brycewang-stanford/Auto-Empirical-Research-Skills / skills/14-luischanci-claude-code-research-starter/dot-claude/skills/stata
stata
VerifiedComprehensive Stata reference for writing correct .do files, data management, econometrics, causal inference, graphics, Mata programming, and 17+ community packages (reghdfe, estout, did, rdrobust, etc.). Covers syntax, options, gotchas, and idiomatic patterns. Use this skill whenever the user asks you to write, debug, or explain Stata code.
packages/did.md
View on GitHub# Modern Difference-in-Differences Packages in Stata
## Introduction
Traditional two-way fixed effects (TWFE) regression can produce biased estimates under staggered treatment adoption with heterogeneous effects. TWFE uses already-treated units as controls for newly-treated units ("forbidden comparisons"), creating negative weights.
**Modern packages solve this by:**
- Using only valid comparisons (never-treated or not-yet-treated as controls)
- Explicitly modeling treatment effect heterogeneity
- Providing transparent aggregation schemes
**Packages covered:**
- **csdid**: Callaway-Sant'Anna group-time ATTs with flexible aggregation
- **did_multiplegt**: de Chaisemartin-D'Haultfoeuille with built-in placebos and dynamics
- **did_imputation**: Borusyak-Jaravel-Spiess imputation-based estimator (most efficient)
- **didregress**: Stata 18 built-in (AIPW, IPW, RA, TWFE)
- **eventstudyinteract**: Sun-Abraham interaction-weighted estimator
- **stackedev**: Stacked event study approach
### Installation
```stata
* Core packages
ssc install csdid
ssc install drdid
ssc install did_multiplegt
ssc install did_imputation
ssc install eventstudyinteract
ssc install bacondecomp
* Supporting packages
ssc install reghdfe
ssc install ftools
ssc install avar
ssc install boottest
ssc install coefplot
ssc install event_plot
```
## Diagnosing the Problem: Goodman-Bacon Decomposition
```stata
ssc install bacondecomp
* Decompose TWFE into component 2x2 DiD comparisons
bacondecomp outcome, ddetail stub(bacon_)
* Visualize: look for negative weights and heterogeneous estimates
scatter bacon_estimate bacon_weight, ///
yline(`twfe_coef', lcolor(red) lpattern(dash)) yline(0)
```
**When is TWFE valid?** Only when treatment effects are homogeneous across units and time, and there are no dynamic effects.
## Data Structure Requirements
All modern DiD packages need:
```stata
* Key variables:
* - unit_id: Panel identifier
* - year: Time period
* - treatment_year: Year unit FIRST receives treatment
* (0 or missing for never-treated units)
* - outcome: Outcome variable
* Create binary treatment indicator
generate treated = (year >= treatment_year & treatment_year > 0)
```
---
## csdid: Callaway and Sant'Anna (2021)
Estimates group-time ATTs (ATT(g,t)) and aggregates them flexibly.
### Syntax
```stata
csdid outcome [covariates], ///
ivar(unit_id) /// // Panel identifier
time(time_var) /// // Time variable
gvar(first_treat_var) /// // First treatment period (0 or . for never-treated)
[method(dripw|reg|ipw)] /// // Estimation method
[notyet | long2] // Control group choice
```
**Key options:**
- `method(dripw)`: Doubly-robust (default, most robust)
- `method(reg)`: Regression-based (faster)
- `method(ipw)`: Inverse probability weighting
- `notyet`: Use not-yet-treated as controls (default)
- `long2`: Use only never-treated as controls
- `wboot rseed(#)`: Wild bootstrap SEs
- `wboot_reps(#)`: Number of bootstrap reps
### Basic Usage
```stata
csdid wage_outcome age grade, ///
ivar(idcode) time(year) gvar(treatment_year) ///
method(dripw) notyet
```
### Aggregation (estat commands)
```stata
* Overall ATT
estat simple
matrix simple_att = r(table)
local att = simple_att[1,1]
local se = simple_att[2,1]
* By treatment cohort
estat group
csdid_plot, group
* By calendar time
estat calendar
csdid_plot, calendar
* Event study
estat event, window(-5 10) estore(cs_event)
csdid_plot, name(event_study, replace)
* Pre-trend test
estat pretrend, pre(5)
```
### Sensitivity to Control Group
```stata
* Compare not-yet-treated vs never-treated
csdid outcome, ivar(id) time(year) gvar(treat_year) method(dripw) notyet
estat simple
local att_notyet = r(table)[1,1]
csdid outcome, ivar(id) time(year) gvar(treat_year) method(dripw) long2
estat simple
local att_never = r(table)[1,1]
display "Not-yet: " `att_notyet' " Never: " `att_never'
```
---
## did_multiplegt: de Chaisemartin and D'Haultfoeuille (2020)
Heterogeneity-robust estimator with built-in placebo tests and dynamic effects. Can handle treatment switchers.
