Combining exact and propensity score matching

[Tslil Aloni]

Hi all,

Would appreciate help with matching on propensity score within bins.

I have either:

teffects psmatch (wage) (treatment i.year i.industry var1 var2 var3, logit), nn(1)

or

psmatch2 treatment i.year i.industry var1 var2 var3, outcome(wage) n(1) noreplacement logit

But I would like the matching to occur at the year X industry level. So that each treated observation in year t industry i is matched to 1 observation in year t industry i.


Thanks!!!

[Sandra Bloem]

I am just a student so I can't give you any solid advice, but how I've been taught this is by creating a grouping variable over year and industry, and then "scale" the propensity scores such that each group has it's own designated interval with propensity scores. Then you can use a caliper width such that an observation in group 1 is not matched to an observation in group 5 for example.

egen groupyrind = group(year industry)
probit treatment i.year i.industry var1 var2 var3
predict double pscore if e(sample)
generate double pscore2 = groupyrind*10+pscore
rsort
psmatch2 treatment, outcome(wage) n(1) noreplacement caliper(2) pscore(pscore2)

[Tslil Aloni]

Oh! That's clever. Thanks a lot for your help [=

[Felix Bittmann]

Another option is using kmatch, like

kmatch ps treatment var1 var2 var3 (wage), wor nn(1) exact(year industry)

However, I suggest using kernel matching with (potential) multiple controls, like:

kmatch ps treatment var1 var2 var3 (wage), exact(year industry)

[Ondrej Dvoulety]

Hello, I tried to use "exact function", but it says "option exact() not allowed" any hint what may be wrong?

Another option is using kmatch, like

kmatch ps treatment var1 var2 var3 (wage), wor nn(1) exact(year industry)

However, I suggest using kernel matching with (potential) multiple controls, like:

kmatch ps treatment var1 var2 var3 (wage), exact(year industry)

[Sandra Bloem]

Ondrej Dvoulety

I believe the option is –ematch– instead of –exact–

[Felix Bittmann]

Indeed, Sandra is correct. It is ematch(varlist). My apologies.

[Ondrej Dvoulety]

Thanks a lot for your help, both of you!

[Phyu Zin]

Hi Sandra & Felix

Can you advise to subset the matched cohort after Kernel matching?

Thanks
Phyu