Appendix E — Instructor Guide
This appendix is for instructors. It collects the material that explains why the book is built as it is, where each chapter’s content comes from, and how its parts can be assigned. Students can skip it.
E.1 Course length
The book is written for a two-quarter sequence at about two chapters a week. The first quarter covers Parts I to III: why reproducibility matters and with whom, the working toolkit, and getting data in and into shape. The second quarter covers Parts IV to VIII: documents and environments, analysis practice, industry standards, and the case studies, ending in the capstone (Chapter 34). For a one-quarter course, the preface gives a core path of twenty chapters, two a week for ten weeks, and lists the chapters suited to a second course or independent reading.
E.2 Why the book covers what it covers
Several chapters exist because comparable graduate programs teach the topic and the author’s students reported needing it in their first jobs. The survey of peer programs behind those decisions is in Appendix B.
E.3 Chapter sources and rationale
Each entry records the lecture notes, courses, and prior material a chapter draws on, and the reason it is in the book. This material previously opened each chapter; it moved here so that chapters open with the problem they solve. Entries follow the book’s order, Parts I to VIII, followed by the two appendices that this edition split from chapters.
E.3.1 Why Reproducible Research
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics. The cases are from Baggerly and Coombes (2009), the Institute of Medicine (2012) report on the Duke trials, and Herndon, Ash, and Pollin (2014); the taxonomy is Goodman, Fanelli, and Ioannidis (2016), with the vocabulary of the NASEM (2019) report.
Rationale. The chapter motivates the rest of the book, and every later chapter is a specific tool for clearing the reproducibility bar in a specific way. It opens with the Duke genomics case because it is a biostatistical failure with patient consequences, found by biostatisticians, and it names the book’s motto, ‘make silent failures loud’, in the cases section, pointing to the habit list in Section 1.6. The federal-mandate argument is reduced to two sentences and a pointer to Chapter 33, which owns the policy detail. The former ‘Orientation’ and ‘statistician’s contribution’ material is kept: the entropy paragraph opens Section 2.4, and the three workflow judgments are a subsection of the anatomy section. The executed ‘break it, then fix it’ example uses the HF-HOME baseline file and base R only, so it runs offline in under a second; its unseeded half deliberately produces different output on every build. The former Further reading items not kept in the chapter are useful for instructors who want more: Peng (2011) in Science; Sandve et al. (2013) and Wilson et al. (2017), two widely assigned ‘simple rules’ papers; Gentleman and Temple Lang (2007), which introduced the research compendium; Marwick’s rrtools (2018); Wilkinson et al. (2016) on the FAIR principles; and the TIER Protocol (projecttier.org), a compendium template for the social sciences. The former exercise asking students to write a one-page reproducibility policy for a research group was cut; it works well as an in-class group activity.
E.3.2 Team Science for Biostatisticians
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics; Slade et al. (Slade et al., 2023) for the survey evidence on team-science skills; the ICMJE recommendations (International Committee of Medical Journal Editors, 2024) and the CRediT taxonomy (Brand et al., 2015) for authorship.
Rationale. The technical skills elsewhere in the book are useful only if students can deploy them on a team, and the team context is governed by social and procedural conventions more than by software. The chapter is built around one recurring collaborator, Dr. X, the heart-failure investigator whose first email opens it, carried from intake (where the immortal-time and confounding red flags turn an observational question into the HF-HOME trial) through the scope of work, a disagreement, and authorship of the trial paper. Results communication was reduced to the points other chapters rely on (the three registers of a claim, effect size before p-value, confidence versus prediction intervals) and a pointer to Chapter 32, which owns the memo, talk, and reviewer response. The exercises are scenarios with model answers, so the chapter can be assigned before students have collaborations of their own; a single ‘own work’ exercise remains.