### Syntax
```stata
did_multiplegt outcome unit_id time_var treatment_var, ///
[robust_dynamic] /// // Robust SEs with dynamic effects
[dynamic(#)] /// // Number of post-treatment periods
[placebo(#)] /// // Number of pre-treatment placebo periods
[breps(#)] /// // Bootstrap replications for SEs
[cluster(varname)] /// // Cluster variable
[controls(varlist)] /// // Time-varying controls
[trends_nonparam(varname)] /// // Unit-specific nonparametric trends
[jointtestplacebo] /// // Joint test of all placebos
[only_never_switchers] // Restrict to absorbing treatment
```
### Basic Usage
```stata
did_multiplegt price hospital_id year merged, ///
robust_dynamic dynamic(5) placebo(3) breps(100) cluster(hospital_id)
* Results stored in e()
display "Instantaneous: " e(effect) " (" e(se_effect) ")"
forvalues k = 1/5 {
display "Dynamic +`k': " e(dynamic_`k') " (" e(se_dynamic_`k') ")"
}
forvalues k = 1/3 {
display "Placebo -`k': " e(placebo_`k') " (" e(se_placebo_`k') ")"
}
```
### Joint Placebo Test
```stata
did_multiplegt price hospital_id year merged, ///
robust_dynamic dynamic(5) placebo(3) breps(200) ///
cluster(hospital_id) jointtestplacebo
display "Joint placebo p-value: " e(p_jointplacebo)
```
### Plotting Dynamic Effects
```stata
* Extract results and build dataset manually
preserve
clear
local n_placebo = 3
local n_dynamic = 5
set obs `= `n_placebo' + 1 + `n_dynamic''
generate period = _n - `n_placebo' - 1
generate estimate = .
generate se = .
forvalues k = 1/`n_placebo' {
local row = `n_placebo' - `k' + 1
replace estimate = e(placebo_`k') in `row'
replace se = e(se_placebo_`k') in `row'
}
local row = `n_placebo' + 1
replace estimate = e(effect) in `row'
replace se = e(se_effect) in `row'
forvalues k = 1/`n_dynamic' {
local row = `n_placebo' + 1 + `k'
replace estimate = e(dynamic_`k') in `row'
replace se = e(se_dynamic_`k') in `row'
}
generate ci_lower = estimate - 1.96 * se
generate ci_upper = estimate + 1.96 * se
twoway (scatter estimate period) (rcap ci_lower ci_upper period), ///
xline(-0.5, lcolor(red) lpattern(dash)) yline(0) ///
xtitle("Periods Relative to Treatment") ytitle("Effect") legend(off)
restore
```
### Saving Results
```stata
did_multiplegt price hospital_id year merged, ///
robust_dynamic dynamic(5) placebo(3) breps(100) ///
cluster(hospital_id) save_results("did_multiplegt_output")
* Creates did_multiplegt_output.dta
```
---
## did_imputation: Borusyak, Jaravel, and Spiess (2021)
Imputation-based approach: imputes counterfactual outcomes for treated units using untreated observations. Most efficient estimator under homoskedasticity.
