稳健性检验是什么？
稳健性检验回答的是主结论在合理替代口径下是否仍然成立。它不是重新找一个更显著的模型，而是围绕同一个研究问题，替换变量定义、样本范围、标准误或估计方法来检查结论是否脆弱。
先用一句话理解
把稳健性检验想成给主规格做压力测试。主规格是一把尺子，稳健性不是换题目，而是换几把合理的尺子看看同一个关系会不会马上消失。
什么时候用
| 判断项 | 教科书式判断 |
|---|---|
| 适合用在 | 主规格已经清楚，需要回应变量定义、样本选择、极端值、标准误或模型设定风险。 |
| 先不要用在 | 主规格本身还没固定，或者只是为了找到显著结果不断试口径。 |
| 最关键的前提 | 每个稳健性口径都要对应一个具体风险，不能只说“换了很多方法”。 |
| 读表顺序 | 先看替代了什么，再看方向、量级和推断是否与主规格一致。 |
| 论文表述边界 | 稳健性支持的是主结论稳定，不应伪装成新的主规格显著性。 |
把它和相近方法分清楚
| 容易混淆的对象 | 怎么区分 | 初学者记法 |
|---|---|---|
| 主规格 | 预先说明的核心模型。 | 全文参照点。 |
| 稳健性检验 | 围绕主规格改变合理口径。 | 检验结论是否脆弱。 |
| 显著性优化 | 搜索能让结果更显著的规格。 | 不应包装成稳健性。 |
核心直觉和模型公式
同一结论的替代口径
稳健性检验可以理解为同一个核心关系在不同合理设定下的重复估计：
\[Y^{(m)}_{it}=\beta^{(m)}X^{(m)}_{it}+\boldsymbol{\gamma}^{(m)'}\mathbf{Z}^{(m)}_{it}+FE^{(m)}+\varepsilon^{(m)}_{it}\]
上标 m 表示第 m 个合理替代口径；重点是 β 的方向、量级和推断是否稳定。
公式里的符号怎么读
| 符号 | 含义 | 在本页怎么理解 |
|---|---|---|
| m | 替代口径 | 例如换变量定义、样本、标准误或估计方法。 |
| β^(m) | 稳健性系数 | 第 m 个口径下的核心估计量。 |
| X^(m) | 替代核心变量 | 同一理论概念的另一种测量方式。 |
| Z^(m) | 替代控制集合 | 合理增减或重新定义的控制变量。 |
| FE^(m) | 固定效应口径 | 不同维度或更细固定效应。 |
专业使用口径
专业稳健性表要让读者知道每一列在回应哪一种质疑。
| 口径 | 专业写法 |
|---|---|
| 变量替代 | 说明新变量和主变量测量的是同一理论概念。 |
| 样本替代 | 说明删除或限制样本的理由，不是事后挑样本。 |
| 推断替代 | 标准误、聚类或 bootstrap 变化应对应误差结构风险。 |
| 结论处理 | 不显著或方向变化的稳健性不能静默删除。 |
一步一步做：从问题到结果的五步
| 步骤 | 这一页具体做什么 | 做到什么程度才算读懂 |
|---|---|---|
| 1. 固定主规格 | 先有一条清楚的主模型。 | 主规格不固定，稳健性就没有参照点。 |
| 2. 写出风险清单 | 变量、样本、极端值、标准误、模型形式分别有什么风险。 | 每个检验要对应一个风险。 |
| 3. 设计替代口径 | 只改变少数合理设定。 | 一次改太多会无法解释。 |
| 4. 对照主系数 | 看方向、量级和 p 值是否大幅变化。 | 不要只看是否仍有星号。 |
| 5. 诚实写边界 | 冲突结果要解释，而不是隐藏。 | 稳健性也可以削弱主结论。 |
结果怎么读：先读口径，再读数字
本页案例读表时，先记住这个口径：dfi_index = 0.5660***；0.5530***。
| 读表顺序 | 本页怎么读 | 不要这样读 |
|---|---|---|
| 先看列名 | 每一列到底替换了什么。 | 不要直接跳到星号。 |
| 再看方向和量级 | 是否和主规格相近。 | 只保留显著性不够。 |
| 然后看推断口径 | 标准误或聚类是否变化。 | 不同推断口径不能混写。 |
| 最后看冲突项 | 哪些检验不支持主结论。 | 不要把冲突结果说成全部稳健。 |
初学者最容易错在哪里
| 误读 | 为什么错 |
|---|---|
| 把稳健性当显著性搜索 | 错。稳健性先来自风险清单，不来自结果好坏。 |
| 一次改很多东西 | 错。读者不知道结论变化来自哪里。 |
| 只展示通过的检验 | 错。选择性展示会削弱可信度。 |
| 把专用稳健性当主规格 patch | 错。PSM、CSDID、Permutation 等应作为专门模块披露。 |
先看这个案例的结论
- dfi_index = 0.5660***；0.5530***。
- roa = 0.3040***；0.2924***。
- lev = -0.0548；-0.0545。
- 这些数字来自页面里的结果表；写论文时先解释数值含义，再讨论理论含义。
变量要求和案例口径
| 字段 | 口径 |
|---|---|
| 数据 | CSMAR 风格 A 股企业创新面板 |
| 原始样本 | 196 家上市公司，2015-2020 年，约 1200 个公司-年观测；各方法有效样本以本页输出表 N 为准 |
| 因变量 | patent_count；回归页通常使用 ln(1 + patent_count) |
| 核心解释变量 | dfi_index，数字普惠金融指数；部分案例输出展示的是标准化后的 dfi_index |
| 控制变量 | roa、lev、size、growth、cashflow、tobinq、top1、dual、board、indep、soe、age |
| 输出文件 | regression_table_稳健性检验.csv |
| 角色要求 | 因变量、核心解释变量 |
| 依赖包 | 无额外 Stata 社区包要求 |
实际代码
下面是本页案例里的最小 Stata 代码。它的作用是帮助你看清变量怎么进入模型、结果文件怎么生成，而不是要求你在网页里手动运行。
log using "/root/workspace/empirical-wizard/workspace/2a35153b/analysis.log", replace text
global JOB_DIR "/root/workspace/empirical-wizard/workspace/2a35153b"
set more off
adopath + "/root/ado/plus"
global DATA_PATH "/root/workspace/empirical-wizard/workspace/test_e2e/csmar_innovation.csv"
import delimited "/root/workspace/empirical-wizard/workspace/test_e2e/csmar_innovation.csv", clear case(preserve)
capture confirm global JOB_DIR
if _rc global JOB_DIR "."
