R Econometrics

R Econometrics

Generate rigorous, modern, reproducible R code for causal inference and panel econometrics. Default to estimators that are robust to the failure modes of 2010-era TWFE, weak instruments, and naive cluster-robust inference.

Operating Principles

1. Identification before code. State the estimand, the source of identifying variation, and the threats to identification before writing any model. Write these as comments at the top of the script.
2. Modern estimators by default. Plain TWFE under staggered treatment is biased; weak IV with F just above 10 is unreliable; RDD without rdrobust understates uncertainty. Use heterogeneity-robust DiD, weak-IV-robust inference, and bias-corrected RDD unless the user explicitly opts out.
3. Match inference to design. Cluster at the level of treatment assignment or sampling. With few clusters (rule of thumb < 30), use wild cluster bootstrap, not cluster-robust SEs.
4. Make every script runnable. Set seeds, declare paths, write to results/, separate one-time data prep from estimation, and never silently drop observations.
5. Show the alternatives. When two reasonable estimators exist (TWFE vs Callaway-Sant'Anna; CR1 vs wild bootstrap), report both and explain divergence.

Decision Policy

This skill follows the repo-wide Agent Policy.

ASK before proceeding (blocking):

1. Estimand and identification strategy (RCT / IV / DiD / RDD / descriptive).
2. Treatment timing — uniform shock vs staggered adoption (drives the choice between TWFE and CS / SA / BJS).
3. Cluster level for inference (treatment-assignment level by default).
4. Sample restrictions to apply (each is irreversible without re-running).

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

• feols from fixest over lm for panel models.
• Heterogeneity-robust DiD (Callaway-Sant'Anna by default) when timing is staggered; TWFE only when the user opts in.
• rdrobust with MSE-optimal bandwidth for sharp RDD; report sensitivity at h/2, h, 2h.
• Wild cluster bootstrap (fwildclusterboot) when there are fewer than 30 clusters.
• Output to results/ (tables) and results/figures/ (vector PDF).

DOCUMENT and proceed (write into the decisions log):

• The chosen estimator within an approved family (e.g. CS vs SA vs BJS).
• The omitted period in event studies (default: -1).
• Inference type (CR1, CR2, wild bootstrap) and cluster level.
• Any composition shifts that emerged (e.g. dropped never-treated cohorts because too small).

PROCEED items: set.seed(20240101), here::here() paths, theme_paper() for figures, asserting panel structure with n_distinct before estimation.

Pre-flight Checklist

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

• Estimand. ATT, LATE, ITT, sharp/fuzzy RD effect, or descriptive?
• Unit of observation and panel keys. Assert uniqueness (fixest::demean will silently aggregate; dplyr::n_distinct(unit, time) should equal nrow(df) for a balanced panel).
• Treatment timing. Single shock, simultaneous adoption, staggered adoption, or reversible?
• Identifying variation. Which comparisons drive the coefficient? What is the implicit control group?
• Clustering level. Where is treatment assigned, or where does within-group correlation live?
• Sample restrictions. Document every filter and report the resulting N at each step.
• Output target. Working paper, slides, journal submission, or exploratory.

Estimator Decision Tree

Difference-in-Differences

Is treatment timing the same across all treated units?
├── YES (one shock, two groups)
│   └── feols(y ~ treat_post | unit + time, cluster = ~unit)   # OK
│
└── NO (staggered adoption)
    ├── Are treatment effects plausibly homogeneous?
    │   └── If you really believe this: TWFE with caveats.
    │
    └── Otherwise (the realistic case):
        ├── Callaway & Sant'Anna  (did::att_gt + aggte)
        ├── Sun & Abraham         (fixest::sunab)
        ├── Borusyak/Jaravel/Spiess (didimputation::did_imputation)
        └── de Chaisemartin & D'Haultfoeuille (DIDmultiplegt)

   Report at least two estimators when possible.

Instrumental Variables

Always estimate and report the first stage explicitly.

Effective first-stage F (Olea-Pflueger):
├── F >= 100  → conventional 2SLS inference is reliable
├── 10 <= F < 100 → conventional t is unreliable; report:
│   ├── Anderson-Rubin confidence set (ivmodel, ivDiag)
│   └── tF-adjusted CI (Lee 2022)
└── F < 10  → instrument is weak; do not interpret 2SLS as causal
              without weak-IV-robust inference.

