If I understand correctly, the kmatch, bwidth() argument refers to the half kernel width in the "distance" space of either md or ps, but in a relative sense (I guess in standard deviation units). That is standard and usually desired behavior, but means that we'll never have the behavior as for em: Matching on set a b c first, and then matching on a b c again using common support units only, there should be no $D_A$ and $D_B$. At the moment, keeping bwidth constant, there are new unmatched units because the distance space in reestimated and therefore the kernel width.
Issue would be solved with absolute specification of distance, but I guess this is not something we can do. So, this is just a thought dump for now.
If I understand correctly, the
kmatch, bwidth()
argument refers to the half kernel width in the "distance" space of eithermd
orps
, but in a relative sense (I guess in standard deviation units). That is standard and usually desired behavior, but means that we'll never have the behavior as forem
: Matching on seta b c
first, and then matching ona b c
again using common support units only, there should be no $D_A$ and $D_B$. At the moment, keepingbwidth
constant, there are new unmatched units because the distance space in reestimated and therefore the kernel width.Issue would be solved with absolute specification of distance, but I guess this is not something we can do. So, this is just a thought dump for now.