quietly duplicates drop
local dvvar "patent_count"
local ivvar "dfi_index"
local controls "roa lev size growth cashflow tobinq top1 dual board indep soe age"
local idvar "stkcd"
local timevar "year"
local industryvar "ind"
local geovar ""
local allvars "dfi_index roa lev size growth cashflow tobinq top1 dual board indep soe age"
capture confirm variable `dvvar'
if _rc {
di as error "Dependent variable not found: `dvvar'"
exit 111
}
capture confirm variable `ivvar'
if _rc {
di as error "Core independent variable not found: `ivvar'"
exit 111
}
local absorb_id "`idvar'"
if "`idvar'" != "" {
capture confirm numeric variable `idvar'
if _rc {
tempvar __ewiz_id
encode `idvar', gen(`__ewiz_id')
local absorb_id "`__ewiz_id'"
}
}
local cluster_id "`absorb_id'"
local fe_time "`timevar'"
if "`timevar'" != "" {
capture confirm numeric variable `timevar'
if _rc {
tempvar __ewiz_time
encode `timevar', gen(`__ewiz_time')
local fe_time "`__ewiz_time'"
}
}
local industry_fe "`industryvar'"
if "`industryvar'" != "" {
capture confirm numeric variable `industryvar'
if _rc {
tempvar __ewiz_industry
encode `industryvar', gen(`__ewiz_industry')
local industry_fe "`__ewiz_industry'"
}
}
local cluster_industry "`industry_fe'"
local geo_fe "`geovar'"
if "`geovar'" != "" {
capture confirm numeric variable `geovar'
if _rc {
tempvar __ewiz_geo
encode `geovar', gen(`__ewiz_geo')
local geo_fe "`__ewiz_geo'"
}
}
local cluster_geo "`geo_fe'"
tempfile ewiz_summary
tempname ewiz_sum
postfile `ewiz_sum' str32 spec double coef se pvalue str4 stars double N R2 using `ewiz_summary', replace
foreach v of local allvars {
local cell_r1_`v' ""
local secell_r1_`v' ""
local cell_r2_`v' ""
local secell_r2_`v' ""
local cell_r3_`v' ""
local secell_r3_`v' ""
local cell_r4_`v' ""
local secell_r4_`v' ""
local cell_r7_`v' ""
local secell_r7_`v' ""
}
preserve
quietly areg `dvvar' `ivvar' roa lev size growth cashflow tobinq top1 dual board indep soe age i.`fe_time', absorb(`absorb_id') vce(cluster `cluster_id' `fe_time')
restore
capture local N_r1 : display %12.0f e(N)
if "`N_r1'"=="" local N_r1 = .
capture local R2_r1 : display %9.4f e(r2)
if "`R2_r1'"=="" local R2_r1 = .
capture scalar __ewiz_main_b_r1 = _b[`ivvar']
capture scalar __ewiz_main_se_r1 = _se[`ivvar']
capture local coefmain_r1 : display %9.4f __ewiz_main_b_r1
capture if abs(__ewiz_main_b_r1) < 0.00005 & __ewiz_main_b_r1 != 0 local coefmain_r1 : display %9.2e __ewiz_main_b_r1
capture local semain_r1 : display %9.4f __ewiz_main_se_r1
capture if abs(__ewiz_main_se_r1) < 0.00005 & __ewiz_main_se_r1 != 0 local semain_r1 : display %9.2e __ewiz_main_se_r1
local pmain_r1 = .
local starsmain_r1 ""
local __ewiz_model_valid_r1 = 1
capture scalar __ewiz_coef_abs_r1 = abs(_b[`ivvar'])
if _rc local __ewiz_model_valid_r1 = 0
capture scalar __ewiz_se_abs_r1 = abs(_se[`ivvar'])
if _rc local __ewiz_model_valid_r1 = 0
capture if scalar(__ewiz_se_abs_r1) <= 1e-12 local __ewiz_model_valid_r1 = 0
capture if scalar(__ewiz_coef_abs_r1) <= 1e-12 & scalar(__ewiz_se_abs_r1) <= 1e-12 local __ewiz_model_valid_r1 = 0
capture local __df_r1 = e(df_r)
capture local __z_r1 = abs(_b[`ivvar'] / _se[`ivvar'])
if !_rc & `__ewiz_model_valid_r1' == 1 {
if "`__df_r1'"=="" | "`__df_r1'"=="." local pmain_r1 = 2*(1-normal(`__z_r1'))
else local pmain_r1 = 2*ttail(`__df_r1', `__z_r1')
}
if `__ewiz_model_valid_r1' == 1 local starsmain_r1 = cond(`pmain_r1'<0.01,"***",cond(`pmain_r1'<0.05,"**",cond(`pmain_r1'<0.1,"*","")))
if `__ewiz_model_valid_r1' == 1 {
capture post `ewiz_sum' ("(R1)") (_b[`ivvar']) (_se[`ivvar']) (`pmain_r1') ("`starsmain_r1'") (`N_r1') (`R2_r1')
if _rc post `ewiz_sum' ("(R1)") (.) (.) (.) ("") (`N_r1') (`R2_r1')
}
else post `ewiz_sum' ("(R1)") (.) (.) (.) ("") (`N_r1') (`R2_r1')
if `__ewiz_model_valid_r1' == 1 {
local __included_vars "`ivvar' dfi_index roa lev size growth cashflow tobinq top1 dual board indep soe age"
foreach v of local allvars {
local __is_included = 0
foreach __inc of local __included_vars {
if "`v'" == "`__inc'" local __is_included = 1
}
if `__is_included' == 0 continue
local model_term "`v'"
if "`v'" == "`ivvar'" local model_term "`ivvar'"
capture scalar __ewiz_term_b = _b[`model_term']
capture scalar __ewiz_term_se = _se[`model_term']
if _rc | missing(__ewiz_term_se) | __ewiz_term_se <= 1e-12 {
local cell_r1_`v' "omitted"
local secell_r1_`v' "(absorbed)"
continue
}
capture local __coef : display %9.4f __ewiz_term_b
capture if abs(__ewiz_term_b) < 0.00005 & __ewiz_term_b != 0 local __coef : display %9.2e __ewiz_term_b
if !_rc {
capture local __se : display %9.4f __ewiz_term_se
capture if abs(__ewiz_term_se) < 0.00005 & __ewiz_term_se != 0 local __se : display %9.2e __ewiz_term_se
local __p = .
