根据首次发生冲击的时间，将staggered setting或general setting下的treatment group和control group进行重构，并生成_group以及_newid。
本命令旨在为手动panel-match提供思路
Content
1.面板数据的生成
首先，我写了一个函数，用于生成三种类型的面板数据：standard setting,staggered setting以及general setting。
这个函数依赖Hongyu Mou & Yiqing Xu老师的panelview包，需要先安装下:
net install grc1leg, from(http://www.stata.com/users/vwiggins) replace
net install gr0075, from(http://www.stata-journal.com/software/sj18-4) replace
ssc install labutil, replace
ssc install sencode, replace
net install panelview, all replace from("https://yiqingxu.org/packages/panelview_stata")
下面就是面板数据生成的函数，
可以根据需要选择standard setting,staggered setting或general setting进行生成
cap program drop panel_gen
program define panel_gen
syntax , type(string)
qui{
clear
set obs 200
egen id = repeat(),v(1/20)
sort id
egen t = repeat(),v(2005/2014)
if "`type'" == "standard"{
// standard-setting.
gen estimator = 1 if t >= 2011 & id <= 7
replace estimator = 0 if estimator == .
panelview estimator, i(id) t(t) type(treat)
}
if "`type'" == "staggered"{
// staggered-spec.
gen estimator = 1 if id <= 2 & t >= 2007
replace estimator = 1 if id == 3 & t >= 2008
replace estimator = 1 if id == 4 & t >= 2009
replace estimator = 1 if (id == 5|id == 6) & t >= 2010
replace estimator = 1 if (id == 7|id == 8) & t >= 2011
replace estimator = 1 if id == 9 & t >= 2012
replace estimator = 0 if estimator == .
panelview estimator, i(id) t(t) type(treat)
}
if "`type'" == "general"{
// general-setting.
gen estimator = 0
replace estimator = 1 if id == 1 & (t==2007|t==2008|t==2009|t==2014|t==2011)
replace estimator = 1 if id == 2 & (t==2008|t==2009|t==2011|t==2013|t==2014)
replace estimator = 1 if id == 3 & (t==2011|t==2009|t==2012)
replace estimator = 1 if id == 4 & (t==2008|t==2013|t==2014)
replace estimator = 1 if id == 5 & (t==2010|t==2011)
replace estimator = 1 if id == 6 & (t==2011|t==2012|t==2014)
replace estimator = 1 if id == 7 & (t==2010|t==2014)
replace estimator = 1 if id == 8 & (t==2012)
replace estimator = 1 if id == 9 & (t==2013|t==2014)
panelview estimator, i(id) t(t) type(treat)
}
}
end
在定义上述函数之后，即可通过以下命令生成相应面板
(1) standard setting
panel_gen, type(standard)
(2) staggered setting
panel_gen, type(staggered)
(3) general setting
panel_gen, type(general)
2.面板数据重构命令
根据panel_gen命令生成的面板数据，我们可以用panel_restruct命令进行重构。
下面是我写的panel_restruct函数用法，该命令能够针对staggered-setting和general-setting两种类型的数据一键进行重构。
panel_restruct, id(id) time(year) did(did_estimator) pre(2) post(2)
只需要定义四个参数
- id：panel id变量
- time：panel time变量
- did：did estimator variable
- pre：希望生成的事前期数
- post：希望生成的事后期数
3.使用示例
先调用panel_gen，生成特定spec的面板数据，
panel_gen, type(general)
然后调用panel_restruct，进行重构。
panel_restruct, id(id) time(t) did(estimator) pre(2) post(2)
至此，面板重构完成，输入数据为general setting，如左图所示
输出数据即重构后的数据如右图所示，已经将左图data打散，并将符合条件的treatment group 和 control group进行了组合
4.panel_restruct源码分享
cap program drop panel_restruct
program define panel_restruct
syntax , id(varlist) time(varlist) did(string) pre(string) post(string)
egen _id = group(`id')
xtset _id `time'
local offsets1 ""
forv i = 0/$pre {
local j_`i' = $pre - `i' + 1
local offsets1 "`offsets1' -`j_`i''"
}
local offsets11 ""
forv i = 1/$pre {
local j_`i' = $pre - `i' + 1
local offsets11 "`offsets11' -`j_`i''"
}
local offsets2 ""
forv i = 1/$post {
local j_`i' = $post - `i' + 1
local offsets2 "`j_`i'' `offsets2'"
}
global relat_time "`offsets11' 0 `offsets2'" //
global relat_time1 "`offsets1' 0 `offsets2'" // 往前多找一期
di "$relat_time1"
di "$relat_time"
su `time',d
local start_year = r(min) + $pre
local end_year = r(max) - $post
global gunit = `end_year' - `start_year' + 1
global length_windows = abs($pre + $post + 1)
forv tt = `start_year'/`end_year'{
preserve
forv i = -$pre/$post{
cap drop Dummyyear`i'
local kp = `tt' + `i'
di `kp'
gen Dummyyear`kp' = (`time' == `kp')
}
egen Dummyyear= rsum(Dummyyear*)
keep if Dummyyear == 1
drop Dummyyear*
// 确定treat==1的unit
cap drop _ident
cap drop _treat
gen _ident=.
su _id,d
forv i =`r(min)'/`r(max)'{
forv j = -$pre/-1{
local checkyear = `tt' + `j'
replace _ident = 1 if (_id == `i') & (`time'== `checkyear') & (`did' == 0)
}
forv j = 0/$post{
local checkyear = `tt' + `j'
replace _ident = 1 if (_id == `i') & (`time'== `checkyear') & (`did' > 0)
}
}
bys _id : egen _treat = count(_ident)
gen _treat_selected = (_treat == $length_windows)
// 确定control==1的unit
cap drop _ident
cap drop _treat
gen _ident=.
su _id,d
forv i =`r(min)'/`r(max)'{
forv j = -$pre/$post{
local checkyear = `tt' + `j'
replace _ident = 1 if (_id == `i') & (`time'== `checkyear') & (`did' == 0)
}
}
bys _id : egen _treat = count(_ident)
gen _control_selected = (_treat == $length_windows)
cap drop _type
cap drop _ident
cap drop _treat
gen _groupyear = `tt' if (_treat_selected==1| _control_selected==1)
gen _type = "treat" if _treat_selected == 1
replace _type = "control" if _control_selected == 1
drop if _type == ""
su _treat_selected,d
global conditionalways = r(mean)
if $conditionalways == 0 {
di "all treat, no panel-data combinated in group_year `tt'"
local dim = _N
drop in 1/`dim'
}
if $conditionalways == 1 {
di "all control, no panel-data combinated in group_year `tt'"
local dim = _N
drop in 1/`dim'
}
if $conditionalways < 1 & $conditionalways > 0 {
di "panel data generated in group_year `tt'"
}
cap drop _id
cap drop _treat_selected
cap drop _control_selected
save group_`tt'.dta,replace
restore
}
use group_`start_year'.dta,clear
local nstart_year = `start_year' + 1
forv i = `nstart_year'/`end_year'{
append using group_`i'.dta
}
forv i = `start_year'/`end_year'{
erase group_`i'.dta
}
cap drop _rt
gen _rt = `time' - _groupyear
egen _groupid = group(`id' _groupyear)
panelview `did', i(_groupid) t(_rt) type(treat)
end
