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.
references/difference-in-differences.md
View on GitHub# Difference-in-Differences Analysis in Stata
## Table of Contents
1. [Core Concepts and Assumptions](#core-concepts-and-assumptions)
2. [Basic Two-Period DiD](#basic-two-period-did)
3. [Two-Way Fixed Effects (TWFE) Estimation](#two-way-fixed-effects-twfe-estimation)
4. [Event Study Designs](#event-study-designs)
5. [Testing and Visualizing Parallel Trends](#testing-and-visualizing-parallel-trends)
6. [Staggered Treatment Adoption](#staggered-treatment-adoption)
7. [Heterogeneous Treatment Effects](#heterogeneous-treatment-effects)
8. [Synthetic Control Methods](#synthetic-control-methods)
9. [Triple Differences](#triple-differences)
10. [Standard Errors and Clustering](#standard-errors-and-clustering)
11. [Recent Advances and Pitfalls](#recent-advances-and-pitfalls)
12. [Complete Workflow Examples](#complete-workflow-examples)
---
## Core Concepts and Assumptions
### Key Assumptions
#### 1. Parallel Trends (Critical)
In the absence of treatment, treated and control groups would have followed parallel trends. Cannot be directly tested but can be assessed with pre-treatment data.
```stata
* Visual inspection of pre-treatment trends
preserve
keep if year < treatment_year
collapse (mean) outcome, by(treated year)
twoway (line outcome year if treated == 0) ///
(line outcome year if treated == 1), ///
legend(label(1 "Control") label(2 "Treated"))
restore
* Formal test with leads
regress outcome i.treated##ib(last).pre_period_year i.year, cluster(state)
test 1.treated#1.pre_period_year 1.treated#2.pre_period_year
```
#### 2. No Anticipation
Treatment does not affect outcomes before implementation. Violations: firms adjusting in anticipation, migration before implementation.
#### 3. Common Shocks
Time-varying shocks affect treated and control groups similarly.
```stata
* Check for differential time trends
regress outcome i.treated##c.year i.year, cluster(state)
test 1.treated#c.year // Should be insignificant
```
---
## Basic Two-Period DiD
### Regression-Based DiD
```stata
* Standard DiD regression
regress wage_outcome i.treated##i.post
* With controls
regress wage_outcome i.treated##i.post age education, robust
* Interpretation:
* - treated: baseline difference between groups
* - post: time trend for control group
* - treated#post: DiD estimate (treatment effect)
```
### Manual DiD Calculation
```stata
quietly: summarize wage_outcome if treated == 1 & post == 1
local y11 = r(mean)
quietly: summarize wage_outcome if treated == 1 & post == 0
local y10 = r(mean)
quietly: summarize wage_outcome if treated == 0 & post == 1
local y01 = r(mean)
quietly: summarize wage_outcome if treated == 0 & post == 0
local y00 = r(mean)
display "DiD Estimate: " (`y11' - `y10') - (`y01' - `y00')
```
---
## Two-Way Fixed Effects (TWFE) Estimation
TWFE extends DiD to multiple periods and/or multiple treatment groups.
