Reproducible Research Pipeline

Project roadmap

1. 01

   Design the pipeline

   ~2h

   Choose a public dataset and a research question (for example, which factors predict house prices). Write DESIGN.md: the stages (ingest, validate, clean, analyze, model, figures and report), what each produces, how outputs are versioned (a run id, the data checksum, the git commit and the package lock file), and what must be identical when rerun. Set up renv in the project. Done when: DESIGN.md describes every stage, and renv.lock exists.

2. 02

   Ingest and validate the data

   ~2h

   Write functions that read the raw files (never modifying them), record their checksums, and validate them against written rules (expected columns and types, value ranges, allowed categories and no duplicate ids, for example with the pointblank or validate package). Stop the pipeline with a clear report when validation fails. Done when: the raw data passes validation, and a deliberately broken copy fails with a clear report.

3. 03

   Build the pipeline with targets

   ~3h

   Define the stages as targets in _targets.R: cleaning, summary tables, figures with ggplot2 (saved as files and tracked with format = "file"), and a model with tidymodels (a recipe, a model and cross-validation with a fixed seed). Run it with tar_make() and view the dependency graph with tar_visnetwork(). Done when: tar_make() builds every output, and changing only a figure's code reruns only that figure.

4. 04

   Produce a versioned report

   ~2h

   Write a Quarto report that reads the pipeline's outputs (with tar_read) instead of recomputing anything, and include a reproducibility section with the run id, date, data checksums, git commit and R and package versions. Save each run's report and outputs into a folder named by run id. Done when: each run produces a complete, self-describing report in its own folder.

5. 05

   Check reproducibility

   ~2h

   Write a check that runs the pipeline twice from scratch (in a clean copy of the project, restored with renv::restore) and compares the outputs: tables and model results must match exactly, and figures by their data. Document anything that legitimately differs, such as timestamps. Done when: two clean runs produce identical results.

6. 06

   Share results with Shiny

   ~2h

   Build a small Shiny app that loads a chosen run's outputs and lets collaborators browse the tables, figures and model results, and compare two runs side by side. The app only reads saved outputs and never reruns the analysis. Done when: collaborators can open any run's results in the app.

7. 07

   Test it

   ~1.5h

   Write testthat tests for the validation rules, the cleaning functions and the version information, and add a quick run of the pipeline on a small sample of the data. Document in README.md how someone else can reproduce the results from a fresh clone. Done when: the tests pass, and README.md has complete reproduction instructions.

Resources

• DocsThe targets R package user manual
• Docsrenv: introduction
• Docstidymodels: get started
• Docspointblank: data validation
• DocsQuarto documentation