local __stars ""
capture local __df = e(df_r)
capture local __z = abs(_b[`model_term'] / _se[`model_term'])
if !_rc {
if "`__df'"=="" | "`__df'"=="." local __p = 2*(1-normal(`__z'))
else local __p = 2*ttail(`__df', `__z')
}
if !_rc local __stars = cond(`__p'<0.01,"***",cond(`__p'<0.05,"**",cond(`__p'<0.1,"*","")))
local cell_r1_`v' "`__coef'`__stars'"
local secell_r1_`v' "(`__se')"
}
}
}
di "(R1): `ivvar'=`coefmain_r1'`starsmain_r1' (se=`semain_r1', p=`pmain_r1'), N=`N_r1', R2=`R2_r1'"
tempvar dv_w iv_w
gen double `dv_w' = `dvvar'
gen double `iv_w' = `ivvar'
quietly summarize `dvvar', detail
local dv_p1 = r(p1)
local dv_p99 = r(p99)
replace `dv_w' = `dv_p1' if !missing(`dv_w') & `dv_w' < `dv_p1'
replace `dv_w' = `dv_p99' if !missing(`dv_w') & `dv_w' > `dv_p99'
quietly summarize `ivvar', detail
local iv_p1 = r(p1)
local iv_p99 = r(p99)
replace `iv_w' = `iv_p1' if !missing(`iv_w') & `iv_w' < `iv_p1'
replace `iv_w' = `iv_p99' if !missing(`iv_w') & `iv_w' > `iv_p99'
quietly areg `dv_w' `iv_w' roa lev size growth cashflow tobinq top1 dual board indep soe age i.`fe_time', absorb(`absorb_id') vce(robust)
capture local N_r2 : display %12.0f e(N)
if "`N_r2'"=="" local N_r2 = .
capture local R2_r2 : display %9.4f e(r2)
if "`R2_r2'"=="" local R2_r2 = .
capture scalar __ewiz_main_b_r2 = _b[`iv_w']
capture scalar __ewiz_main_se_r2 = _se[`iv_w']
capture local coefmain_r2 : display %9.4f __ewiz_main_b_r2
capture if abs(__ewiz_main_b_r2) < 0.00005 & __ewiz_main_b_r2 != 0 local coefmain_r2 : display %9.2e __ewiz_main_b_r2
capture local semain_r2 : display %9.4f __ewiz_main_se_r2
capture if abs(__ewiz_main_se_r2) < 0.00005 & __ewiz_main_se_r2 != 0 local semain_r2 : display %9.2e __ewiz_main_se_r2
local pmain_r2 = .
local starsmain_r2 ""
local __ewiz_model_valid_r2 = 1
capture scalar __ewiz_coef_abs_r2 = abs(_b[`iv_w'])
if _rc local __ewiz_model_valid_r2 = 0
capture scalar __ewiz_se_abs_r2 = abs(_se[`iv_w'])
if _rc local __ewiz_model_valid_r2 = 0
capture if scalar(__ewiz_se_abs_r2) <= 1e-12 local __ewiz_model_valid_r2 = 0
capture if scalar(__ewiz_coef_abs_r2) <= 1e-12 & scalar(__ewiz_se_abs_r2) <= 1e-12 local __ewiz_model_valid_r2 = 0
capture local __df_r2 = e(df_r)
capture local __z_r2 = abs(_b[`iv_w'] / _se[`iv_w'])
if !_rc & `__ewiz_model_valid_r2' == 1 {
if "`__df_r2'"=="" | "`__df_r2'"=="." local pmain_r2 = 2*(1-normal(`__z_r2'))
else local pmain_r2 = 2*ttail(`__df_r2', `__z_r2')
}
if `__ewiz_model_valid_r2' == 1 local starsmain_r2 = cond(`pmain_r2'<0.01,"***",cond(`pmain_r2'<0.05,"**",cond(`pmain_r2'<0.1,"*","")))
if `__ewiz_model_valid_r2' == 1 {
capture post `ewiz_sum' ("(R2)") (_b[`iv_w']) (_se[`iv_w']) (`pmain_r2') ("`starsmain_r2'") (`N_r2') (`R2_r2')
if _rc post `ewiz_sum' ("(R2)") (.) (.) (.) ("") (`N_r2') (`R2_r2')
}
else post `ewiz_sum' ("(R2)") (.) (.) (.) ("") (`N_r2') (`R2_r2')
if `__ewiz_model_valid_r2' == 1 {
local __included_vars "`iv_w' dfi_index roa lev size growth cashflow tobinq top1 dual board indep soe age"
foreach v of local allvars {
local __is_included = 0
foreach __inc of local __included_vars {
if "`v'" == "`__inc'" local __is_included = 1
}
if `__is_included' == 0 continue
local model_term "`v'"
if "`v'" == "`ivvar'" local model_term "`iv_w'"
capture scalar __ewiz_term_b = _b[`model_term']
capture scalar __ewiz_term_se = _se[`model_term']
if _rc | missing(__ewiz_term_se) | __ewiz_term_se <= 1e-12 {
local cell_r2_`v' "omitted"
local secell_r2_`v' "(absorbed)"
continue
}
capture local __coef : display %9.4f __ewiz_term_b
capture if abs(__ewiz_term_b) < 0.00005 & __ewiz_term_b != 0 local __coef : display %9.2e __ewiz_term_b
if !_rc {
capture local __se : display %9.4f __ewiz_term_se
capture if abs(__ewiz_term_se) < 0.00005 & __ewiz_term_se != 0 local __se : display %9.2e __ewiz_term_se
local __p = .
local __stars ""
capture local __df = e(df_r)
capture local __z = abs(_b[`model_term'] / _se[`model_term'])
if !_rc {
if "`__df'"=="" | "`__df'"=="." local __p = 2*(1-normal(`__z'))
else local __p = 2*ttail(`__df', `__z')
}
if !_rc local __stars = cond(`__p'<0.01,"***",cond(`__p'<0.05,"**",cond(`__p'<0.1,"*","")))
local cell_r2_`v' "`__coef'`__stars'"
local secell_r2_`v' "(`__se')"
}
}
}
di "(R2): `iv_w'=`coefmain_r2'`starsmain_r2' (se=`semain_r2', p=`pmain_r2'), N=`N_r2', R2=`R2_r2'"
tempvar dv_w5 iv_w5
gen double `dv_w5' = `dvvar'
gen double `iv_w5' = `ivvar'
quietly summarize `dvvar', detail
local dv_p5 = r(p5)
local dv_p95 = r(p95)
replace `dv_w5' = `dv_p5' if !missing(`dv_w5') & `dv_w5' < `dv_p5'
replace `dv_w5' = `dv_p95' if !missing(`dv_w5') & `dv_w5' > `dv_p95'
quietly summarize `ivvar', detail
local iv_p5 = r(p5)
local iv_p95 = r(p95)
replace `iv_w5' = `iv_p5' if !missing(`iv_w5') & `iv_w5' < `iv_p5'
replace `iv_w5' = `iv_p95' if !missing(`iv_w5') & `iv_w5' > `iv_p95'
quietly areg `dv_w5' `iv_w5' roa lev size growth cashflow tobinq top1 dual board indep soe age i.`fe_time', absorb(`absorb_id') vce(robust)
capture local N_r3 : display %12.0f e(N)
if "`N_r3'"=="" local N_r3 = .