### Syntax
```stata
did_imputation outcome unit_id time_var first_treat_var, ///
[horizons(numlist)] /// // Event times (e.g., 0/10)
[pretrend(#)] /// // Pre-treatment periods to test
[autosample] /// // Automatically balance sample
[minn(#)] /// // Min obs per group-time
[controls(varlist)] /// // Control variables (partialled out)
[cluster(varname)] /// // Cluster SEs
[agg(simple|event|cohort|time)] // Aggregation type
```
### Basic Usage
```stata
did_imputation employed idcode year first_treatment, autosample minn(0)
```
### Event Study
```stata
did_imputation employed idcode year first_treatment, ///
horizons(0/10) pretrend(5) autosample cluster(idcode)
* Plot
event_plot, ///
default_look stub_lag(tau#) stub_lead(pre#) together ///
graph_opt(xtitle("Periods Since Treatment") ytitle("Effect") xlabel(-5(1)10))
```
### Pre-trend Testing
```stata
did_imputation employed idcode year first_treatment, ///
horizons(0/10) pretrend(5) autosample cluster(idcode)
* Joint test of pre-treatment coefficients
test pre1 pre2 pre3 pre4 pre5
```
### Aggregation Types
```stata
* Cohort-specific effects
did_imputation employed idcode year first_treatment, horizons(0/5) agg(cohort) autosample
* Calendar-time effects
did_imputation employed idcode year first_treatment, horizons(0/5) agg(time) autosample
```
---
## didregress: Stata 18 Built-in Command
Requires Stata 18+. Multiple estimators in one command.
### Syntax
```stata
didregress (outcome [covariates]) (treatment_var [covariates]), ///
group(unit_id) time(time_var) ///
[aipw | ipw | ra | twfe] ///
[vce(cluster varname)]
```
- Covariates in outcome equation: for RA and AIPW
- Covariates in treatment equation: for IPW and AIPW
### Usage
```stata
* AIPW (default, doubly robust)
didregress (outcome gdp_growth) (treated gdp_growth), ///
group(state) time(year) aipw vce(cluster state)
* Post-estimation aggregations
estat aggregation, event window(-5 10)
estat aggregation, cohort
estat aggregation, time
estat grlevelsof
```
**Advantages:** Official support, multiple estimators, clean post-estimation.
**Limitations:** Stata 18 only, fewer diagnostics than specialized packages, no built-in event study plots.
---
## eventstudyinteract: Sun and Abraham (2021)
Interaction-weighted estimator avoiding negative weights.
```stata
ssc install eventstudyinteract
* Create interaction terms for each cohort x relative-time
levelsof treatment_year if treatment_year > 0, local(cohorts)
foreach g of local cohorts {
forvalues k = -5/10 {
if `k' != -1 {
generate treat_`g'_`k' = (treatment_year == `g' & rel_time == `k')
}
}
}
eventstudyinteract outcome treat_*, ///
absorb(unit_id year) ///
cohort(treatment_year) ///
control_cohort(never_treated) ///
vce(cluster unit_id)
event_plot, default_look ///
graph_opt(xtitle("Event Time") ytitle("Effect"))
```
---
## Stacked Event Study (Manual)
Stack separate datasets per cohort, each with its own never-treated control group.
```stata
levelsof treatment_year if !missing(treatment_year), local(cohorts)
tempfile stacked
local first = 1
foreach cohort of local cohorts {
preserve
keep if treatment_year == `cohort' | missing(treatment_year)
generate stack_cohort = `cohort'
generate stack_id = unit_id * 10000 + `cohort'
generate stack_reltime = year - `cohort'
if `first' {
save `stacked', replace
local first = 0
}
else {
append using `stacked'
save `stacked', replace
}
restore
}
use `stacked', clear
* Create event time dummies (omit -1)
forvalues k = 5(-1)1 {
generate lead`k' = (stack_reltime == -`k')
}
forvalues k = 0/10 {
generate lag`k' = (stack_reltime == `k')
}
reghdfe outcome lead* lag*, absorb(stack_id stack_cohort#year) cluster(unit_id)
```
---
## When to Use Each Package
| Scenario | Package | Why |
|----------|---------|-----|
| Single treatment period | Any / TWFE | All equivalent |
| Flexible aggregation needed | `csdid` | Group, calendar, event study aggregations |
| Built-in placebos + dynamics | `did_multiplegt` | Comprehensive diagnostics |
| Maximum efficiency | `did_imputation` | Lowest SEs under homoskedasticity |
| Official Stata 18 | `didregress` | Integrated post-estimation |
| Cohort-specific event studies | `eventstudyinteract` | Transparent weights |
| Teaching / transparency | Stacked event study | Easy to explain |
**Decision rule:** If staggered adoption, always use at least one robust method. Report TWFE alongside for comparison. If estimates diverge, heterogeneity is present and TWFE is biased.