### Panel Data Setup
```stata
xtset unit_id year
* TWFE regression with xtreg
xtreg outcome post_treatment, fe vce(cluster unit_id)
* Equivalent using reghdfe (faster for large datasets)
ssc install reghdfe
reghdfe outcome post_treatment, absorb(unit_id year) vce(cluster unit_id)
* With time-varying controls
reghdfe outcome post_treatment x1 x2, ///
absorb(unit_id year) vce(cluster unit_id)
```
### Unit-Specific and Group-Specific Time Trends
```stata
* Linear unit-specific trends
reghdfe outcome post_treatment, ///
absorb(unit_id year unit_id#c.year) vce(cluster unit_id)
* Group-specific time trends
reghdfe outcome post_treatment, ///
absorb(unit_id year group#c.year) vce(cluster unit_id)
```
---
## Event Study Designs
### Basic Event Study
```stata
* Create relative time variable
generate rel_time = year - treatment_year
replace rel_time = -999 if missing(treatment_year) // Never treated
* Create event time dummies (omit -1 as base period)
quietly: summarize rel_time
local min = r(min)
local max = r(max)
forvalues k = `min'/`max' {
if `k' != -1 & `k' != -999 {
generate lead_lag_`k' = (rel_time == `k')
}
}
* Event study regression
reghdfe outcome lead_lag_*, absorb(unit_id year) vce(cluster unit_id)
* Plot event study coefficients
preserve
parmest, saving(event_study.dta, replace)
use event_study.dta, clear
keep if regexm(parm, "lead_lag")
generate event_time = real(regexr(parm, "lead_lag_", ""))
sort event_time
twoway (scatter estimate event_time) ///
(rcap min95 max95 event_time), ///
xline(-0.5, lcolor(red) lpattern(dash)) ///
yline(0, lcolor(black)) ///
xlabel(-5(1)5) ///
xtitle("Periods Relative to Treatment") ///
ytitle("Treatment Effect") legend(off)
restore
```
### Event Study with Binned Endpoints
```stata
* Bin distant periods to reduce noise
generate rel_time_binned = rel_time
replace rel_time_binned = -5 if rel_time < -5 & rel_time != -999
replace rel_time_binned = 5 if rel_time > 5
forvalues k = -5/5 {
if `k' != -1 {
generate event_`k' = (rel_time_binned == `k')
}
}
reghdfe outcome event_*, absorb(unit_id year) vce(cluster unit_id)
```
### Automated Event Study (csdid)
```stata
ssc install csdid
ssc install drdid
* Callaway and Sant'Anna (2021)
csdid outcome x1 x2, ivar(unit_id) time(year) gvar(treatment_year) ///
method(dripw) notyet
* Plot event study
estat event, window(-5 5) estore(cs_results)
csdid_plot, name(event_study, replace)
* Pre-trend test
estat pretrend
```
---
## Testing and Visualizing Parallel Trends
### Visual Inspection
```stata
preserve
collapse (mean) outcome, by(treated year)
twoway (connected outcome year if treated == 0, msymbol(circle)) ///
(connected outcome year if treated == 1, msymbol(square)), ///
xline(`treatment_year', lcolor(red)) ///
legend(order(1 "Control" 2 "Treated"))
restore
```
### Formal Tests for Pre-Trends
```stata
reghdfe outcome event_*, absorb(unit_id year) vce(cluster unit_id)
* Joint test of pre-treatment leads
test event_-5 event_-4 event_-3 event_-2
```
### Placebo Tests
```stata
* Assign fake treatment earlier, use only pre-treatment data
generate placebo_treatment = (year >= treatment_year - 3) if year < treatment_year
keep if year < treatment_year
reghdfe outcome placebo_treatment, absorb(unit_id year) vce(cluster unit_id)
* Should find no significant effect
```
---
## Staggered Treatment Adoption
### The Problem with TWFE
Standard TWFE with staggered adoption uses already-treated units as controls ("forbidden comparisons"), can have negative weights, and is biased under heterogeneous treatment effects.