capture local R2_r3 : display %9.4f e(r2)
if "`R2_r3'"=="" local R2_r3 = .
capture scalar __ewiz_main_b_r3 = _b[`iv_w5']
capture scalar __ewiz_main_se_r3 = _se[`iv_w5']
capture local coefmain_r3 : display %9.4f __ewiz_main_b_r3
capture if abs(__ewiz_main_b_r3) < 0.00005 & __ewiz_main_b_r3 != 0 local coefmain_r3 : display %9.2e __ewiz_main_b_r3
capture local semain_r3 : display %9.4f __ewiz_main_se_r3
capture if abs(__ewiz_main_se_r3) < 0.00005 & __ewiz_main_se_r3 != 0 local semain_r3 : display %9.2e __ewiz_main_se_r3
local pmain_r3 = .
local starsmain_r3 ""
local __ewiz_model_valid_r3 = 1
capture scalar __ewiz_coef_abs_r3 = abs(_b[`iv_w5'])
if _rc local __ewiz_model_valid_r3 = 0
capture scalar __ewiz_se_abs_r3 = abs(_se[`iv_w5'])
if _rc local __ewiz_model_valid_r3 = 0
capture if scalar(__ewiz_se_abs_r3) <= 1e-12 local __ewiz_model_valid_r3 = 0
capture if scalar(__ewiz_coef_abs_r3) <= 1e-12 & scalar(__ewiz_se_abs_r3) <= 1e-12 local __ewiz_model_valid_r3 = 0
capture local __df_r3 = e(df_r)
capture local __z_r3 = abs(_b[`iv_w5'] / _se[`iv_w5'])
if !_rc & `__ewiz_model_valid_r3' == 1 {
if "`__df_r3'"=="" | "`__df_r3'"=="." local pmain_r3 = 2*(1-normal(`__z_r3'))
else local pmain_r3 = 2*ttail(`__df_r3', `__z_r3')
}
if `__ewiz_model_valid_r3' == 1 local starsmain_r3 = cond(`pmain_r3'<0.01,"***",cond(`pmain_r3'<0.05,"**",cond(`pmain_r3'<0.1,"*","")))
if `__ewiz_model_valid_r3' == 1 {
capture post `ewiz_sum' ("(R3)") (_b[`iv_w5']) (_se[`iv_w5']) (`pmain_r3') ("`starsmain_r3'") (`N_r3') (`R2_r3')
if _rc post `ewiz_sum' ("(R3)") (.) (.) (.) ("") (`N_r3') (`R2_r3')
}
else post `ewiz_sum' ("(R3)") (.) (.) (.) ("") (`N_r3') (`R2_r3')
if `__ewiz_model_valid_r3' == 1 {
local __included_vars "`iv_w5' dfi_index roa lev size growth cashflow tobinq top1 dual board indep soe age"
foreach v of local allvars {
local __is_included = 0
foreach __inc of local __included_vars {
if "`v'" == "`__inc'" local __is_included = 1
}
if `__is_included' == 0 continue
local model_term "`v'"
if "`v'" == "`ivvar'" local model_term "`iv_w5'"
capture scalar __ewiz_term_b = _b[`model_term']
capture scalar __ewiz_term_se = _se[`model_term']
if _rc | missing(__ewiz_term_se) | __ewiz_term_se <= 1e-12 {
local cell_r3_`v' "omitted"
local secell_r3_`v' "(absorbed)"
continue
}
capture local __coef : display %9.4f __ewiz_term_b
capture if abs(__ewiz_term_b) < 0.00005 & __ewiz_term_b != 0 local __coef : display %9.2e __ewiz_term_b
if !_rc {
capture local __se : display %9.4f __ewiz_term_se
capture if abs(__ewiz_term_se) < 0.00005 & __ewiz_term_se != 0 local __se : display %9.2e __ewiz_term_se
local __p = .
local __stars ""
capture local __df = e(df_r)
capture local __z = abs(_b[`model_term'] / _se[`model_term'])
if !_rc {
if "`__df'"=="" | "`__df'"=="." local __p = 2*(1-normal(`__z'))
else local __p = 2*ttail(`__df', `__z')
}
if !_rc local __stars = cond(`__p'<0.01,"***",cond(`__p'<0.05,"**",cond(`__p'<0.1,"*","")))
local cell_r3_`v' "`__coef'`__stars'"
local secell_r3_`v' "(`__se')"
}
}
}
di "(R3): `iv_w5'=`coefmain_r3'`starsmain_r3' (se=`semain_r3', p=`pmain_r3'), N=`N_r3', R2=`R2_r3'"
tempvar _resid_keep
quietly areg `dvvar' `ivvar' roa lev size growth cashflow tobinq top1 dual board indep soe age i.`fe_time', absorb(`absorb_id') vce(robust)
capture predict double `_resid_keep', rstudent
if _rc == 0 {
replace `_resid_keep' = . if abs(`_resid_keep') > 3
}
preserve
capture drop if missing(`_resid_keep')
quietly areg `dvvar' `ivvar' roa lev size growth cashflow tobinq top1 dual board indep soe age i.`fe_time', absorb(`absorb_id') vce(robust)
restore
capture local N_r4 : display %12.0f e(N)
if "`N_r4'"=="" local N_r4 = .
capture local R2_r4 : display %9.4f e(r2)
if "`R2_r4'"=="" local R2_r4 = .
capture scalar __ewiz_main_b_r4 = _b[`ivvar']
capture scalar __ewiz_main_se_r4 = _se[`ivvar']
capture local coefmain_r4 : display %9.4f __ewiz_main_b_r4
capture if abs(__ewiz_main_b_r4) < 0.00005 & __ewiz_main_b_r4 != 0 local coefmain_r4 : display %9.2e __ewiz_main_b_r4
capture local semain_r4 : display %9.4f __ewiz_main_se_r4
capture if abs(__ewiz_main_se_r4) < 0.00005 & __ewiz_main_se_r4 != 0 local semain_r4 : display %9.2e __ewiz_main_se_r4
local pmain_r4 = .
