regression after teffects psmatch ..., nn(3)

[Kurt Wichtelsacker]

Dear Statalist members,

I have the following problem and cant find an answer by myself: I want to run a propensity score weighted regression by using the command teffects psmatch to estimate the propensity scores.
For the default value of nearest neighbor = 1 it is no problem since the help of the guide http://www.ssc.wisc.edu/sscc/pubs/stata_psmatch.htm. But if I want to estimate the propensity scores with nearest neighbor values of 3 or 5 I dont know how to adjust the algorithm:


teffects psmatch (y) (treatment x x x x x x), atet gen(match) nn(3 or 5)

predict ps0 ps1, ps

predict y0 y1, po
*Because observation 1 is in the control group, y0 contains its observed value of y. y1 is the observed value of y for observation X's match, observation XX

predict te
*The treatment effect is simply the difference between y1 and y0. You could calculate the ATE yourself (but emphatically not its standard error) with:

sum te

*and the ATET with:

sum te if treatment==1

*regression preparation

gen ob=_n //store the observation numbers for future use
save fulldata,replace // save the complete data set

keep if treatment // keep just the treated group
keep match1 // keep just the match1 variable (the observation numbers of their matches)
bysort match1: gen weight=_N // count how many times each control observation is a match
by match1: keep if _n==1 // keep just one row per control observation
ren match1 ob //rename for merging purposes

merge 1:m ob using fulldata // merge back into the full data
replace weight=1 if gdum // set weight to 1 for treated observations

regress y treatment x x x xx x x x x z z z [fweight=weight], robust


so how must the middle part be modified so that the weights are adjusted to the nn(3 or 5) option?

Thank you very much!

Cheers,

Kurt

[Dano Hano]

Hello Kurt I have the same problem now I do not know how to adjust the weight in case of using more than one nearest neighbor, please could you help me if you have an idea ??thank you

[Wazzoi Daniels]

Hi Dano Hano Kurt Wichtelsacker I have run in to the same problem, did you manage to solve it?

[eric_a_booth]

clear
set obs 1000
 
g y = runiform()*10
g treatment = rbinomial(1, .66)
for A in num 1/5 : g xA = rbeta(1,100)

su *
regress y i.treatment x*  , robust
est sto before


**

local matches =3 //change this to number of nn 

g obs = _n
teffects psmatch (y) (treatment x*), atet gen(match) nn(`matches') 
teffects overlap

predict ps0 ps1, ps

predict y0 y1, po

predict te

sum te

sum te if treatment==1
sort obs
l y* treatment obs match* ps* te in 1/5, noobs clean
 
l y* treatment obs   match* ps* te  if inlist(24, obs, match1, match2, match3), noobs clean

sa fulldata,replace 


forval n = 1/`matches' {
keep if treatment  
keep match`n'  
bysort match`n': gen weight=_N  
by match`n': keep if _n==1 
rename match`n' obs
merge 1:m obs using fulldata.dta 
assert _m!=1
cap drop _m
replace weight=1 if treatment ==1   
}

tabstat weight ps* , by(treatment)


*take a look
tw   (scatter y ps1 if treatment ==0 )  (scatter y ps1 if treatment==1)  (lfit y ps1 if treatment ==0 , lwidth(thick) )  (lfit y ps1 if treatment==1, lwidth(thick))
*also  
cap which binscatter
if _rc   ssc install binscatter, replace //from SSC if not already installed
binscatter y ps1, by(treatment)  



qui regress y i.treatment x* [fweight=weight], robust
est sto after


est table before after, star

[Maria Ribeiro]

Could you please explain me what are the main goals of doing the following and how should I interpret them when comparing with teffects psmatch...?

 
 predict ps0 ps1, ps  predict y0 y1, po  predict te  sum te

And what about the code "teffects overlap"?

Thank you in advance.