### Goodman-Bacon Decomposition
```stata
ssc install bacondecomp
bacondecomp outcome, ddetail
bacondecomp outcome, stub(bacon_)
scatter bacon_estimate bacon_weight, ///
xlabel(0(0.1)0.5) ylabel(-2(1)2) xline(0) yline(0)
```
### Callaway and Sant'Anna (2021) Estimator
```stata
ssc install csdid
* Group-time average treatment effects
csdid outcome, ivar(unit_id) time(year) gvar(treatment_year) ///
method(dripw) notyet
* Aggregations
estat simple // Overall ATT
estat group // By cohort
estat calendar // By time period
estat event, window(-5 5) estore(event_results) // Event study
* With covariates
csdid outcome x1 x2, ivar(unit_id) time(year) gvar(treatment_year) ///
method(dripw) notyet
* Use never-treated as control instead of not-yet-treated
csdid outcome, ivar(unit_id) time(year) gvar(treatment_year) ///
method(dripw) long2
```
### Sun and Abraham (2021) Estimator
```stata
ssc install eventstudyinteract
eventstudyinteract outcome treat_*, ///
cohort(treatment_year) control_cohort(never_treated) ///
absorb(unit_id year) vce(cluster unit_id)
event_plot, default_look graph_opt(xtitle("Periods to Treatment") ///
ytitle("Average Effect") xlabel(-5(1)5))
```
### de Chaisemartin and D'Haultfoeuille (2020)
```stata
ssc install did_multiplegt
did_multiplegt outcome unit_id year treatment, ///
robust_dynamic dynamic(5) placebo(3) breps(100) cluster(unit_id)
* With controls
did_multiplegt outcome unit_id year treatment, ///
controls(x1 x2) robust_dynamic dynamic(5) breps(100) cluster(unit_id)
* Results: Placebo_X (pre-treatment, should be 0), Effect (at treatment),
* Dynamic_X (X periods after treatment)
matrix list e(effect)
matrix list e(N_effect)
```
### Imputation Methods
```stata
ssc install did_imputation
* Borusyak, Jaravel, Spiess (2021)
did_imputation outcome unit_id year treatment_year, ///
autosample minn(0)
* Event study version
did_imputation outcome unit_id year treatment_year, ///
horizons(0/5) pretrend(5)
event_plot, default_look
```
---
## Heterogeneous Treatment Effects
### Heterogeneity by Subgroup
```stata
did_multiplegt outcome unit_id year treatment if subgroup == 1, ///
robust_dynamic dynamic(5) breps(100) cluster(unit_id)
estimates store subgroup1
did_multiplegt outcome unit_id year treatment if subgroup == 2, ///
robust_dynamic dynamic(5) breps(100) cluster(unit_id)
estimates store subgroup2
estimates table subgroup1 subgroup2
```
### Quantile DiD
```stata
ssc install didq
didq outcome, tr(treated) time(post) quantile(0.25 0.50 0.75)
didq outcome, tr(treated) time(post) quantile(0.1(0.1)0.9)
didq_plot
```
---
## Synthetic Control Methods
### Basic Synthetic Control
```stata
ssc install synth
tsset unit_id year
synth outcome outcome(1990) outcome(1995) outcome(2000) ///
x1 x2 x3, ///
trunit(treated_unit_id) trperiod(treatment_year) ///
keep(synth_results, replace) fig
* View weights
matrix list e(W_weights) // Unit weights
matrix list e(V_matrix) // Predictor weights
```
### Placebo Tests for Synthetic Control
```stata
* In-space placebo: Apply to each control unit
local control_units "2 3 4 5 6"
foreach unit of local control_units {
capture synth outcome outcome(1990) outcome(1995) x1 x2, ///
trunit(`unit') trperiod(treatment_year) ///
keep(placebo_`unit', replace)
}
* In-time placebo: Fake treatment date
local fake_treatment = treatment_year - 5
synth outcome outcome(1990) outcome(1995) x1 x2, ///
trunit(treated_unit) trperiod(`fake_treatment') ///
keep(placebo_time, replace)
```
### Inference via RMSPE Ratios
```stata
use synth_results, clear
* Pre-treatment RMSPE
generate sq_error = (_Y_treated - _Y_synthetic)^2 if _time < treatment_year
quietly: summarize sq_error
local rmspe_pre = sqrt(r(mean))
* Post-treatment RMSPE
replace sq_error = (_Y_treated - _Y_synthetic)^2 if _time >= treatment_year
quietly: summarize sq_error
local rmspe_post = sqrt(r(mean))
local rmspe_ratio = `rmspe_post' / `rmspe_pre'
display "RMSPE Ratio: " `rmspe_ratio'
* Compare to placebo distribution of RMSPE ratios for p-value
```
---
## Triple Differences
Triple differences (DDD) adds a third difference to control for group-specific trends.
