35 Course Synthesis and Review

We shall not cease from exploration, and the end of all our exploring will be to arrive where we started and know the place for the first time.

T. S. Eliot, Four Quartets: Little Gidding (1942)

Sources

Adapted from author’s lecture notes and supporting materials for a graduate practicum in biostatistics.

35.1 Prerequisites

No Prerequisites quiz for this chapter. Quiz answers appear anyway, in case you want to self-assess what you have taken away from the book. See Section 35.8.

35.2 What the Practicum covered

We have taken the reader, over the course of this book, through eight parts:

The reproducibility mindset. Why reproducibility matters and what the word means (Chapter 2), what the HIPAA Privacy Rule permits you to share (Chapter 3), and the team norms within which all of it is practiced (Chapter 4).
The workstation. Setting up R, the editor, Git, and dotfiles (Chapter 5); the Unix shell beneath them (Chapter 6); Git for solo work (Chapter 7) and for a team (Chapter 8); and remote compute (Chapter 9).
Reproducible infrastructure. Research compendia (Chapter 10), package version pinning with renv (Chapter 11), environment reproducibility with Docker (Chapter 12), the all-in-one zzcollab framework (Chapter 13), and dependency-aware pipelines with targets (Chapter 14).
Reproducible reporting. Quarto for literate documents (Chapter 15) and the R Markdown workflow it succeeds (Chapter 16).
Data wrangling and graphics. Getting data in (Chapter 17), tidyverse wrangling (Chapter 18), type handling (Chapter 19), joins (Chapter 20), database extraction with SQL (Chapter 21), publication graphics (Chapter 22), and missing data (Chapter 23).
Analysis practice. Statistical analysis plans (2), CDISC data standards (Chapter 25), testing analytic workflows (Chapter 26), continuous integration (Chapter 27), AI-assisted coding (Chapter 28), and the SAS bridge (Chapter 29) for cross-language regulatory work.
Case studies. Palmer Penguins
1. for an end-to-end small analysis, ADNI MCI prediction (Chapter 31) for a longitudinal exercise, and the communication of a finished result (Chapter 32).
Obligations. What you owe once the analysis exists: the federal requirements and the mechanics of deposition (Chapter 33), and the failures of inference that no amount of compliance prevents (Chapter 34). Both were held to the end because both presuppose a finished artifact, one to hand over and one to be skeptical about.

Two appendices sit outside the sequence: Appendix A, for reading the base R, data.table, and arrow code that a tidyverse education does not prepare you for, and Appendix B, which records the curricular evidence on which several chapters rest their case for existing.

The parts are progressive: each builds on the previous. Reproducibility is impossible without infrastructure; infrastructure is wasted without disciplined reporting; reporting without an analysis plan is performative; and an analysis nobody can read is an analysis nobody will act on.

35.3 The habits you should take forward

The book has taught a great many tools, and tools date. The habits do not, and Figure 35.1 is the shorter thing worth carrying out of it: seven practices, and the part of the book each one came from.

flowchart TD
  H1["<b>The compendium is the unit</b><br/><i>Part III</i>"]
  H2["<b>Pin the environment</b><br/><i>Part III</i>"]
  H3["<b>Version control from day one</b><br/><i>Part II</i>"]
  H4["<b>Pre-specify, then tag</b><br/><i>Part VI</i>"]
  H5["<b>Test the statistics,<br/>not the shape</b><br/><i>Part VI</i>"]
  H6["<b>Every number in the prose<br/>is computed</b><br/><i>Part IV</i>"]
  H7["<b>Every model output is a<br/>hypothesis until verified</b><br/><i>Part VI</i>"]
  H1 --> D["<b>An analysis that survives<br/>scrutiny, and a reviewer,<br/>and a decade</b>"]
  H2 --> D
  H3 --> D
  H4 --> D
  H5 --> D
  H6 --> D
  H7 --> D

Figure 35.1: Seven habits and where the book argued for each. If the tools in this book are all replaced within a decade, which is likely, these are the commitments that survive the replacement, because each is a statement about what makes an analysis defensible rather than about any particular package.

A concise list of practices that every chapter pointed at:

Every analysis lives in a compendium. A research-compendium structure with R/, analysis/, and a Dockerfile is the unit of reproducibility.
Every compendium is Dockerized before it leaves your machine. renv pins R packages; Docker pins everything else.
Every chunk of analysis code is under version control. Git from day one; tag at submission.
Every clinically meaningful analysis has a pre-specified SAP, tagged in Git before data lock. Pre-specification is what distinguishes confirmatory from exploratory work.
Every non-trivial function has at least one unit test. Tests catch regressions on refactor and document expected behavior.
Every figure in a paper is regenerable by make render (or its equivalent). Manual steps in figure generation are reproducibility holes.
Every LLM suggestion is treated as a hypothesis until verified. AI-amplified workflow demands amplified verification.

