stata-regression

name	stata-regression
description	Generates rigorous, reproducible Stata regression workflows that follow World Bank DIME Analytics conventions (`ietoolkit`, `iefolder`, master do-files, `ieboilstart`, dynamic absolute paths, `iebaltab`, `ieddtab`, `esttab`). Defaults to modern estimators (`reghdfe`, `ivreg2`/`ivreghdfe`, `csdid`, `eventstudyinteract`, `did_imputation`, `boottest`, `rdrobust`) with weak-IV-robust inference and wild cluster bootstrap when needed. Use when the user asks for OLS, logit, fixed effects, panel regression, DiD, IV, RDD, event study, balance tables, esttab/outreg2, reghdfe, csdid, ivreg2, boottest, ietoolkit, iefolder, master do-file, or publication-ready Stata output.
workflow_stage	analysis
compatibility	["claude-code","cursor","codex","gemini-cli"]
author	JonasWeinert
version	2.0.0
tags	["stata","regression","dime","ietoolkit","reghdfe","csdid","did","iv","rdd","econometrics"]

Stata Regression

Generate rigorous, reproducible Stata code that follows DIME Analytics impact-evaluation conventions and uses modern econometric estimators. The default style is ietoolkit + reghdfe + esttab, with heterogeneity-robust DiD and weak-IV-robust inference baked in.

Operating Principles

1. Reproducibility before convenience. Every script starts with ieboilstart, sets a seed, declares dynamic absolute paths via globals, and writes outputs to a Results/ folder. Never cd, never use relative paths, never commit .dta files containing PII.
2. Modern estimators by default. Plain xi: reg y treat##post under staggered timing is biased; use csdid, eventstudyinteract, did_imputation, or jwdid. For weak instruments, use ivreg2 ... weakid and weakivtest. With few clusters, use boottest.
3. ietoolkit is the house standard. Use iebaltab for balance, ieddtab for simple two-period DiD, iefolder for project setup, iegraph for regression-result figures, and iedropone for safe drops.
4. One do-file, one purpose. Master do-file installs packages and routes to sub-do-files for cleaning, construction, analysis, and output. Each script must run end-to-end from a clean Stata session.
5. Tables are produced, never copy-pasted. All tables go through esttab to .tex, .csv, or .rtf — no manual editing.

Decision Policy

This skill follows the repo-wide Agent Policy.

ASK before proceeding (blocking):

1. Estimand and identification strategy.
2. Treatment timing (uniform vs staggered) — drives csdid / eventstudyinteract / did_imputation vs plain TWFE.
3. Cluster level for inference (matches assignment level by default).
4. DataWork folder layout: which globals to use, where do tables/figures land.

DEFAULT + flag (use this default; tell the user how to override):

• reghdfe over xtreg / areg for high-dim FE.
• csdid (Callaway-Sant'Anna) as preferred staggered-DiD estimator; report TWFE as benchmark in tab_did_estimators.
• ivreg2 + weakivtest with cluster-robust SEs for IV; switch to condivreg when effective F < 100.
• boottest wild cluster bootstrap when G < 30 or treatment is concentrated in a few clusters.
• esttab for tables; coefplot for figures; iebaltab for balance.
• Outputs to ${tabs}/ (.tex) and ${figs}/ (.pdf), declared by master do-file globals.

DOCUMENT and proceed (write into the decisions log):

• Specific estimator within an approved family.
• The cluster level and the number of clusters G.
• Any sample restriction beyond what the user named.
• Choice of bandwidth in RDD and sensitivity range.

PROCEED items: ieboilstart, dynamic absolute paths via globals, isid, iedropone, assert, set seed, never cd, never copy-paste numbers from the Results window.

Pre-flight Checklist

Before writing code, confirm with the user — and write the answers as a header comment block:

• Estimand. ATT, LATE, ITT, sharp/fuzzy RD effect, descriptive, or balance check?
• Unit of observation and panel keys. Assert with isid (e.g. isid hhid year).
• Treatment timing. Single shock, simultaneous, staggered, or reversible?
• Identifying variation. Which comparison drives the coefficient?
• Clustering level. Where is treatment assigned, or where does within-group correlation live?
• Sample. Document each filter and report the resulting N.
• Output target. Working paper, slides, journal submission, or internal report.

