causing
bias etc), and B) does it sense given that there is no rare events *multinomial*
logit?
thanks for any suggestions,
anya
_______________________________________________
gov2001-l mailing list
gov2001-l at
lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l