If you take only one habit from this book, take the first: every analysis lives in a compendium. Everything else flows from it.

35.4 What this book did not cover

The scope of biostatistical computing exceeds any one book. Topic areas explicitly out of scope here but worth pursuing:

Statistical theory and inference. Covered in the companion book Statistical Computing in the Age of AI. Linear models, GLMs, mixed-effects, survival, Bayesian computation, simulation, and bootstrap each get a chapter there.
Bayesian computation. A topic area of comparable weight to this entire book. Covered briefly in the companion’s Bayesian chapter.
Modern machine learning. tidymodels, deep learning, MLOps, and the broader ML pipeline are one-book subjects on their own. The Practicum’s reproducibility habits transfer; the methods do not.
Causal inference. Pre-specification
1. is necessary but not sufficient. The potential-outcomes framework, instrumental variables, sensitivity analysis to unmeasurement, and modern econometric methods belong elsewhere.
Collaboration at industrial scale. Multi- sponsor pharma collaborations and CROs operate under CDISC conventions beyond what Chapter 25 introduces.

The omissions are deliberate. The book aimed to cover what every biostatistician needs in their first two years; it cannot cover what they may need over a career.

35.5 What to read next

A curated reading list, roughly two years of material:

Year one (consolidation).

R Packages, 2nd ed. (Wickham & Bryan, 2023), package development. Read end to end; build a package with each chapter.
R for Data Science, 2nd ed. (Wickham et al., 2023), tidyverse depth. The Practicum touched many of these chapters; the full book has more.
Advanced R, 2nd ed. (Wickham, 2019), the R language. Memory model, functional programming, environments, S4. Read when you hit a corner case.
Happy Git with R (Bryan, 2019), Git for R users in depth.
Bayesian Data Analysis, 3rd ed. (Gelman et al., 2013) or Statistical Rethinking, 2nd ed. (McElreath, 2020), Bayesian computation. Pick one based on preferred level.

Year two (specialization).

Modeling Survival Data (Therneau & Grambsch, 2000) for survival analysis depth.
Mixed-Effects Models in S and S-PLUS (Pinheiro & Bates, 2000) for the canonical mixed- models reference (still relevant in 2026).
Tidy Modeling with R (Kuhn & Silge, 2022) for the tidymodels framework.
Reproducible Research with R and RStudio, 3rd ed. (Gandrud, 2020) for the publishing pipeline.
Whatever else your domain demands. Genomics: Modern Statistics for Modern Biology (Holmes & Huber, 2019). Imaging: the various Bioconductor and FSL/AFNI tutorials. Time series: Forecasting Principles and Practice (Hyndman & Athanasopoulos, 2021).

The plan: in year one, consolidate the practicum’s coverage into deeper proficiency. In year two, specialize toward your applied area.

35.6 A note on judgment

The Practicum has emphasized tools and habits. What it could not teach directly is judgment, the small decisions about which test, which covariate, which transformation, which audience. Judgment comes from practice: working on real problems, being wrong sometimes, learning the patterns of when each tool fits.

The ‘statistician’s contribution’ callouts in each chapter were an attempt to articulate the judgments explicitly. Re-read them after you have practiced the tools for six months; they will read differently.

35.7 Sign-off

You have finished a book that took its infrastructure as seriously as its statistics. The arc was deliberate: the biostatistician of 2026 spends a substantial fraction of professional time on workflow, environment, and collaboration. A graduate course that taught only methods would leave the rest as on-the-job learning. This book covered the rest.

The habits in this book pay for themselves at about the second project. By the fifth, the investment is invisible: it is just how you work. The hardest part is the first three projects, where the new infrastructure feels like overhead. The reward is permanence: a research record that ages well, an analysis you can re-run a decade later, a paper a skeptical reviewer cannot break.

Good luck with the next analysis.

35.8 Prerequisites answers

(No quiz for this chapter; this section exists to keep cross-reference anchors consistent across the book. If you would like a self-assessment, write a two-sentence answer to each of the following.)

If a former collaborator emails in three years asking you to reproduce one of your figures, what three files do you need to have kept?
What is one practice from this book you will carry into the next project you start?
What is the next book on your reading list?