DIME Project Layout

Every analysis script assumes a DataWork/ folder created by iefolder:

ProjectABC/
├── DataWork/
│   ├── MasterDoFile.do            # globals, packages, routes to sub-master files
│   ├── Baseline/
│   │   ├── DataSets/
│   │   │   ├── Raw/
│   │   │   ├── Intermediate/
│   │   │   └── Final/
│   │   ├── Dofiles/
│   │   │   ├── Cleaning/
│   │   │   ├── Construction/
│   │   │   └── Analysis/
│   │   ├── Output/
│   │   │   ├── Tables/
│   │   │   └── Figures/
│   │   └── Documentation/
│   └── Endline/
└── README.md

Globals are user-specific and defined only in the master do-file:

* Master do-file globals
if c(username) == "jonas"      global root "/Users/jonas/Dropbox/ProjectABC"
if c(username) == "coauthor"   global root "C:/Users/coauthor/Dropbox/ProjectABC"

global dataWork  "${root}/DataWork"
global baseline  "${dataWork}/Baseline"
global results   "${baseline}/Output"

Estimator Decision Tree

Difference-in-Differences

Two periods, two groups, no staggered adoption
└── ieddtab y, t(time) treatment(treated)  // DIME-style two-period DD
    or:    reghdfe y treat_post, absorb(unit time) vce(cluster unit)

Staggered adoption (the realistic case)
└── Do NOT use plain TWFE. Choose one of:
    ├── csdid           → Callaway & Sant'Anna (2021)
    ├── eventstudyinteract → Sun & Abraham (2021)
    ├── did_imputation  → Borusyak, Jaravel & Spiess (2024)
    ├── did_multiplegt_dyn → de Chaisemartin & D'Haultfoeuille
    └── jwdid           → Wooldridge (2021), TWFE re-derivation

Continuous or non-binary treatment
└── did_multiplegt_dyn (de Chaisemartin & D'Haultfoeuille)

Instrumental Variables

Always estimate and report the first stage explicitly.

For 2SLS with diagnostics:
  ivreg2 y x_exog (endog = z), cluster(cluster_var) first

Weak-IV diagnostics:
  - ivreg2 reports Kleibergen-Paap rk Wald F
  - weakivtest after ivreg2 → Olea-Pflueger effective F
  - condivreg or rivtest → Anderson-Rubin / CLR confidence sets

Effective F (Olea-Pflueger):
  - F >= 100 → conventional t is fine
  - 10 <= F < 100 → use AR or tF-adjusted CI
  - F < 10  → instrument is weak; do not interpret 2SLS causally

Panel IV with high-dim FE:
  ivreghdfe y x_exog (endog = z), absorb(unit time) cluster(cluster_var) first

Regression Discontinuity

Sharp RDD:
  rdrobust y x, c(0)                            // Calonico-Cattaneo-Titiunik
  rdplot   y x, c(0) binselect(esmv)
  rddensity x, c(0)                              // density manipulation test

Fuzzy RDD:
  rdrobust y x, c(0) fuzzy(treatment)

Bandwidth sensitivity:
  Loop over h_opt/2, h_opt, 2*h_opt and report the table.

Inference

Number of clusters G:
  G >= 50      → vce(cluster cl) is fine
  30 <= G < 50 → reg2hdfe or boottest with t-statistic correction
  G < 30       → wild cluster bootstrap via boottest, type(rademacher)

Multi-way clustering:
  Use vce2way (or `, vce(cluster cl1 cl2)` in some commands).
  Only when there is real correlation along both dimensions.

Few-cluster small-sample fix in reghdfe:
  reghdfe y x, absorb(unit time) vce(cluster unit, suite(boot))

Output Skeleton

Every analysis do-file should follow this structure:

*-------------------------------------------------------------*
* Project   : ProjectABC
* Purpose   : Main DiD specification on outcome Y
* Author    : First Last (email)
* Created   : 2026-05-05
* Inputs    : ${baseline}/DataSets/Final/analysis.dta
* Outputs   : ${results}/Tables/table_did_main.tex
*             ${results}/Figures/event_study.pdf
* Estimand  : ATT of program on outcome Y
* Cluster   : at the unit level (treatment assignment)
*-------------------------------------------------------------*

* 0. Settings
ieboilstart, version(17.0)
`r(version)'

* 1. Load and assert structure
use "${baseline}/DataSets/Final/analysis.dta", clear
isid unit_id year

* 2. Sample construction (log every drop)
iedropone if missing(outcome), error  // hard-fail if pattern is unexpected

* 3. Descriptives + balance
iebaltab age gender baseline_y, grpvar(treated) ///
    save("${results}/Tables/balance.xlsx") replace

* 4. Main specification
eststo clear
reghdfe outcome treat_post, absorb(unit_id year) vce(cluster unit_id)
eststo main

* 5. Robustness ladder
reghdfe outcome treat_post, absorb(unit_id year) vce(cluster unit_id year)
eststo twoway

reghdfe outcome treat_post controls, absorb(unit_id year) vce(cluster unit_id)
eststo controls

* 6. Export
esttab main twoway controls using "${results}/Tables/table_did_main.tex", ///
    replace booktabs label se ///
    stats(N r2_within, fmt(%9.0fc %9.3f) labels("Observations" "Within R2")) ///
    star(* 0.10 ** 0.05 *** 0.01) ///
    notes("Cluster-robust SEs in parentheses.")