E.3.3 De-identification and Data Ethics
Sources. Karl Broman’s Tools for Reproducible Research (Broman, 2019) and the Johns Hopkins statistical-computing sequence (Johns Hopkins Bloomberg School of Public Health, 2024), the two peer curricula that treat de-identification and human-subjects protection explicitly; the HHS de-identification guidance (US Department of Health and Human Services, 2012); Sweeney (Sweeney, 2002) for the Weld case and k-anonymity.
Rationale. A biostatistician working with clinical data handles protected health information on the first day, and the obligations that attach to it are not optional. The survey of peer curricula (Appendix B) found the topic conspicuous by its absence from most programs; for a book aimed at students who will spend their careers with clinical data, the omission would be indefensible. The chapter opens with the Weld re-identification because it shows the mechanism (quasi-identifiers combine) without any technical apparatus. Its compliance content was corrected during the revision (Safe Harbor versus Expert Determination labeling, the 164.514(c) rule on re-identification codes, the limited data set with a DUA) and should be read by a second reader with privacy or compliance expertise, such as a privacy officer, before it is assigned. This edition split the chapter: sections 1 to 4 (promise, pathways, transformations, obligations) remain here in Part I, with no executed code, and the two HF-HOME releases and the residual-risk section moved to Chapter 18 in Part III. Part I gained a decisions cheat sheet and an intake-reply cumulative exercise.
E.3.4 The Unix Shell
Sources. The working subset taught by the Software Carpentry lesson ‘The Unix Shell’, by Berkeley STAT 243 (University of California, Berkeley Department of Statistics, 2024), and by Karl Broman’s reproducibility course (Broman, 2019); the peer survey (Appendix B) found the command line taught explicitly in all three and in Data Carpentry. The Practicum had previously assumed the shell rather than taught it.
Rationale. The shell is the substrate under the rest of the toolkit: Docker is driven from it, Git began as a set of shell commands, and every computing cluster is reached through it. The chapter now opens with the HF-HOME CRF export, whose planted byte-order mark and CRLF line endings make two quick grep counts return 0, so that the first thing students learn about the shell is the book’s motto in miniature: a correct command on bytes you cannot see returns a plausible wrong answer. Every closed exercise runs against the shipped HF-HOME files, and each key is computed in R from the same bytes at render time, which also teaches the habit of confirming a shell count in R (the shell-versus-R table is the bridge to Chapter 12). Shell commands that only read files run at render as bash chunks; the one that demonstrates truncation by redirection works on a copy in a mktemp -d directory and deletes it, and the check-raw.sh script is displayed once and executed from the same source through knitr::knit_code. The file command is demoted to a footnote, because versions with a CSV test (5.41 on macOS) report only ‘CSV text’ for a file with a BOM and CRLF endings, and minimal containers lack it; od -c and cat -e are taught instead. The former ‘statistician’s contribution’ judgments (script rather than click, respect destructive commands, inspect before loading, keep the shell out of the analysis proper) are placed in the sections where each arises. The former open-ended exercises 1, 2, and 4 became closed exercises on HF-HOME, so that each has a key; exercise 3 (which large files belong out of Git) is folded into the search section’s prose, and exercise 5 (compare grep with an editor’s search) was cut. The ‘echo before rm’ rule is given as a standing instruction for a coding agent (4.6). The chapter precedes Chapter 6 in the part order, so ‘You’ll need’ points Windows students to Git Bash or WSL there.
E.3.5 Setting Up Your Workstation
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics.