local starsmain_r4 ""
local __ewiz_model_valid_r4 = 1
capture scalar __ewiz_coef_abs_r4 = abs(_b[`ivvar'])
if _rc local __ewiz_model_valid_r4 = 0
capture scalar __ewiz_se_abs_r4 = abs(_se[`ivvar'])
if _rc local __ewiz_model_valid_r4 = 0
capture if scalar(__ewiz_se_abs_r4) <= 1e-12 local __ewiz_model_valid_r4 = 0
capture if scalar(__ewiz_coef_abs_r4) <= 1e-12 & scalar(__ewiz_se_abs_r4) <= 1e-12 local __ewiz_model_valid_r4 = 0
capture local __df_r4 = e(df_r)
capture local __z_r4 = abs(_b[`ivvar'] / _se[`ivvar'])
if !_rc & `__ewiz_model_valid_r4' == 1 {
if "`__df_r4'"=="" | "`__df_r4'"=="." local pmain_r4 = 2*(1-normal(`__z_r4'))
else local pmain_r4 = 2*ttail(`__df_r4', `__z_r4')
}
if `__ewiz_model_valid_r4' == 1 local starsmain_r4 = cond(`pmain_r4'<0.01,"***",cond(`pmain_r4'<0.05,"**",cond(`pmain_r4'<0.1,"*","")))
if `__ewiz_model_valid_r4' == 1 {
capture post `ewiz_sum' ("(R4)") (_b[`ivvar']) (_se[`ivvar']) (`pmain_r4') ("`starsmain_r4'") (`N_r4') (`R2_r4')
if _rc post `ewiz_sum' ("(R4)") (.) (.) (.) ("") (`N_r4') (`R2_r4')
}
else post `ewiz_sum' ("(R4)") (.) (.) (.) ("") (`N_r4') (`R2_r4')
if `__ewiz_model_valid_r4' == 1 {
local __included_vars "`ivvar' dfi_index roa lev size growth cashflow tobinq top1 dual board indep soe age"
foreach v of local allvars {
local __is_included = 0
foreach __inc of local __included_vars {
if "`v'" == "`__inc'" local __is_included = 1
}
if `__is_included' == 0 continue
local model_term "`v'"
if "`v'" == "`ivvar'" local model_term "`ivvar'"
capture scalar __ewiz_term_b = _b[`model_term']
capture scalar __ewiz_term_se = _se[`model_term']
if _rc | missing(__ewiz_term_se) | __ewiz_term_se <= 1e-12 {
local cell_r4_`v' "omitted"
local secell_r4_`v' "(absorbed)"
continue
}
capture local __coef : display %9.4f __ewiz_term_b
capture if abs(__ewiz_term_b) < 0.00005 & __ewiz_term_b != 0 local __coef : display %9.2e __ewiz_term_b
if !_rc {
capture local __se : display %9.4f __ewiz_term_se
capture if abs(__ewiz_term_se) < 0.00005 & __ewiz_term_se != 0 local __se : display %9.2e __ewiz_term_se
local __p = .
local __stars ""
capture local __df = e(df_r)
capture local __z = abs(_b[`model_term'] / _se[`model_term'])
if !_rc {
if "`__df'"=="" | "`__df'"=="." local __p = 2*(1-normal(`__z'))
else local __p = 2*ttail(`__df', `__z')
}
if !_rc local __stars = cond(`__p'<0.01,"***",cond(`__p'<0.05,"**",cond(`__p'<0.1,"*","")))
local cell_r4_`v' "`__coef'`__stars'"
local secell_r4_`v' "(`__se')"
}
}
}
di "(R4): `ivvar'=`coefmain_r4'`starsmain_r4' (se=`semain_r4', p=`pmain_r4'), N=`N_r4', R2=`R2_r4'"
capture noisily poisson `dvvar' `ivvar' roa lev size growth cashflow tobinq top1 dual board indep soe age i.`timevar' i.`industry_fe', vce(cluster `cluster_id')
if _rc != 0 {
capture noisily nbreg `dvvar' `ivvar' roa lev size growth cashflow tobinq top1 dual board indep soe age i.`timevar' i.`industry_fe', vce(cluster `cluster_id')
}
capture local N_r7 : display %12.0f e(N)
if "`N_r7'"=="" local N_r7 = .
capture local R2_r7 : display %9.4f e(r2)
if "`R2_r7'"=="" local R2_r7 = .
capture scalar __ewiz_main_b_r7 = _b[`ivvar']
capture scalar __ewiz_main_se_r7 = _se[`ivvar']
capture local coefmain_r7 : display %9.4f __ewiz_main_b_r7
capture if abs(__ewiz_main_b_r7) < 0.00005 & __ewiz_main_b_r7 != 0 local coefmain_r7 : display %9.2e __ewiz_main_b_r7
capture local semain_r7 : display %9.4f __ewiz_main_se_r7
capture if abs(__ewiz_main_se_r7) < 0.00005 & __ewiz_main_se_r7 != 0 local semain_r7 : display %9.2e __ewiz_main_se_r7
local pmain_r7 = .
local starsmain_r7 ""
local __ewiz_model_valid_r7 = 1
capture scalar __ewiz_coef_abs_r7 = abs(_b[`ivvar'])
if _rc local __ewiz_model_valid_r7 = 0
capture scalar __ewiz_se_abs_r7 = abs(_se[`ivvar'])
if _rc local __ewiz_model_valid_r7 = 0
capture if scalar(__ewiz_se_abs_r7) <= 1e-12 local __ewiz_model_valid_r7 = 0
capture if scalar(__ewiz_coef_abs_r7) <= 1e-12 & scalar(__ewiz_se_abs_r7) <= 1e-12 local __ewiz_model_valid_r7 = 0
capture local __df_r7 = e(df_r)
capture local __z_r7 = abs(_b[`ivvar'] / _se[`ivvar'])
if !_rc & `__ewiz_model_valid_r7' == 1 {
if "`__df_r7'"=="" | "`__df_r7'"=="." local pmain_r7 = 2*(1-normal(`__z_r7'))
else local pmain_r7 = 2*ttail(`__df_r7', `__z_r7')
}
if `__ewiz_model_valid_r7' == 1 local starsmain_r7 = cond(`pmain_r7'<0.01,"***",cond(`pmain_r7'<0.05,"**",cond(`pmain_r7'<0.1,"*","")))
if `__ewiz_model_valid_r7' == 1 {
capture post `ewiz_sum' ("(R7)") (_b[`ivvar']) (_se[`ivvar']) (`pmain_r7') ("`starsmain_r7'") (`N_r7') (`R2_r7')
if _rc post `ewiz_sum' ("(R7)") (.) (.) (.) ("") (`N_r7') (`R2_r7')
}
else post `ewiz_sum' ("(R7)") (.) (.) (.) ("") (`N_r7') (`R2_r7')
if `__ewiz_model_valid_r7' == 1 {
local __included_vars "`ivvar' dfi_index roa lev size growth cashflow tobinq top1 dual board indep soe age"
foreach v of local allvars {
local __is_included = 0
foreach __inc of local __included_vars {
if "`v'" == "`__inc'" local __is_included = 1
}
if `__is_included' == 0 continue
local model_term "`v'"
if "`v'" == "`ivvar'" local model_term "`ivvar'"
capture scalar __ewiz_term_b = _b[`model_term']
capture scalar __ewiz_term_se = _se[`model_term']
if _rc | missing(__ewiz_term_se) | __ewiz_term_se <= 1e-12 {
local cell_r7_`v' "omitted"
local secell_r7_`v' "(absorbed)"
continue
}
capture local __coef : display %9.4f __ewiz_term_b
capture if abs(__ewiz_term_b) < 0.00005 & __ewiz_term_b != 0 local __coef : display %9.2e __ewiz_term_b
if !_rc {
capture local __se : display %9.4f __ewiz_term_se
capture if abs(__ewiz_term_se) < 0.00005 & __ewiz_term_se != 0 local __se : display %9.2e __ewiz_term_se
local __p = .