Common Pitfalls

• Plain xi: reg y i.treat##i.post with staggered timing — biased; use csdid or eventstudyinteract.
• Conditioning on first-stage F > 10 — biased; report Olea-Pflueger F and AR.
• vce(cluster cl) with very few clusters — undercoverage; use boottest.
• Hardcoded paths and cd — break collaboration; use globals from the master do-file.
• Copy-pasting tables into Word — kills reproducibility; always go through esttab.
• Quietly dropping observations — every drop should be logged or use iedropone.
• Saving .dta with PII or running scripts on raw confidential data — see DIME PII guidance.
• xtreg, fe for high-dim FE — slow and limited; use reghdfe.

Additional Resources

• reference.md — extended code patterns for each estimator family.
• examples/ — runnable do-files:
  • examples/master.do — DIME-style master do-file
  • examples/reghdfe_baseline.do — reghdfe + esttab workflow
  • examples/iebaltab_balance.do — balance tables with iebaltab
  • examples/csdid_staggered.do — Callaway-Sant'Anna in Stata
  • examples/eventstudyinteract_sa.do — Sun-Abraham event study
  • examples/did_imputation_bjs.do — BJS imputation estimator
  • examples/ivreg2_weak.do — IV with weakivtest
  • examples/rdrobust_sharp.do — sharp RDD with rdrobust
  • examples/boottest_few_clusters.do — wild cluster bootstrap

Requirements

• Stata >= 16 (Stata 17+ recommended for frames and faster matrix ops).
• DIME packages: ietoolkit, iefieldkit.
• Estimation: reghdfe, ivreg2, ivreghdfe, weakivtest.
• Modern DiD: csdid, did_imputation, did_multiplegt_dyn, eventstudyinteract, jwdid.
• RDD: rdrobust, rddensity.
• Inference: boottest, vce2way.
• Tables and figures: estout (provides esttab), coefplot, outreg2.

* Run once, or include with `, replace` in the master do-file.
ssc install ietoolkit,        replace
ssc install iefieldkit,       replace
ssc install reghdfe,          replace
ssc install ftools,           replace        // dependency for reghdfe
ssc install ivreg2,           replace
ssc install ivreghdfe,        replace
ssc install ranktest,         replace        // dependency for ivreg2
ssc install weakivtest,       replace
ssc install csdid,            replace
ssc install drdid,            replace        // dependency for csdid
ssc install did_imputation,   replace
ssc install did_multiplegt_dyn, replace
ssc install eventstudyinteract, replace
ssc install jwdid,            replace
ssc install rdrobust,         replace
ssc install rddensity,        replace
ssc install boottest,         replace
ssc install estout,           replace
ssc install coefplot,         replace
ssc install outreg2,          replace

References

DIME Conventions

• DIME Analytics, Stata Coding Practices
• DIME Analytics, Master Do-files
• DIME Analytics, ietoolkit
• DIME Analytics, Reproducible Research
• DIME Analytics, Data Analysis
• DIME Analytics, Difference-in-Differences
• Development Research in Practice (DIME Analytics handbook).

Methods

• Goodman-Bacon (2021). DiD with Variation in Treatment Timing. JoE.
• Callaway & Sant'Anna (2021). DiD with Multiple Time Periods. JoE.
• Sun & Abraham (2021). Estimating Dynamic Treatment Effects in Event Studies. JoE.
• Borusyak, Jaravel & Spiess (2024). Revisiting Event Study Designs. ReStud.
• Olea & Pflueger (2013). A Robust Test for Weak Instruments. JBES.
• Lee, McCrary, Moreira & Porter (2022). Valid t-ratio Inference for IV. AER.
• Calonico, Cattaneo & Titiunik (2014). Robust Nonparametric CIs for RDD. Econometrica.
• MacKinnon & Webb (2018). Wild Bootstrap for Few (Treated) Clusters. EJ.