Propensity score matching and DID

[Ashani Abayasekara]

Hi everyone,

I'm trying to estimate the effect of car plant closures on affected workers' income and wages, by comparing outcomes between car plant workers and similar unaffected workers in other manufacturing industries in a DID framework. To ensure that I am comparing workers in similar occupations in the treatment and control groups, I want to match these workers based on their occupation codes and age, prior to running DID models.

I am unsure how to proceed with propensity score matching. I was thinking of using the user-contributed 'psmatch2' command for this purpose. My initial code would be:

psmatch2 treat occ age, out(income wage) common

were treat=1 for car industry workers and =0 for other manufacturing workers and occ refers to the occupation code.

How would I proceed after this? Would I limit the DID regressions to observations for which there is common support (_support==1)? Or do I need to consider the estimate propensity scores, using the _pscore variable generated by Stata? My current DID regression is as follows:

xtreg income i.year treat08-treat20, fe i(id) r

where treat08-treat20 are indicator variables equal to one if the individual was in the treated group in the particular year.

Would I just add in a condition to this to limit the sample to those matched based on the matching exercise? Relatedly, I'm also unsure of which option to use, among the many matching options available under psmatch2.

Any guidance would be much appreciated. Unfortunately I am unable to share the data which is confidential.

Many thanks.
Ashani

[George Ford]

with different treatment dates, you'd probably want to use csdid (which includes ps matching).

[Ashani Abayasekara]

Hi George,

Thanks for your message. To clarify, the treatment dates are not necessarily different, but in this setting I am unsure of the exact start of treatment for different individuals. Hence, I'm estimating changes in effects in each year and then comparing them against a base year. Would you till recommend -csdid- under this setup?

[George Ford]

This may be useful. You can weight by the propensity score.

webuse cattaneo2 , clear

teffects ipwra (bweight prenatal1 mmarried mage fbaby) (mbsmoke foreign alcohol mage medu fage fedu , logit)  

psmatch2 mbsmoke mmarried c.mage##c.mage fbaby medu , logit
regress bweight mbsmoke prenatal1 mmarried mage fbaby [aw=_pscore] , r

logit mbsmoke foreign alcohol mage medu fage fedu
qui predict double ps if e(sample)
qui gen double ipw = 1.mbsmoke/ps + 0.mbsmoke/(1-ps)
reg bweight mbsmoke prenatal1 mmarried mage fbaby [aw=ipw] , r

kmatch ps mbsmoke mmarried c.mage##c.mage fbaby medu , generate
reg bweight mbsmoke prenatal1 mmarried mage fbaby [aw=_KM_ps] , r

drop _KM*
kmatch eb mbsmoke mmarried c.mage##c.mage fbaby medu , generate
reg bweight mbsmoke prenatal1 mmarried mage fbaby [aw=_KM_mw] , r

g magesq = mage*mage  //em won't take the interactions
drop _KM*
kmatch em mbsmoke mmarried mage magesq fbaby medu , generate
reg bweight mbsmoke prenatal1 mmarried mage fbaby [aw=_KM_mw] , r