local __stars ""
capture local __df = e(df_r)
capture local __z = abs(_b[`model_term'] / _se[`model_term'])
if !_rc {
if "`__df'"=="" | "`__df'"=="." local __p = 2*(1-normal(`__z'))
else local __p = 2*ttail(`__df', `__z')
}
if !_rc local __stars = cond(`__p'<0.01,"***",cond(`__p'<0.05,"**",cond(`__p'<0.1,"*","")))
local cell_r7_`v' "`__coef'`__stars'"
local secell_r7_`v' "(`__se')"
}
}
}
di "(R7): `ivvar'=`coefmain_r7'`starsmain_r7' (se=`semain_r7', p=`pmain_r7'), N=`N_r7', R2=`R2_r7'"
postclose `ewiz_sum'
capture file close f_robust
file open f_robust using "$JOB_DIR/regression_table_稳健性检验.csv", write replace
file write f_robust "变量,(R1),(R2),(R3),(R4),(R7)" _n
file write f_robust ",双重聚类（个体+时间）,1% Winsor,5% Winsor,剔除异常残差 |rstudent|>3,Poisson/NB count model" _n
file write f_robust "dfi_index,`cell_r1_dfi_index',`cell_r2_dfi_index',`cell_r3_dfi_index',`cell_r4_dfi_index',`cell_r7_dfi_index'" _n
file write f_robust ",`secell_r1_dfi_index',`secell_r2_dfi_index',`secell_r3_dfi_index',`secell_r4_dfi_index',`secell_r7_dfi_index'" _n
file write f_robust "roa,`cell_r1_roa',`cell_r2_roa',`cell_r3_roa',`cell_r4_roa',`cell_r7_roa'" _n
file write f_robust ",`secell_r1_roa',`secell_r2_roa',`secell_r3_roa',`secell_r4_roa',`secell_r7_roa'" _n
file write f_robust "lev,`cell_r1_lev',`cell_r2_lev',`cell_r3_lev',`cell_r4_lev',`cell_r7_lev'" _n
file write f_robust ",`secell_r1_lev',`secell_r2_lev',`secell_r3_lev',`secell_r4_lev',`secell_r7_lev'" _n
file write f_robust "size,`cell_r1_size',`cell_r2_size',`cell_r3_size',`cell_r4_size',`cell_r7_size'" _n
file write f_robust ",`secell_r1_size',`secell_r2_size',`secell_r3_size',`secell_r4_size',`secell_r7_size'" _n
file write f_robust "growth,`cell_r1_growth',`cell_r2_growth',`cell_r3_growth',`cell_r4_growth',`cell_r7_growth'" _n
file write f_robust ",`secell_r1_growth',`secell_r2_growth',`secell_r3_growth',`secell_r4_growth',`secell_r7_growth'" _n
file write f_robust "cashflow,`cell_r1_cashflow',`cell_r2_cashflow',`cell_r3_cashflow',`cell_r4_cashflow',`cell_r7_cashflow'" _n
file write f_robust ",`secell_r1_cashflow',`secell_r2_cashflow',`secell_r3_cashflow',`secell_r4_cashflow',`secell_r7_cashflow'" _n
file write f_robust "tobinq,`cell_r1_tobinq',`cell_r2_tobinq',`cell_r3_tobinq',`cell_r4_tobinq',`cell_r7_tobinq'" _n
file write f_robust ",`secell_r1_tobinq',`secell_r2_tobinq',`secell_r3_tobinq',`secell_r4_tobinq',`secell_r7_tobinq'" _n
file write f_robust "top1,`cell_r1_top1',`cell_r2_top1',`cell_r3_top1',`cell_r4_top1',`cell_r7_top1'" _n
file write f_robust ",`secell_r1_top1',`secell_r2_top1',`secell_r3_top1',`secell_r4_top1',`secell_r7_top1'" _n
file write f_robust "dual,`cell_r1_dual',`cell_r2_dual',`cell_r3_dual',`cell_r4_dual',`cell_r7_dual'" _n
file write f_robust ",`secell_r1_dual',`secell_r2_dual',`secell_r3_dual',`secell_r4_dual',`secell_r7_dual'" _n
file write f_robust "board,`cell_r1_board',`cell_r2_board',`cell_r3_board',`cell_r4_board',`cell_r7_board'" _n
file write f_robust ",`secell_r1_board',`secell_r2_board',`secell_r3_board',`secell_r4_board',`secell_r7_board'" _n
file write f_robust "indep,`cell_r1_indep',`cell_r2_indep',`cell_r3_indep',`cell_r4_indep',`cell_r7_indep'" _n
file write f_robust ",`secell_r1_indep',`secell_r2_indep',`secell_r3_indep',`secell_r4_indep',`secell_r7_indep'" _n
file write f_robust "soe,`cell_r1_soe',`cell_r2_soe',`cell_r3_soe',`cell_r4_soe',`cell_r7_soe'" _n
file write f_robust ",`secell_r1_soe',`secell_r2_soe',`secell_r3_soe',`secell_r4_soe',`secell_r7_soe'" _n
file write f_robust "age,`cell_r1_age',`cell_r2_age',`cell_r3_age',`cell_r4_age',`cell_r7_age'" _n
file write f_robust ",`secell_r1_age',`secell_r2_age',`secell_r3_age',`secell_r4_age',`secell_r7_age'" _n
file write f_robust "核心解释变量口径,dfi_index,dfi_index,dfi_index,dfi_index,dfi_index" _n
file write f_robust "估计模型,OLS/FE,OLS/FE,OLS/FE,OLS/FE,OLS/FE" _n
file write f_robust "控制变量,是,是,是,是,是" _n
file write f_robust "个体固定效应,是,是,是,是,否" _n
file write f_robust "时间固定效应,是,是,是,是,否" _n
file write f_robust "聚类标准误,双重聚类（个体+时间）,稳健标准误,稳健标准误,稳健标准误,稳健标准误" _n
file write f_robust "N,`N_r1',`N_r2',`N_r3',`N_r4',`N_r7'" _n
file write f_robust "R²,`R2_r1',`R2_r2',`R2_r3',`R2_r4',`R2_r7'" _n
file close f_robust
