Preface
Why this book?
Most biostatistics graduate curricula teach statistical theory thoroughly. They teach practical data-analysis skill casually. Students learn likelihood theory and the fine points of hypothesis testing. Unfortunately, they arrive at their first collaborative project unsure how to set up a project directory, commit changes with git, parameterize a report, or write a defensible statistical analysis plan.
We see this omission as structural rather than local to any one program. A curriculum committee allocates credit hours to what can be examined. The practice of analysis is hard to examine. There is no closed-form answer to how a project should be organized. Nor is there a problem set that separates a compendium a stranger can rebuild from a directory of scripts that produced the right numbers once. As a result, the skills a biostatistician uses every working day are the ones the degree least prepared them for. They are learned, if at all, by apprenticeship and accident.
This book aims to close that gap. We teach the practical skills needed to do reproducible biostatistical work in 2026: the tools, the workflows, and the professional norms. We assume throughout that the reader is also studying a theory-heavy statistical computing course alongside this one, and we treat modeling as that course’s territory rather than our own.
Why now
A book of this kind would have been useful in 2015. We consider it necessary in 2026. Three developments have changed what a competent biostatistician is expected to do.
Reproducibility stopped being a virtue and became a requirement. For two decades the case for reproducible analysis was made on scientific grounds (Goodman et al., 2016; Peng, 2011). It was widely read as good advice. It was not read as an obligation. That changed when funders made reproducibility a formal requirement. The 2022 memorandum from the White House Office of Science and Technology Policy (Office of Science and Technology Policy, 2022) directs every federal research funder to require immediate public access to publications and to the data underlying them. Agencies were to implement the policy by the end of 2025. Since January 2023, every NIH application has carried a Data Management and Sharing Plan that becomes a term of the award. The FDA has required CDISC-conformant submissions since 2016. A graduate entering the field now will be asked, in their first year, to name a repository, produce a persistent identifier, and state how the computational environment will be preserved. A theory sequence does not prepare students to answer these questions. Yet the statistician is the team member most likely to be asked them. 34 Federal Requirements and Deposition covers these policies directly. The rest of the book is, in effect, the technical answer to it.
Large language models shifted the bottleneck from writing code to auditing it. By 2026 a competent model produces working R for most of the exercises in this book. It also produces a substantial share of the analysis code a junior statistician is asked to write. That changes what expertise consists of. The parts of the work a model does well are the mechanical parts this book teaches quickly. The parts it does badly are the judgments this book spends most of its pages on. A model will not tell you that your discussion section overstates what your design supports. It will not tell you that the join you just wrote inflated the denominator. Nor will it tell you that the study was not powered for the subgroup you are about to report. It will, however, write the join. By 2026 much of that code is written by an agent. An agent is a tool that reads the codebase, edits several files, and runs the result. It does more than a chat window that returns a single answer to copy. An agent lets a mistake spread faster and further. The analyst is still responsible for catching it. If we ignored these tools, students would discover their failure modes on real projects. Teaching the tools without teaching verification would be worse. 12 AI-Assisted Coding addresses the question directly. Every chapter also closes with prompts paired with an explicit verification step.
The tooling finally settled. Until recently, the tools this book teaches changed from year to year. Quarto matured and took over much of the role of R Markdown. The renv package (Ushey & Wickham, 2026) replaced packrat and became the default for project-level dependency pinning. The rocker images (Boettiger & Eddelbuettel, 2017) made containerized R routine. The targets package (Landau, 2021) gave R a dependency-aware build tool that fits how analysts actually work. A book written in 2018 would have taught tools that are now deprecated. The tools described here are stable enough to be worth learning. They are also recent enough that most working biostatisticians were never taught them.
What the peer curricula do and do not cover
Several chapters of this book are motivated by a survey of the published graduate curricula of peer United States biostatistics programs. We list the sources in Appendix B — The Peer-Program Survey so that the reader can check the claim.
The published curricula concentrate, reasonably, on statistical computing in the algorithmic sense: optimization, simulation design, the bootstrap, and Bayesian computation. The gaps tend to fall in the practical work between a method and a result a reader can trust. The command line, SQL and database extraction, principled missing-data practice, dependency-aware pipelines, continuous integration, and de-identification each appear in a minority of published curricula. Each is something a biostatistician will be asked to do within months of starting work. When a chapter exists because of such a gap, it says so at the start and cites the evidence.
Two topics deserve particular mention. Alumni report directly that their absence has a cost. The first is CDISC (28 Clinical Data Standards: CDISC, SDTM, and ADaM). Graduates entering pharmaceutical and contract-research positions encounter SDTM and ADaM in their first week and find that no course prepared them. The second is SAS (29 SAS for R Programmers), which remains the screening filter for a large share of United States industry positions, regardless of what one thinks of the language. Neither chapter argues that these standards are good. Both argue that a graduate who cannot read them is shut out of work they are otherwise qualified for.
