stata工具变量笔记代码

工具变量法操作

use mroz.dta,clear
describe
summarize 
## 清理数据
## drop nonworking women and summarize
drop if lfp==0
summarize wage educ exper
# create variables
gen lwage = ln(wage)
gen exper2 = exper^2
## 用普通回归作对比
#Least squares estimation
reg lwage educ exper exper2

#Example 1: (only 1 instrument) using only mothereduc as IV
## 手动方法
# first stage regression
reg educ exper exper2 mothereduc
##test relevance/weak instrument
test mothereduc
# obtain predicted values
predict educhat
# 2sls using 2-stages
reg lwage educhat exper exper2
-------------------------------------------------------------------
## 直接代码
# IV estimation using automatic command
ivregress 2sls lwage (educ=mothereduc) exper exper2
ivregress 2sls lwage (educ=mothereduc) exper exper2,small
ivregress 2sls lwage (educ=mothereduc) exper exper2, vce(robust) small

内生性检验

#test endogenous
qui ivregress 2sls lwage (educ=mothereduc) exper exper2
estimates store IV
qui reg lwage educ exper exper2
estimates store OLS
hausman IV OLS, constant sigmamore
### （sig怕统计量出现负值，最好加一下）
### hausman检验前面加的是在任何情况都能使用的模型，后面一个是在特定情况才成立
#the alternative way
qui reg educ exper exper2 mothereduc
predict res, residuals
reg lwage exper exper2 educ res
-------------------------------------------------------------------
# This test is equivalent to:
qui ivregress 2sls lwage exper exper2 (educ = mothereduc)
estat endogenous

结果的解读：看p值。此题情况p=0.08.即在10%显著性水平下通过，有理由拒绝存在内生性的假设，ols存在内生性问题，有必要使用2阶段最小二乘法估计。

弱工具变量检验

#Example 2: (more than 1 instrument)fathereduc, mothereduc
## 手动
##first stage
reg educ exper exper2 mothereduc fathereduc
##test relevance/weak instrument
test mothereduc fathereduc
ivregress 2sls lwage (educ=mothereduc fathereduc) exper exper2
##Testing for weak instruments using estat
estat firststage

qui ivregress 2sls lwage (educ=mothereduc fathereduc) exper exper2,vce(robust)
-------------------------------------------------------------------
## 直接代码
## Testing for weak instruments using estat
estat firststage

结果的解读：
当F值=36.14大于10时，可以通过检验，认为这不是一个弱工具变量。

过度识别检验

##overidentification test
qui ivregress 2sls lwage (educ=mothereduc fathereduc) exper exper2
estat overid

结果的解读：看p。但是要注意原假设！
这个情况是，p=0.5>0.1。没有通过过度识别检验，因此认为这个估计没有过度识别的倾向，使用工具变量的方法是恰当的。

两个变量的内生性检验

##test endogenous
qui ivregress 2sls lwage (educ=mothereduc fathereduc) exper exper2
estimates store IV
qui reg lwage educ exper exper2
estimates store OLS
hausman IV OLS,sigmamore