Rationale. A poorly configured machine does not announce itself; it appears as friction that accumulates over years, and its defects are the hardest in the book to attribute. The chapter keeps its original hook and is cut to an essentials path (R, an IDE, Quarto, Git, and renv), with per-operating-system installation in tabs that include Windows under WSL, and it ends the path with the ‘Is my machine ready?’ checklist (Table 6.4), which later chapters can cite instead of restating setup. The editor survey beyond the three IDEs, the shell configuration, starship, tmux, fonts, dotfiles, and the fresh-laptop script moved to an optional ‘Level up’ section at the end rather than to an appendix; the earlier corrections in that material (a private dotfiles repository, one R installation route, the brew shellenv line and the caveat on piping a download into bash, zsh rather than bash completion, and the font cask names) are unchanged. The chapter follows Chapter 5, which removes the forward dependency on cd and .zshrc. The restored-workspace failure, which the earlier LLM prompt named only in passing, is now an executed demonstration: an .RData file in a temporary directory and two R sessions, one restoring it and one started with --vanilla. This edition adds a short section on secrets (~/.Renviron, Sys.getenv(), and keyring) and a pointer to Posit Public Package Manager binaries for Linux; dated snapshots belong to Chapter 20. Devcontainers and pre-commit were not added here; they fit Chapter 21 and Chapter 8 better. The former open-ended exercises became exercises with stated expected results (versions that agree, find.package() locations, the restore demonstration, and a check-machine.sh script with a model solution); the dotfiles exercise was dropped with the material it exercised, and the hard-coded-paths exercise is the ‘own work’ item. The bypass quiz’s questions on dotfiles and terminal editors were replaced by questions on the restored workspace and the checklist.
E.3.6 Git and GitHub for Solo Developers
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics. The companion textbook, Statistical Computing in the Age of AI (chapter 2), covers Git’s mechanics in team contexts; this chapter covers the same tools from the solo analyst’s perspective. Jenny Bryan’s Happy Git and GitHub for the useR and her 2018 American Statistician article are the models for the argument.
Rationale. Solo Git is a harder sell than team Git, because every standard argument for version control is about collaboration. The chapter therefore opens with the March and November analysts and makes the case once, as a table of November’s questions, in place of the three overlapping ‘why’ passes of earlier drafts. The habits that make the history useful (atomic commits, messages that say why, tags for irreversible moments, a remote as backup) are taught where each command appears. The ‘Oh no, I…’ table is meant as a lookup card for students’ first months. Every exercise uses one throwaway practice repository that the student builds in Exercise 1, so that solutions can state the expected git log output; Git does not run at render time, so the keys were produced by running the exercises against Git 2.55 and abbreviating hashes, which will differ for every student. The zzgit plugin moved to a footnote beside the portable alternative (gitleaks as a pre-commit hook), and the Atlassian tutorials were dropped from Further reading.
E.3.7 Git for Teams
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics; Jenny Bryan’s Happy Git and GitHub for the useR and chapters 3 and 5 of Pro Git for the mechanics, and the pull-request and code-review material of R Packages (Wickham and Bryan), which transfers to research compendia.
Rationale. Chapter 7 taught Git for one analyst working alone, and Chapter 3 taught the social norms of working with others; between them sits the mechanics of two statisticians committing to one repository, which the continuous-integration workflow of Section 26.7 and both case studies assume. The chapter opens with the two analysts whose branches each give a correct count and whose merge gives a third count neither has seen; the three counts come from a synthetic cohort generated in the chapter’s source, so the before-and-after conflict table is computed rather than typed, and the worked example’s failing CI assertion uses the same numbers. The earlier rebuild of the conflict (a three-way disagreement whose resolution matches neither side) is kept verbatim, and the chapter also states the full difference between the resolution and each branch, including the patients the diastolic ceiling removes from her branch’s cohort, which the former LLM prompt’s verification omitted. The former ‘statistician’s contribution’ judgments are placed where each arises: conflicts as disagreements about the analysis in the conflicts section, protecting main in the branch-protection subsection, preventing conflicts in the generated-files subsection, and reviewing the analysis in the review section, whose six questions are now a table. Detail on resolving renv.lock conflicts is reduced to a pointer to Chapter 20, which does not cover it in depth either, so the short treatment here is the book’s only one; instructors who need more should supplement it. The earlier corrections on --theirs under a rebase (now a Pitfall) and on renv.lock merge=ours are unchanged. The three LLM prompts became one callout on recovering a shared branch (force push, --force-with-lease, and revert rather than rewrite), with a caution that background fetches can defeat the lease; the conflict-resolution prompt depended on the defective example, and the diff-review prompt is superseded by the checklist. The formal register (‘we would urge the reader’) is replaced by second person, as throughout the book. The exercises are checkable, each with a key: a planted conflict and a git bisect run hunt in throwaway repositories, a planted diff with one deliberate distractor (a moved set.seed() that changes no draw), and a cumulative exercise that simulates two analysts with a bare repository and two clones, so that no collaborator or GitHub account is needed. The Git output in the solutions was produced by running each exercise with Git 2.55; hashes differ for every student. The branch-protection exercise still needs GitHub, and its key describes the expected rejection. The references to Section 26.7, Chapter 26, Chapter 25, Chapter 19, Chapter 20, and Chapter 14 point forward in the current part order. R Packages was dropped from Further reading to meet the limit of three.