Craft is not separable from correctness
Analysis is error-prone. Many of its errors are caught by infrastructure and missed by vigilance. That is why the infrastructure in this book exists. Consider three failures. A join silently multiplies rows. A package update changes a function’s behavior between the analysis and the revision. A figure is built from an intermediate file that was not rebuilt after the cleaning script changed. None of these is visible when it happens, and being careful does not catch them. Each is caught by an assertion, a lockfile, or a dependency graph. Reproducible infrastructure is a courtesy to future readers. Its more immediate benefit is practical. The work of a statistician consists of ten or fifteen simultaneous projects (Slade et al., 2023). The infrastructure keeps each of them from deteriorating unnoticed.
These tools also have a limit. A perfectly reproducible, fully compliant, numerically correct analysis can still mislead every reader it reaches. No automated check catches that kind of failure. 32 Ethics Beyond Compliance addresses this problem. Its argument is that none of the infrastructure in this book determines whether an analysis misleads. We place the chapter late because the argument makes the most sense after the reader has built that infrastructure.
What is different here
- Reproducibility is the organizing principle. Each chapter is framed around producing analyses that a collaborator can rerun and extend.
- Judgment is taught directly. Every chapter contains a section on what the tool cannot decide for you. These decisions separate defensible practice from competent tool use. Neither a package nor a model makes them for you.
- AI-assisted workflows are covered. Students in 2026 will use large language models in their daily work, often through coding agents that read and edit a project directly. One chapter covers using these tools responsibly in R work. Every chapter ends with prompts designed to expose their limits, each paired with a verification step.
- Real case studies. Two full analyses (Palmer Penguins and ADNI MCI prediction) are carried end-to-end, from project scaffolding to final report.
- Federal context. Biostatisticians working on federally funded research must meet NIH/NSF reproducibility requirements. A dedicated chapter covers what those requirements are and how to satisfy them.
- The collaborative reality. The biostatistician’s work is done on teams, across several projects at once, with colleagues who cannot read the code. 3 Team Science for Biostatisticians treats working this way as a skill that can be learned.
Prerequisites
Readers should have:
- An undergraduate statistics background.
- Basic R familiarity (subsetting, the pipe, writing functions). Readers who lack this should first work through the early chapters of the companion book or R for Data Science (Wickham et al., 2023).
- Access to R 4.4+, RStudio (optional), and Git.
No prior Docker, renv, or Quarto experience is assumed.
Conventions
See the Conventions page for the visual cues used throughout the book.
Assigning parts of this book
Instructors are welcome to assign individual chapters rather than the whole book. The book is designed to allow this. Every chapter opens with a card listing what it assumes, what the reader will be able to do, and an outline. A short quiz follows for readers who may be able to skip the chapter. Because the book is free, assigning a few chapters costs students nothing.
However, the chapters do depend on one another, as Figure 1.1 in 1 Introduction shows. Parts I through III give the reader the reasons, the toolkit, and data in hand. Part IV makes that work reproducible. Parts V and VI apply it to analysis plans, testing, and industry standards. A chapter from Part V or VI can be assigned on its own to a reader who already has a working environment. A reader without one will need to go back several parts. The case studies in Part VII and the capstone in Part VIII assume everything above them.
Course length and the one-quarter core
The book is written for a two-quarter sequence at about two chapters a week. The first quarter covers why reproducibility matters, the working toolkit, and getting data in and into shape. The toolkit includes renv, Docker, and the research compendium, so a project rebuilds on another machine from the first weeks. The second quarter covers Quarto documents and pipelines, analysis practice, industry standards, and the case studies, and ends in a capstone. A semester course can cover most of the book at the same pace.
A one-quarter course should use the core path below: twenty chapters, two a week for ten weeks. It includes the skills a new analyst needs in the first week on the job, and the habits the rest of the book builds on. The remaining chapters are written to be picked up later, alone or in a second course.
Outside the core, and suited to the second quarter or to independent reading: 3 Team Science for Biostatisticians, 14 APIs, Scraping, and Freezing What You Receive, 16 Factors, Strings, and Dates, 22 De-identification in Practice, 24 Reproducible Pipelines with targets, 25 Cloud Compute and Remote Servers, 28 Clinical Data Standards: CDISC, SDTM, and ADaM, 29 SAS for R Programmers, 31 Case Study: ADNI MCI Prediction, 34 Federal Requirements and Deposition, and the capstone in 35 Capstone: HF-HOME from Export to Deposit. Two chapters are optional for everyone: 5 The Unix Shell, for readers new to the terminal, and 19 Databases and SQL for Biostatisticians, for readers whose data live in a database. Students heading to the pharmaceutical industry should add 28 Clinical Data Standards: CDISC, SDTM, and ADaM and 29 SAS for R Programmers to the core.
The prose, figures, and exercises are licensed CC BY 4.0, so adaptation, translation, and excerpting are permitted with credit. See the colophon for the full license split.
Acknowledgments
This book draws heavily on Jenny Bryan’s STAT 545 (Bryan & Stephens, 2019) and Happy Git with R (Bryan, 2019), Ben Marwick and colleagues’ rrtools (Marwick, 2018; Marwick et al., 2018), Karl Broman’s Tools for Reproducible Research (Broman, 2019), Frank Harrell’s R Workflow (Harrell, 2025), and Hadley Wickham’s book family (Wickham et al., 2023; Wickham & Bryan, 2023).
We thank the graduate students whose questions shaped much of the material.
rgtlab curriculum team
La Jolla, California
Summer 2026