Panel IV: feols(y ~ x1 | fe | endog ~ z, data = df)
First stage: summary(model, stage = 1)

Regression Discontinuity

Sharp RDD:
├── Default: rdrobust::rdrobust(y, x, c = cutoff)
├── Visualization: rdrobust::rdplot(y, x, c = cutoff)
├── Density manipulation test: rddensity::rddensity(x, c = cutoff)
└── Bandwidth sensitivity: report estimates over [h/2, h, 2h]

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

Never pick a bandwidth by eye. Use rdbwselect or rdrobust's MSE-optimal default.

Inference

Number of clusters G:
├── G >= 50    → vcov = cluster-robust (CR1) is fine
├── 30 <= G < 50 → use clubSandwich::vcovCR(..., type = "CR2")
└── G < 30     → use wild cluster bootstrap (fwildclusterboot::boottest)

Multi-way clustering only when there is real correlation along
both dimensions (e.g. industry x year shocks). Adding a second
dimension always inflates SEs and is not "more conservative" — it
is a different estimator.

Output Skeleton

Every analysis script should follow this structure:

# 0. Header: estimand, identification, sample, clustering, data version
# 1. Setup: packages, seed, paths
# 2. Load + assert panel structure (isid-equivalent)
# 3. Sample construction with documented filters and Ns
# 4. Descriptives (Table 1, balance table)
# 5. Main specification(s)
# 6. Robustness ladder (alt FE, alt cluster, alt sample, alt estimator)
# 7. Diagnostics (pre-trends, first-stage, density, placebo)
# 8. Tables: etable() or modelsummary() to results/
# 9. Figures: iplot(), coefplot() to results/
# 10. Save fitted models with versioned filenames

Common Pitfalls

• Plain TWFE with staggered timing — biased when effects are heterogeneous (Goodman-Bacon 2021, de Chaisemartin & D'Haultfoeuille 2020).
• Reporting first-stage F > 10 as a sufficient pre-test — biased; use AR or tF (Andrews, Stock & Sun 2019; Lee 2022).
• Cluster-robust SEs with a handful of clusters — undercoverage; use wild bootstrap (Cameron, Gelbach & Miller 2008; MacKinnon & Webb 2018).
• Picking an RDD bandwidth by eye — use rdrobust / rdbwselect.
• Using lm() for panel data — slow, no within transform, fragile clustering. Use feols.
• Quietly dropping observations — log every filter and report sample sizes.
• Treating a coefficient on treated * post as the "DiD" without thinking about composition when treatment is staggered.

Additional Resources

• reference.md — extended code patterns for each estimator family.
• examples/ — runnable R scripts:
  • examples/did_twfe_baseline.R — single-shock DiD with fixest
  • examples/did_callaway_santanna.R — staggered DiD with did
  • examples/event_study_sun_abraham.R — heterogeneity-robust event study
  • examples/iv_weak_instruments.R — first-stage diagnostics, AR set, tF
  • examples/rdd_robust.R — sharp and fuzzy RDD with rdrobust
  • examples/wild_cluster_bootstrap.R — inference with few clusters

Requirements

• R >= 4.1
• Core: fixest, modelsummary, dplyr, ggplot2
• Modern DiD: did, didimputation, DIDmultiplegt
• IV: ivmodel, ivDiag
• RDD: rdrobust, rddensity
• Inference: clubSandwich, fwildclusterboot, sandwich, lmtest

install.packages(c(
  "fixest", "modelsummary", "dplyr", "ggplot2",
  "did", "didimputation", "DIDmultiplegt",
  "ivmodel", "ivDiag",
  "rdrobust", "rddensity",
  "clubSandwich", "fwildclusterboot", "sandwich", "lmtest"
))

References

• Goodman-Bacon (2021). Difference-in-Differences with Variation in Treatment Timing. JoE.
• Callaway & Sant'Anna (2021). Difference-in-Differences with Multiple Time Periods. JoE.
• Sun & Abraham (2021). Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Effects. JoE.
• Borusyak, Jaravel & Spiess (2024). Revisiting Event Study Designs: Robust and Efficient Estimation. ReStud.
• de Chaisemartin & D'Haultfoeuille (2020). Two-way Fixed Effects Estimators with Heterogeneous Treatment Effects. AER.
• 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 Confidence Intervals for RDD. Econometrica.
• MacKinnon & Webb (2018). The Wild Bootstrap for Few (Treated) Clusters. EJ.
• Roth, Sant'Anna, Bilinski & Poe (2023). What's Trending in Difference-in-Differences? JoE.