Four Ways to Work with R Code

The Plan

• The same analysis can be run four ways
• From the console, one big script, notebook, to a main.r pipeline
• Each way has its own strengths and weaknesses

Reproducibility is an important goal

• Could someone else reproduce this result from scratch? Could you, a year from now?
• Reproducible means: fresh session, run the code, get the identical numbers

Way 1: The Console

Type a line, hit enter, see the result. Repeat.

dt <- fread(here("raw_data", "wage.csv"))
dt[, mean(wage)]
feols(log(wage) ~ educ, data = dt)

• Great for exploring — poke at the data, test an idea, check a value
• Nothing is saved. Close RStudio and the steps are gone
• You cannot reproduce what you did, because there is no record of it

Way 2: One Big Script

Move the good lines into a saved .r file and run it top to bottom.

# analysis.r — everything, in order
dt <- fread(here("raw_data", "wage.csv"))
dt[, lwage := log(wage)]
# ... 300 more lines ...
est <- feols(lwage ~ educ + exper, data = dt)

• Now it is saved, ordered, and rerunnable. A real step up from the console
• It grows. At 500 lines you cannot find anything and is hard to debug

Way 3: Notebooks

Quarto, R Markdown, or Jupyter — code, prose, and output woven together.

## Returns to Education

We estimate a wage regression.

```{r}
est <- feols(lwage ~ educ, data = dt)
etable(est)
```

The coefficient on `educ` is about 0.08.

• Code, its output, and your writing live in one document that renders to HTML or PDF
• Seems Ideal for reports, teaching, and anything a human needs to read alongside results
• These slides are a notebook

Notebooks: The Fine Print

• Hidden state: you can run cells out of order and get numbers your saved file can’t reproduce
• Awkward as the home for a big multi-stage pipeline: one giant document gets unwieldy
• Best use: a communication layer that sits on top of a pipeline

Way 4: main.r + Numbered Scripts

Split the work into small scripts, each doing one job. One file runs them in order.

# main.r — reproduce everything from scratch
library(here)

source(here("code", "01_clean_data.r"))
source(here("code", "02_make_panel.r"))
source(here("code", "03_estimation.r"))
source(here("code", "04_figures.r"))

message("main.r complete")

• Each numbered script has one responsibility and a header saying its input and output
• main.r is the recipe: run it in a fresh session and the whole analysis rebuilds
• This is the research standard

Why the Pipeline Wins for Research

• One job per script means you can find and fix the cleaning without touching estimation
• Numbered order is optional but makes the flow of data obvious: 01 feeds 02 feeds 03
• ** main.r** builds and rebuilds every number in the paper