倾向得分匹配法方面的问题

（1）请问大家，我第一步使用logit计算完倾向得分后，我有五个协变量不显著，四个协变量显著。在下一步匹配时，我是用四个显著的协变量进行匹配还是9个协变量全部放进去进行匹配？我用的命令是psmatch2具体命令如下：
logit policy a b c d e f g h i 根据policy将数据分为对照组和实验组，其余a至i是用于匹配的协变量，stata结果显示a b c d显著，其余不显著。
下一步要开始进行核匹配了，问题就是我是将显著的放进去匹配还是全部放进去匹配？

（2）同志们，再追加一个问题啊！psmatch2命令如何 进行平衡性检验？我原来是使用pscore命令的，stata会自动进行平衡性检验，结果会说Balance is satisfied。但是现在我改用psmatch2的命令，然后不知道怎么检验平衡性？是输入命令然后stata自动显示
Balance is satisfied？还是输入什么命令，然后看某些指标然后自己判断？

* 倾向分匹配
* 假定y是因变量

* 个案的随机排序
gen tmp = runiform()
sort tmp

* 匹配方法：nearest neighbor matching within caliper
psmatch2 policy a b c d e f g h i , logit neighbor(1) common caliper(.05) ties
* 检验两者的平衡性
pstest a b c d e f g h i
* 剔除不显著的变量
psmatch2 policy a b c d , logit neighbor(1) common caliper(.05) ties
pstest a b c d
* 或 pscore命令来检验平衡性
pscore policy a b c d , pscore(newpscore) logit

* 估计policy对因变量y的效应
psmatch2 policy a b c d, outcome(y) logit neighbor(1) common caliper(.05)

* 平衡性的图示
psgraph

顶一个！

1) The logistic regression used for estimation of propensity score is "different" (has special characteristics) from ordinary logistic regression:
        a. Don’t care about the predictive ability of model
        b. Don’t  care about collinearity of covariates
        c. Theoretically, your should select all variables related to your treatment assignment (in your cases, policy) and your outcome
        c. Just care about whether it results in balanced matched samples

In your case, at this time, I would include all your 9 covariates (if you believe all of them are related to treatment assignment and especially your outcome) in the logistic model, then test the matched sample balance

2)By the way, if you use psmatch2, you don't have to use a separate logistic regression model to get the propensity score (although you can). You can integrate propensity score production and ATT estimation into psmatch2 like:

psmatch2 policy a b c d e f g h i, outcome(pollution) logit    //assume pollution is your outcome variable

3) Use  pstest to check your matched sample balance after pamatch2 like:

pstest, both

Good luck

luhmann 发表于 2013-10-30 12:48
* 倾向分匹配
* 假定y是因变量

谢谢您的回答！

hplcdadong 发表于 2013-11-2 04:07
1) The logistic regression used for estimation of propensity score is "different" (has special chara ...

谢谢您的回答！我本想将200个论坛币平均分给你和luhmann，但是论坛好像没有这个功能，只能选一个最佳答案。我权衡了一下，是luhmann先回答的，我就把他的选作最佳答案了，希望您能够谅解。

哥来看你了

hplcdadong 发表于 2013-11-2 04:07
1) The logistic regression used for estimation of propensity score is "different" (has special chara ...

请教一下，选择协变量时考虑同时与 treatment assignment and outcome相关吗？如果某个协变量只与outcome相关，而与 treatment assignment 不相关，那么是否包含这个变量进行logit回归？

kz 发表于 2014-6-12 17:11
请教一下，选择协变量时考虑同时与 treatment assignment and outcome相关吗？如果某个协变量只与outcome ...

Covariates associated with treatment assignment and outcome should be in the model. If you know that the covariate is related (only) to outcome, it should be in the model.