new-analysis

name	new-analysis
description	Scaffolds a method-specific analysis notebook (DiD, IV, RDD, LASSO, Panel FE) with boilerplate. Use when starting a new econometric analysis.
argument-hint	<method> [title]
allowed-tools	Bash, Read, Write, Edit, Glob, Grep
version	1.1.0
workflow_stage	analysis
tags	["econometrics","notebook","DiD","IV","RDD","panel-data"]

Scaffold Analysis Notebook

Create a new notebook pre-populated with method-specific boilerplate for a common econometric technique.

Arguments

• $ARGUMENTS — the method name and optional title (e.g., "DiD Event Study", "IV Analysis of Colonial Origins", "RDD Minimum Wage", "LASSO Variable Selection", "Panel FE Growth Regressions")

Steps

1. Parse the method from the arguments. Recognized methods:

   • DiD (difference-in-differences)
   • IV (instrumental variables)
   • RDD (regression discontinuity design)
   • LASSO (regularized regression / variable selection)
   • Panel FE (panel fixed effects)
   • If the method is not recognized, ask the user to clarify.

2. Follow the same notebook creation conventions as /project:new-notebook:

   • Check notebooks/ for existing files to determine the next sequential number
   • Ask the user for the kernel: Python, R, or Stata
   • Create the .qmd with YAML frontmatter (title and jupyter kernel) and the setup cell
   • Include pedagogical markdown narrative explaining each step

3. Add method-specific sections as markdown and code cells:

   All methods include these sections:

   • Data Loading (code cell)
   • Variable Construction (code cell)
   • Summary Statistics (descriptive table exported to ../tables/)
   • Estimation (regression table built manually as pipe-delimited Markdown)
   • Visualization (figure exported at 6×4 inches, 300 DPI, to ../images/)
   • Robustness Checks (markdown header + empty code cell)

   All figures: 6 inches wide × 4 inches tall, 300 DPI, exported to ../images/. All tables: Export to ../tables/ in three formats (CSV + Markdown + LaTeX). Build regression tables manually — do NOT use pf.etable(type="md"), etable(markdown=TRUE), or esttab md as their output doesn't render correctly in the manuscript.

   Method-specific boilerplate:

   • DiD: parallel trends test (event-study coefficients for pre-treatment periods), event study plot, TWFE regression, staggered treatment note. Staggered treatment warning: if treatment timing varies across units, naive TWFE can produce biased estimates due to negative weights. Use did package (R, Callaway & Sant'Anna 2021), csdid (Stata), or did2s (Python, Gardner 2022) instead. Include Bacon decomposition to diagnose heterogeneity in treatment timing.
   • IV: first-stage regression with F-statistic (rule of thumb: F > 10; use Kleibergen-Paap rk Wald F when SEs are clustered), reduced-form regression (as a credibility check — if reduced form is insignificant, IV results are suspect), 2SLS estimation, Anderson-Rubin confidence intervals (robust to weak instruments), Hansen J overidentification test (when overidentified). Always report first-stage F prominently.
   • RDD: running variable histogram (check for manipulation), McCrary density test (required — manipulation of the running variable invalidates the design), bandwidth selection (Calonico-Cattaneo-Titiunik optimal bandwidth preferred over Imbens-Kalyanaraman), local polynomial estimation with robust bias-corrected CIs, RD plot with binned scatter and fitted polynomials both sides of cutoff. Include bandwidth sensitivity plot and donut-hole RDD as robustness checks.
   • LASSO: cross-validation for lambda, coefficient path plot, selected variables, post-LASSO OLS (Belloni, Chernozhukov & Hansen 2014)
   • Panel FE: within estimator, entity and time FE, clustered standard errors (cluster at the entity level by default — clustering at a lower level underestimates SEs), Hausman test (FE vs RE). Report within-R² (not overall R²) as the relevant goodness-of-fit measure.

4. Include language-specific package guidance in the notebook preamble:

   • Python: pyfixest for fixed effects and DiD, linearmodels for IV and panel models, rdrobust for RDD, scikit-learn for LASSO
   • R: fixest (preferred for all FE models — supports sunab() for staggered DiD, feols() for IV), rdrobust for RDD, ivreg for IV, glmnet for LASSO
   • Stata: reghdfe for high-dimensional FE, ivregress for IV, rdrobust for RDD, csdid for staggered DiD, lasso2 for LASSO

5. Register in _quarto.yml under manuscript.notebooks:

   - notebook: notebooks/<name>.qmd
     title: "<title>"
   

6. To embed in index.qmd:

   • Figures: {{< embed notebooks/<name>.qmd#fig-label >}}
   • Tables: {{< include tables/<label>.md >}}

7. Confirm the notebook renders: quarto render notebooks/<name>.qmd

8. Report the file path and list the embed-ready labels created.

Common Pitfalls

• Staggered DiD with TWFE: Naive two-way fixed effects produces biased estimates when treatment timing varies — negative weights on already-treated units contaminate the ATT. Always check for staggered adoption and use appropriate estimators.
• Weak instruments: A first-stage F-statistic below 10 signals weak instruments. Standard 2SLS inference is unreliable — use Anderson-Rubin CIs or LIML instead.
• RDD bandwidth sensitivity: Results that hold only at the MSE-optimal bandwidth are fragile. Always show estimates across a range of bandwidths (50%–200% of optimal).
• Clustering level: Cluster standard errors at the level of treatment assignment. Clustering at a finer level (e.g., individual when treatment is at the state level) underestimates SEs and inflates significance.
• Pre-trends in DiD: Statistically insignificant pre-trend coefficients do not prove parallel trends — low power can mask violations. Discuss the magnitude of pre-trend estimates relative to the treatment effect.

References

• Callaway & Sant'Anna (2021) "Difference-in-Differences with Multiple Time Periods" — staggered DiD
• de Chaisemartin & D'Haultfoeuille (2020) "Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects" — TWFE bias
• Andrews, Stock & Sun (2019) "Weak Instruments in IV Regression" — weak instrument diagnostics
• Cattaneo, Idrobo & Titiunik (2019) "A Practical Introduction to Regression Discontinuity Designs" — RDD best practices
• Belloni, Chernozhukov & Hansen (2014) "Inference on Treatment Effects after Selection among High-Dimensional Controls" — post-LASSO