---
## Event Study Setup
```stata
* Relative time variable
generate event_time = year - treatment_year
replace event_time = -1000 if missing(treatment_year) | treatment_year == 0
* Event time dummies (omit -1 as reference)
forvalues k = 10(-1)2 {
generate lead`k' = (event_time == -`k')
}
forvalues k = 0/10 {
generate lag`k' = (event_time == `k')
}
```
### Binning Distant Endpoints
```stata
generate event_time_binned = event_time
replace event_time_binned = -10 if event_time < -10 & event_time != -1000
replace event_time_binned = 10 if event_time > 10
```
### Testing Pre-Trends
```stata
* Joint test of all pre-treatment coefficients
test lead10 lead9 lead8 lead7 lead6 lead5 lead4 lead3 lead2
* Or after csdid:
estat pretrend, pre(5)
```
---
## Standard Errors and Clustering
**General rule:** Cluster at the level of treatment assignment or higher.
### By Package
```stata
* csdid: auto-clusters at ivar() level; wild bootstrap available
csdid outcome, ivar(id) time(year) gvar(g) wboot rseed(12345)
* did_multiplegt: explicit cluster + bootstrap reps
did_multiplegt outcome id year treat, breps(200) cluster(state_id)
* did_imputation: cluster option
did_imputation outcome id year g, autosample cluster(id)
* didregress: standard vce()
didregress (outcome) (treat), group(id) time(year) vce(cluster state_id)
```
### Few Clusters (< 30)
```stata
* Wild cluster bootstrap (install boottest)
ssc install boottest
reghdfe outcome treatment, absorb(unit_id year) cluster(state_id)
boottest treatment, cluster(state_id) boottype(wild) reps(9999) seed(12345)
```
### Two-Way Clustering
```stata
reghdfe outcome treatment, absorb(unit_id year) vce(cluster state_id year)
```
---
## Side-by-Side Comparison of All Estimators
```stata
* 1. TWFE
reghdfe outcome treated x1 x2, absorb(unit_id year) cluster(unit_id)
local est1 = _b[treated]
* 2. Callaway-Sant'Anna
csdid outcome x1 x2, ivar(unit_id) time(year) gvar(treatment_year) method(dripw) notyet
estat simple
local est2 = r(table)[1,1]
* 3. did_multiplegt
did_multiplegt outcome unit_id year treated, robust_dynamic breps(100) cluster(unit_id)
local est3 = e(effect)
* 4. did_imputation
did_imputation outcome unit_id year treatment_year, autosample cluster(unit_id)
local est4 = _b[tau]
* 5. didregress (Stata 18)
capture didregress (outcome x1 x2) (treated x1 x2), group(unit_id) time(year) aipw
capture local est5 = _b[ATET:r1vs0.treated]
* Compare
display "TWFE: " `est1'
display "CS: " `est2'
display "DM: " `est3'
display "DI: " `est4'
```
---
## Complete Workflow Template (csdid)
```stata
clear all
set seed 98765
use "your_data.dta", clear
* 1. TWFE baseline
reghdfe outcome treated covars, absorb(unit_id year) vce(cluster unit_id)
local twfe = _b[treated]
* 2. Bacon decomposition
bacondecomp outcome, ddetail
* 3. Main estimation
csdid outcome covars, ivar(unit_id) time(year) gvar(treatment_year) ///
method(dripw) notyet
* 4. Aggregations
estat simple // Overall ATT
estat group // By cohort
estat event, window(-5 10) estore(es) // Event study
estat pretrend, pre(5) // Pre-trend test
* 5. Plots
csdid_plot, name(event_study, replace)
* 6. Robustness: never-treated controls
csdid outcome covars, ivar(unit_id) time(year) gvar(treatment_year) ///
method(dripw) long2
estat simple
* 7. Robustness: alternative methods
csdid outcome covars, ivar(unit_id) time(year) gvar(treatment_year) ///
method(reg) notyet
estat simple
csdid outcome covars, ivar(unit_id) time(year) gvar(treatment_year) ///
method(ipw) notyet
estat simple
```
## Pre-Publication Checklist
- [ ] Test parallel trends (visual + formal `estat pretrend`)
- [ ] Show event study plot
- [ ] Report Goodman-Bacon decomposition (if staggered)
- [ ] Use at least one heterogeneity-robust estimator
- [ ] Report treatment effect dynamics
- [ ] Cluster standard errors at treatment level
- [ ] Check robustness to control group choice (not-yet vs never)
- [ ] Report number of treated/control units and cohort sizes
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the work | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |