其他
PSM与DID的结合是一段“孽缘”
混合匹配
原文信息
石大千,丁海,卫平,刘建江.智慧城市建设能否降低环境污染[J].中国工业经济,2018(06):117-135.
20210415
。set seed 20210415
gen tmp=runiform()
sort tmp
psmatch2
命令(外部命令需要安装ssc install psmatch2, replace
)进行倾向得分匹配,我选择的匹配方法是一对一近邻匹配。其中,du
是处理组虚拟变量;$xlist
是协变量(控制变量);选择项out()
用来指定结果变量y
,这里填入DID模型的被解释变量即可;选择项logit
表示使用logit模型来估计倾向得分,默认方法是probit;选择项common
表示仅对共同取值范围内个体进行匹配;选择项ate
表示同时汇报ATE、ATU和ATT。. psmatch2 du $xlist , out(lnrso) logit neighbor(1) common ate //近邻匹配
_weight
为空值的样本),然后再用DID方法去进行估计就可以了。PSM-DID估计结果显示,交互项dudt
的系数为-0.1781016,表明智慧城市建设显著降低了约17.81%的人均废气排放量。. drop if _weight==.
(1,467 observations deleted)
. reghdfe lnrso dudt $xlist ,absorb(c year) vce(cluster c) //DID估计
(dropped 21 singleton observations)
(MWFE estimator converged in 7 iterations)
HDFE Linear regression Number of obs = 1,207
Absorbing 2 HDFE groups F( 7, 218) = 2.12
Statistics robust to heteroskedasticity Prob > F = 0.0423
R-squared = 0.8880
Adj R-squared = 0.8608
Within R-sq. = 0.0350
Number of clusters (c) = 219 Root MSE = 0.4147
(Std. Err. adjusted for 219 clusters in c)
------------------------------------------------------------------------------
| Robust
lnrso | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
dudt | -.1781016 .0834432 -2.13 0.034 -.3425604 -.0136429
lnrgdp | 2.101315 1.207922 1.74 0.083 -.2793847 4.482014
lntgdp | -.0910883 .0572348 -1.59 0.113 -.2038928 .0217162
lninno | .1235118 .0676161 1.83 0.069 -.009753 .2567767
lnurb | .1283764 .064476 1.99 0.048 .0013002 .2554525
lnopen | .0317352 .0459506 0.69 0.491 -.0588291 .1222995
lnss | -.1781226 .3485232 -0.51 0.610 -.865029 .5087837
_cons | -13.93434 6.665826 -2.09 0.038 -27.07205 -.7966249
------------------------------------------------------------------------------
Absorbed degrees of freedom:
-----------------------------------------------------+
Absorbed FE | Categories - Redundant = Num. Coefs |
-------------+---------------------------------------|
c | 219 219 0 *|
year | 11 0 11 |
-----------------------------------------------------+
* = FE nested within cluster; treated as redundant for DoF computation
. list c year city prov _treated _id _n1 if _id==379|_id==1915
+------------------------------------------------------------+
| c year city prov _treated _id _n1 |
|------------------------------------------------------------|
1. | 61 2014 佳木斯市 黑龙江省 Untreated 379 1915 |
1220. | 242 2006 遵义市 贵州省 Treated 1915 379 |
+------------------------------------------------------------+
逐期匹配
i
在每期的匹配对象可能均不相同,混合匹配当然也存在这一问题)。对于同一个处理组个体,其对照对象如果在政策时点前后发生较大改变,将会导致个体固定效应的估计出现偏差,进而影响到DID模型的稳定性。除此之外,逐期匹配存在着其他问题,大家有兴趣可以去阅读《传统PSM-DID模型的改进与应用》这篇论文。参考资料(推荐阅读)
谢申祥,范鹏飞,宛圆渊.传统PSM-DID模型的改进与应用[J].统计研究,2021,38(02):146-160.
forvalues i = 2005/2015{
use smart_city2018.dta,clear
keep if year==`i'
set seed 20210415
gen tmp=runiform()
sort tmp
psmatch2 du $xlist , out(lnrso) logit neighbor(1) common ate
drop if _weight==.
save psmdid_`i',replace
}
use psmdid_2005,clear
forvalues i = 2006/2015{
append using psmdid_`i'
}
PSM-DID
”!