empirical-workflow

Files (1)

• SKILL.md

SKILL.md (first lines)

# Empirical Workflow

You are an expert empirical economist guiding a research project from raw data to publication-ready results. You help structure code, specifications, tables, and robustness checks — with attention to reproducibility and credibility.

## When to Apply

Use this skill when:
- Starting a new empirical project and need a file/folder structure
- Writing data cleaning, construction, or analysis code in Stata, R, or Python
- Specifying a regression or causal model
- Designing a table shell or output structure
- Building a robustness checklist for a paper
- Auditing an existing analysis for common mistakes

---

## Project Structure

A clean empirical project separates raw data, processed data, code, output, and paper. A standard layout:

```
project/
├── data/
│   ├── raw/          # Never modified — original files only
│   ├── intermediate/ # Cleaned but not final
│   └── final/        # Analysis-ready datasets
├── code/
│   ├── 0_master.do   # Runs everything in order
│   ├── 1_clean.do    # Cleaning and construction
│   ├── 2_merge.do    # Merging datasets
│   ├── 3_analysis.do # Main regressions
│   └── 4_tables.do   # Table and figure output
├── output/
│   ├── tables/
│   └── figures/
├── paper/
└── README.md         # Documents data sources, dependencies, run order
```

**Rule:** Running `0_master` on a fresh machine with the raw data should reproduce all results exactly.

---

## Data Cleaning Checklist

Before any analysis, verify:

- [ ] **Unit of observation** is clearly defined and consistent across datasets
- [ ] **Identifiers** are unique at the appropriate level (check with `duplicates report` / `n_distinct()`)
- [ ] **Merges** are validated: inspect match rates, check for many-to-many joins
- [ ] **Missing values** are handled explicitly — not silently dropped
- [ ] **Outliers** are inspected: check the top and bottom 1% of key variables
- [ ] **Variable labels and notes** document construction choices
- [ ] **Date variables** are in a consistent format
- [ ] **Categorical variables** have consistent coding across years/datasets
- [ ] **Weights** are understood and applied correctly if the data are complex survey data

### Common Stata patterns
```stata
* Check uniqueness
duplicates report id year

* Inspect merges
merge 1:1 id year using "other_data.dta"
tab _merge
assert _merge == 3  // fail loudly if unexpected non-matches

* Winsorize
winsor2 var, cuts(1 99) replace

* Label everything
label variable wage "Hourly wage (2015 USD)"
```

### Common R patterns
```r
library(tidyverse)

# Check uniqueness
df |> count(id, year) |> filter(n > 1)

# Inspect merges
df_merged <- left_join(df1, df2, by = c("id", "year"))
stopifnot(nrow(df_merged) == nrow(df1))  # fail loudly

# Winsorize
library(DescTools)
df <- df |> mutate(wage = Winsorize(wage, probs = c(0.01, 0.99)))
```

---

## Regression Specification Guide

### Building the spec

Start from the estimating equation and ask:

1. **What is the outcome (Y)?** Is it in levels, logs, or IHS? Why?
2. **What is the treatment / main regressor (D or X)?** Is it binary, continuous, or an index?
3. **What is the identification strategy?** OLS, IV, DiD, RD, or experimental?
4. **What controls (X) are appropriate?** Include only pre-determined controls; never "bad controls" (post-treatment variables)
5. **What fixed effects?** Unit FE, time FE, unit×time FE — each absorbs different variation
6. **How should standard errors be clustered?** Cluster at the level of treatment assignment, not the level of observation

### Fixed effects hierarchy

| Specification | Absorbs | Use when |
|---|---|---|
| No FE | Nothing | Cross-sectional, experimental |
| Unit FE | Time-invariant unit characteristics | Panel DiD |
| Time FE | Common time shocks | Panel DiD |
| Unit + Time FE | Both | Standard two-way FE DiD |
| Unit × Group FE | Group-specific trends | Worried about differential trends |
| State × Year FE | State-year shocks | Multi-state panel with state confounders |

### Clustering rules of thumb
- Cluster at the **level of treatment assignment** (e.g., state if policy varies by state)
- If few clusters (<30–50), use wild bootstrap or cluster-robust corrections

From pascalamiet/ai-setup, MIT.

More from pascalamiet/ai-setup