Propensity Score Matching

[Pietro Lombardi]

I want to follow a paper and I am struggling with propensity score matching. Essentially, my basic regression to analyze whether a supplier_change has an influence on the quality of a produced car is:

regress  car_quality  supplier_change  working_hours price supplier  i.year i.state, vce(cluster ID)

The three steps indicated in the paper are

• 1) Create propensity score of a supplier_change with all of the control variables (working_hours price supplier)
• 2) Match each treatment observation to the single closest control observation from the same year and state and a caliper distance of max. 0.1
• 3) Re-run the regression with matched sample (although other posts indicate this is problematic).

So far I have tried for step 1):

probit supplier_change working_hours price supplier 
predict pscore

For step 2) I have discovered that many use psmatch2 but I could not get the matching to work...
Also, I think I cannot combine steps 1 and 2 with psmatch because of the calculation requirements.

Help is highly appreciated!

[George Ford]

Stata has a wide variety of PS matching methods.

teffects is useful for most purposes..

teffects psmatch car_quality (supplier_change) (working_hours price supplier i.year i.state)

[Pietro Lombardi]

teffects does indeed offer many advantages. However, I am not sure, whether it is possible to include all of the steps in the paper with teffects.

I ended up with the following solution for my desired three-step-approach:

Step 1: Propensity Score

logit supplier_change working_hours price supplier  predict pscore

Step 2: Matching based on propensity score based on closes observation from the same state and year (and a caliper distance of max. 0.1)

egen group_id = group(year state)
su group_id, meanonly
gen weight = .

forval i = 1/`r(max)' {
psmatch2 supplier_change if group_id == `i', pscore(pscore) caliper(0.1)
replace weight = _weight if group_id == `i'
}

Step 3: Re-run regression with matched sample

 regress car_quality supplier_change  working_hours price supplier i.year i.state [fweight= weight], vce(cluster ID)