```stata
* Three dimensions: Treatment group x Time x Third dimension
generate ddd = treated * post * third_group
reghdfe outcome treated##post##third_group, ///
absorb(state year demographic) vce(cluster state)
* Coefficient on three-way interaction is DDD estimate
```
### Example: Minimum Wage by Age Group
```stata
generate teen = (age >= 16 & age <= 19)
generate treated_state = inlist(state, "CA", "NY", "WA")
generate post_policy = (year >= 2016)
reghdfe employment treated_state##post_policy##teen ///
age education, ///
absorb(state year age_group) vce(cluster state)
* Coefficient on treated_state#post_policy#teen is the DDD estimate
```
---
## Standard Errors and Clustering
### Clustering Considerations
```stata
* Cluster at treatment level (most conservative)
reghdfe outcome post_treatment, absorb(unit_id year) vce(cluster state)
* Two-way clustering
reghdfe outcome post_treatment, absorb(unit_id year) vce(cluster state year)
* Multi-way clustering (requires reghdfe)
reghdfe outcome post_treatment, absorb(unit_id year) vce(cluster state year industry)
```
### Wild Bootstrap for Few Clusters
```stata
ssc install boottest
reghdfe outcome post_treatment, absorb(unit_id year) cluster(state)
boottest post_treatment, cluster(state) boottype(wild) reps(10000)
```
### Randomization Inference
```stata
* Save actual estimate
reghdfe outcome post_treatment, absorb(unit_id year)
local actual_coef = _b[post_treatment]
* Permutation test
set seed 12345
local n_perm = 1000
matrix perm_coefs = J(`n_perm', 1, .)
forvalues i = 1/`n_perm' {
preserve
sort unit_id
generate random = runiform()
bysort year: egen rank = rank(random)
generate perm_treat = (rank <= number_treated)
quietly: reghdfe outcome perm_treat, absorb(unit_id year)
matrix perm_coefs[`i', 1] = _b[perm_treat]
restore
}
svmat perm_coefs
count if abs(perm_coefs1) >= abs(`actual_coef')
local p_value = r(N) / `n_perm'
display "Randomization inference p-value: " `p_value'
```
### Spatial HAC Standard Errors
```stata
ssc install ols_spatial_HAC
ols_spatial_HAC outcome post_treatment x1 x2, ///
lat(latitude) lon(longitude) time(year) ///
distcutoff(500) lagcutoff(5)
```
---
## Recent Advances and Pitfalls
### Common Pitfalls
**1. Negative Weights in TWFE**
```stata
* Check with Bacon decomposition, then use robust estimators
bacondecomp outcome
csdid outcome, ivar(unit_id) time(year) gvar(treatment_year)
```
**2. Violation of Parallel Trends** -- Consider: adding controls, matching on pre-treatment trends, synthetic control, or unit-specific trends (`absorb(unit_id#c.year)`).
**3. Clustering Too Low** -- Cluster at treatment variation level. Too few clusters? Use wild bootstrap.
**4. Choosing Wrong Comparison Group**
```stata
* Not-yet-treated (time-varying control)
csdid outcome, ivar(unit_id) time(year) gvar(treatment_year) notyet
* Never-treated (time-invariant control)
csdid outcome, ivar(unit_id) time(year) gvar(treatment_year) long2
```
### Stacked DiD
```stata
* Create separate datasets for each treatment cohort, stack them
local cohorts "2010 2012 2015"
tempfile stacked
local first = 1
foreach cohort of local cohorts {
use original_data, clear
keep if treatment_year == `cohort' | missing(treatment_year)
generate cohort = `cohort'
generate cohort_id = unit_id * 10000 + `cohort'
if `first' {
save `stacked', replace
local first = 0
}
else {
append using `stacked'
save `stacked', replace
}
}
use `stacked', clear
reghdfe outcome post_treatment, ///
absorb(cohort_id cohort#year) vce(cluster unit_id)
```
---
## Complete Workflow Examples
### Example 1: State Minimum Wage Policy
```stata
use "state_employment.dta", clear
* Define treatment
generate treated = inlist(state_name, "California", "New York", "Massachusetts")
generate treatment_year = 2014 if treated == 1
generate post = (year >= 2014) if treated == 1
replace post = 0 if treated == 0
encode state_name, generate(state_id)
xtset state_id year
* Visual trends
preserve
collapse (mean) teen_employment, by(treated year)
twoway (connected teen_employment year if treated == 0, msymbol(O)) ///