preserve
use `ewiz_summary', clear
export delimited using "$JOB_DIR/regression_table.csv", replace
drop if missing(coef)
if _N == 0 {
capture file close f_base
tempname __ewz_skipfh
file open `__ewz_skipfh' using "$JOB_DIR/regression_table_稳健性检验.csv", write replace
file write `__ewz_skipfh' "状态,值" _n
file write `__ewz_skipfh' "状态,skipped" _n
file write `__ewz_skipfh' "原因,所有规格估计失败（被解释变量无变异 / 控制变量完美共线 / 样本量不足等），无法生成有效回归表" _n
file close `__ewz_skipfh'
di as error "[regression] all specs degenerate; wrote skipped marker"
}
if _N > 0 {
gen spec_order = _n
gen lb = coef - 1.96 * se
gen ub = coef + 1.96 * se
twoway (rcap ub lb spec_order, lcolor(gs8)) (scatter coef spec_order, mcolor(navy) msymbol(D)), yline(0, lpattern(dash) lcolor(maroon)) xlabel(1 "(R1)" 2 "(R2)" 3 "(R3)" 4 "(R4)" 5 "(R7)", angle(45) labsize(small)) xtitle("Specification") ytitle("Coefficient on `ivvar'") title("Specification Comparison")
capture graph export "$JOB_DIR/robustness_coef_plot.png", replace width(1800)
}
restore
di "[EWIZ_STATS] target=dfi_index;coef=`coefmain_r7';se=`semain_r7';pvalue=`pmain_r7';nobs=`N_r7'"
di "回归表输出完成"
* ── 安慰剂检验：随机化 IV 500 次 ──
tempname ph_fh
capture file close `ph_fh'
file open `ph_fh' using "$JOB_DIR/placebo_results.csv", write replace
file write `ph_fh' "iter,placebo_coef,placebo_p" _n
set seed 20250101
local placebo_iters = 500
local true_coef = `coefmain_r7'
local above_count = 0
preserve
forval __ph = 1/`placebo_iters' {
capture drop __ph_iv
gen double __ph_iv = runiform()
* 保持分布形态：用真实 IV 的百分位映射随机数
quietly summarize `ivvar', detail
local __ph_min = r(min)
local __ph_max = r(max)
replace __ph_iv = `__ph_min' + (`__ph_max' - `__ph_min') * __ph_iv
capture quietly reg `dvvar' __ph_iv `controls', vce(robust)
if _rc == 0 {
local __ph_c = _b[__ph_iv]
local __ph_se = _se[__ph_iv]
local __ph_p = 2*(1 - normal(abs(`__ph_c'/`__ph_se')))
if abs(`__ph_c') >= abs(`true_coef') {
local above_count = `above_count' + 1
}
file write `ph_fh' "`__ph',`__ph_c',`__ph_p'" _n
}
}
restore
file close `ph_fh'
local placebo_p_emp = `above_count' / `placebo_iters'
tempname ph_summary
capture file close `ph_summary'
file open `ph_summary' using "$JOB_DIR/placebo_summary.csv", write replace
file write `ph_summary' "指标,值" _n
file write `ph_summary' "真实 IV 系数,`true_coef'" _n
file write `ph_summary' "打乱迭代次数,`placebo_iters'" _n
file write `ph_summary' "打乱后|系数|≥真实|系数|的次数,`above_count'" _n
file write `ph_summary' "经验 p 值（双尾）,`placebo_p_emp'" _n
file close `ph_summary'
di "安慰剂检验完成：经验 p = `placebo_p_emp'"
* ── 经济显著性：标准化 β 与弹性 ──
quietly summarize `ivvar', detail
local iv_sd = r(sd)
local iv_mean = r(mean)
quietly summarize `dvvar', detail
local dv_sd = r(sd)
local dv_mean = r(mean)
local std_beta = (`true_coef' * `iv_sd') / `dv_sd'
local elasticity = (`true_coef' * `iv_mean') / `dv_mean'
tempname es_fh
capture file close `es_fh'
file open `es_fh' using "$JOB_DIR/economic_significance.csv", write replace
file write `es_fh' "指标,值,解读" _n
file write `es_fh' "原系数,`true_coef'," _n
file write `es_fh' "IV 标准差,`iv_sd'," _n
file write `es_fh' "DV 标准差,`dv_sd'," _n
file write `es_fh' "标准化 β,`std_beta',IV 每变动一个标准差 DV 变动多少个标准差" _n
file write `es_fh' "样本均值弹性,`elasticity',IV 变动 1%% DV 变动百分数" _n
file close `es_fh'
di "经济显著性：标准化 β = `std_beta'； 均值弹性 = `elasticity'"
log close实际输出表
这张表是本方法页配套案例的输出结果，用来把前面的公式、变量口径和代码对应到实际数字。
| 变量 | (R1) | (R2) | (R3) | (R4) | (R7) |
|---|---|---|---|---|---|
| 双重聚类（个体+时间） | 1% Winsor | 5% Winsor | 剔除异常残差 |rstudent|>3 | Poisson/NB count model | |
| dfi_index | 0.5660*** | 0.5530*** | 0.5583*** | 0.5660*** | 0.2957*** |
| ( 0.0639) | ( 0.0451) | ( 0.0471) | ( 0.0460) | ( 0.0201) | |
| roa | 0.3040*** | 0.2924*** | 0.2729*** | 0.3040*** | 0.1482*** |
| ( 0.0458) | ( 0.0436) | ( 0.0409) | ( 0.0449) | ( 0.0191) | |
| lev | -0.0548 | -0.0545 | -0.0614 | -0.0548 | -0.0132 |
| ( 0.0488) | ( 0.0456) | ( 0.0437) | ( 0.0473) | ( 0.0231) | |
| size | 0.2375*** | 0.2323*** | 0.2186*** | 0.2375*** | 0.1092*** |
| ( 0.0460) | ( 0.0458) | ( 0.0443) | ( 0.0461) | ( 0.0223) | |
| growth | -0.0819 | -0.0857* | -0.0804* | -0.0819 | -0.0198 |
| ( 0.0715) | ( 0.0494) | ( 0.0471) | ( 0.0509) | ( 0.0232) | |
| cashflow | 0.0407 | 0.0406 | 0.0379 | 0.0407 | 0.0197 |