(connected teen_employment year if treated == 1, msymbol(S)), ///
xline(2014, lcolor(red) lpattern(dash)) ///
legend(order(1 "Control States" 2 "Treated States"))
restore
* Basic DiD
reghdfe teen_employment i.treated##i.post, ///
absorb(state_id year) vce(cluster state_id)
estimates store basic_did
* Event study
generate rel_time = year - treatment_year
replace rel_time = -999 if missing(treatment_year)
forvalues k = 5(-1)2 {
generate lead`k' = (rel_time == -`k')
}
forvalues k = 0/5 {
generate lag`k' = (rel_time == `k')
}
reghdfe teen_employment lead5 lead4 lead3 lead2 lag0-lag5, ///
absorb(state_id year) vce(cluster state_id)
* Pre-trend test
test lead5 lead4 lead3 lead2
* Callaway-Sant'Anna (robust to staggered adoption)
csdid teen_employment gdp_growth population_growth, ///
ivar(state_id) time(year) gvar(treatment_year) ///
method(dripw) notyet
estat simple
estat event, window(-5 5) estore(cs_event)
csdid_plot
* Placebo test (fake treatment at t-2)
preserve
keep if year < 2014
generate placebo_post = (year >= 2012)
reghdfe teen_employment i.treated##i.placebo_post, ///
absorb(state_id year) vce(cluster state_id)
restore
```
### Example 2: Synthetic Control for Single Treatment
```stata
use "smoking.dta", clear
encode state, generate(state_id)
tsset state_id year
local treated_id = 3 // California
local treatment_year = 1989
* Synthetic control estimation
synth cigsale cigsale(1980) cigsale(1988) ///
lnincome(1980(1)1988) retprice(1980(1)1988) age15to24(1980(1)1988), ///
trunit(`treated_id') trperiod(`treatment_year') ///
keep(synth_ca, replace) replace fig
matrix list e(W_weights)
* Gap plot
use synth_ca, clear
generate effect = _Y_treated - _Y_synthetic
twoway (line effect _time, lcolor(black)), ///
xline(`treatment_year', lcolor(red) lpattern(dash)) ///
yline(0, lcolor(gray)) ///
xtitle("Year") ytitle("Gap in Cigarette Sales")
* In-space placebos for inference
* (Run synth for each control unit, compare RMSPE ratios)
```
### Example 3: Staggered DiD with did_multiplegt
```stata
use "paid_leave.dta", clear
bysort state_id (year): egen first_treat = min(year) if paid_leave == 1
bysort state_id: egen treatment_year = min(first_treat)
* Standard TWFE (potentially biased)
reghdfe female_lfp paid_leave gdp_pc unemployment_rate, ///
absorb(state_id year) vce(cluster state_id)
estimates store twfe
* Bacon decomposition to diagnose
bacondecomp female_lfp, ddetail
* Heterogeneity-robust: did_multiplegt
did_multiplegt female_lfp state_id year paid_leave, ///
robust_dynamic dynamic(10) placebo(5) breps(100) cluster(state_id)
* Heterogeneity-robust: Callaway-Sant'Anna
csdid female_lfp gdp_pc unemployment_rate, ///
ivar(state_id) time(year) gvar(treatment_year) ///
method(dripw) notyet
estat simple
estat event, window(-5 10) estore(cs_event)
* Wild cluster bootstrap for TWFE
reghdfe female_lfp paid_leave gdp_pc unemployment_rate, ///
absorb(state_id year) cluster(state_id)
boottest paid_leave, cluster(state_id) boottype(wild) reps(999)
```
---
## Summary: When to Use Each Method
| Method | Use Case |
|--------|----------|
| Basic DiD (2x2) | Two periods, two groups, uniform treatment timing |
| TWFE | Multiple periods/groups; cautious with staggered timing |
| Event Study | Visualize dynamic effects, test parallel trends |
| `csdid` / `did_multiplegt` / `did_imputation` | Staggered adoption, heterogeneous effects |
| Synthetic Control | Single or few treated units, rich pre-treatment data |
| Triple Differences | Additional difference controls for group-specific trends |
### Essential Packages
```stata
ssc install reghdfe
ssc install ftools
ssc install csdid
ssc install drdid
ssc install did_multiplegt
ssc install did_imputation
ssc install eventstudyinteract
ssc install bacondecomp
ssc install synth
ssc install boottest
ssc install avar
ssc install coefplot
ssc install event_plot
```
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: 322k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 322k | 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 |