| ( 0.0455) | ( 0.0450) | ( 0.0442) | ( 0.0456) | ( 0.0215) | |
| tobinq | 0.0247 | 0.0243 | 0.0078 | 0.0247 | 0.0146 |
| ( 0.0374) | ( 0.0477) | ( 0.0457) | ( 0.0488) | ( 0.0208) | |
| top1 | -0.0166 | -0.0271 | -0.0363 | -0.0166 | -0.0230 |
| ( 0.0460) | ( 0.0491) | ( 0.0474) | ( 0.0498) | ( 0.0229) | |
| dual | -0.0272 | -0.0287 | -0.0364 | -0.0272 | -0.0032 |
| ( 0.0513) | ( 0.0448) | ( 0.0435) | ( 0.0459) | ( 0.0197) | |
| board | -0.0383 | -0.0302 | -0.0218 | -0.0383 | -0.0021 |
| ( 0.0657) | ( 0.0434) | ( 0.0427) | ( 0.0447) | ( 0.0242) | |
| indep | -0.0309 | -0.0268 | -0.0167 | -0.0309 | -0.0075 |
| ( 0.0360) | ( 0.0480) | ( 0.0461) | ( 0.0491) | ( 0.0208) | |
| soe | 0.0025 | 0.0023 | -0.0013 | 0.0025 | -0.0055 |
| ( 0.0397) | ( 0.0491) | ( 0.0469) | ( 0.0503) | ( 0.0236) | |
| age | -0.0539 | -0.0523 | -0.0428 | -0.0539 | -0.0061 |
| ( 0.0414) | ( 0.0444) | ( 0.0433) | ( 0.0450) | ( 0.0222) | |
| 核心解释变量口径 | dfi_index | dfi_index | dfi_index | dfi_index | dfi_index |
| 估计模型 | OLS/FE | OLS/FE | OLS/FE | OLS/FE | OLS/FE |
| 控制变量 | 是 | 是 | 是 | 是 | 是 |
| 个体固定效应 | 是 | 是 | 是 | 是 | 否 |
| 时间固定效应 | 是 | 是 | 是 | 是 | 否 |
| 聚类标准误 | 双重聚类（个体+时间） | 稳健标准误 | 稳健标准误 | 稳健标准误 | 稳健标准误 |
| N | 720 | 720 | 720 | 720 | 720 |
| R² | 0.4098 | 0.4034 | 0.3885 | 0.4098 | . |
补充输出
下面这些文件来自同一次案例生成流程，用来补齐主表之外的诊断信息。
economic_significance.csv
| 指标 | 值 | 解读 |
|---|---|---|
| 原系数 | .2957 | |
| IV 标准差 | 1.014028344931995 | |
| DV 标准差 | 1.346176410712608 | |
| 标准化 β | .2227406298388962 | IV 每变动一个标准差 DV 变动多少个标准差 |
| 样本均值弹性 | -.0092732140896711 | IV 变动 1%% DV 变动百分数 |
placebo_results.csv
| iter | placebo_coef | placebo_p |
|---|---|---|
| 1 | .0422755054569741 | .0970925062563102 |
| 2 | -.0065886697061428 | .8008666111798852 |
| 3 | .0040773466649033 | .8727249389093115 |
| 4 | .0367103784152566 | .1388927961891919 |
| 5 | .0338720727626226 | .1744955386427338 |
| 6 | .0316865554752359 | .2420682323029761 |
| 7 | -.045687137571697 | .0695783024294538 |
| 8 | -.0421014733843154 | .0931455454485262 |
| 9 | -.0005652588075882 | .9820618340989864 |
| 10 | .0150228098974096 | .544185211418162 |
| 11 | .012211326861772 | .6467912938829625 |
| 12 | -.0494454278266741 | .0440870374472586 |
| 13 | .0350923725880614 | .1806690339590573 |
| 14 | -.0055093124090006 | .8300929012336378 |
| 15 | .0278231223840121 | .2654802124103861 |
| 16 | .005068688753756 | .8444113990905013 |
| 17 | .0179324524164106 | .468420924875327 |
| 18 | -.029792581403417 | .2489301642154378 |
| 19 | .0063703468936055 | .8019263946215645 |
| 20 | .0121415432401664 | .628011576333392 |
| 21 | .0257868915876518 | .2964418241316977 |
| 22 | -.0249144499412879 | .3311852520231255 |
| 23 | .0228448723389608 | .3848312934454989 |
| 24 | .0197430782200771 | .4419589760747307 |
| 25 | -.033662822970117 | .1868657196671086 |
| 26 | .006494889433034 | .8061992116001064 |
| 27 | -.0011002707828248 | .9670925975219595 |
| 28 | -.0105787285247736 | .6757942035032509 |
| 29 | .0154102174570747 | .5549087014038609 |
| 30 | -.018559019091525 | .4633283943526829 |
| 31 | .01748290355602 | .4908557570304632 |
| 32 | .0205975826156778 | .4094703461632099 |
| 33 | .0320784798988639 | .2464671443443844 |
| 34 | -.0104049567516386 | .6795776080894909 |
| 35 | .0137289053644033 | .5942775899892003 |
| 36 | -.0015537527351706 | .9531343678062938 |
这里只展示前 36 行；页面保留截断预览，避免过长表格影响阅读。
placebo_summary.csv
| 指标 | 值 |
|---|---|
| 真实 IV 系数 | .2957 |
| 打乱迭代次数 | 500 |
| 打乱后|系数|≥真实|系数|的次数 | 0 |
| 经验 p 值（双尾） | 0 |
regression_table.csv
| spec | coef | se | pvalue | stars | N | R2 |
|---|---|---|---|---|---|---|
| (R1) | .566040838796411 | .063915245 | .00030515241 | *** | 720 | .40979999 |
| (R2) | .5530192008772081 | .045082174 | 6.2751406e-31 | *** | 720 | .4034 |
| (R3) | .5582563962233911 | .047093663 | 3.4581854e-29 | *** | 720 | .38850001 |
| (R4) | .5660408387964105 | .04600032 | 4.3109577e-31 | *** | 720 | .40979999 |
| (R7) | .29570926818386 | .020119438 | 0 | *** | 720 |
案例图
这是一张由同一份案例数据生成的页面内诊断图。
论文里怎么写
本文在共用企业面板样本上报告稳健性检验，核心输出见 regression_table_稳健性检验.csv。结果解释时同时关注样本口径、变量构造、系数方向、标准误和适用前提，避免只凭单个 p 值完成方法选择。
检查清单
- 每个稳健性检验都要对应一个具体风险：变量度量、样本窗口、极端值或标准误。
- 稳健性不是重新寻找显著结果；方向、量级和理论解释是否稳定更重要。
- 不一致的稳健性结果要解释原因，不能只保留支持主结论的表。