E.3.8 AI-Assisted Coding
Sources. The author’s blog posts 46-ellmerinRcoding (the ellmer package) and 36-simpleshinyappwithchatgpt; the ellmer package from Posit; the btw package documentation; direct experience with Claude Code during the development of this book; and the author’s review of the LLM sections across the book.
Rationale. Language models now produce working R for most of the book’s exercises, and the chapter’s position, kept from the original, is that they are an amplifier and not a replacement: every line of generated code is a hypothesis to test. The chapter is ordered so that its durable core comes first, and this edition moved it as a unit to Part II, after Git; its card names Chapter 7 and Chapter 5 as prerequisites. The core comprises the four Cs (Context, Constraints, Criteria, Check) as the organizing protocol, with a table and the verify-first loop redrawn as a four-Cs loop; the failure modes, now a table of failure, example, and check, with ten executed mini-examples on the HF-HOME files, each with a planted bug and the assertion that catches it, which point forward to Chapter 13, Chapter 14, and Chapter 26; data governance and agent safety (Section 9.3, which every chapter’s LLM callouts cite), with a what-may-go-where table, the tests-off-limits rule, and hallucinated package names (slopsquatting); project instruction files, live documentation, and provenance and journal disclosure; and ‘when not to reach for an LLM’. The perishable material (the ellmer API, coding-agent product names, the model identifier, and the btw and MCP setup) was moved in this edition to a dated appendix, Appendix F, which keeps the section id sec-ai-tooling so that every cross-reference still resolves. The earlier corrections are kept: chat_anthropic() rather than chat_claude(), chat$stream() with coro::loop(), the model identifier in one variable with an ‘as of’ comment, cor()‘s default of use = "everything", and stats::filter as the time-series function masked by dplyr::filter. The ’Prompt patterns that work’ section was folded into the four-Cs table; the ‘statistician’s contribution’ judgments into the opening of the first section; ‘Readable code is auditable code’ moved before the planted bugs, with an executed lintr example; the book’s own disclosure became the worked example of the provenance subsection. The first prerequisite question, which asked about ellmer, now asks what may be sent to a model, so that the quiz survived the move of the tooling section. Exercises have keys: a four-Cs rewrite of a one-line request, an executed planted risk-difference bug whose sign depends on factor level order, a data-classification exercise, and instruction-file rules; the former cumulative tinytest exercise, run against both versions, is relabeled ‘Looking ahead’ (return after Chapter 26), and the ellmer exercise moved to the appendix with its section. A new reference, spracklen2025package, supports the slopsquatting paragraph.
E.3.9 Getting Data In: Files
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics. The checking argument follows Harrell’s R Workflow (Harrell, 2025), Chapters 7 and 8; the import mechanics follow R for Data Science (Wickham et al., 2023).
Rationale. Every later data chapter begins with a data frame already in memory, and this chapter exists because getting it there is where a substantial share of the errors that survive into publication are introduced: a value silently misread at import is one no downstream test will catch. The chapter sets out the three ways data arrive (a file, an interface, a page) and the discipline they share, and it owns the raw-data rule (Habit 2) that Chapter 12 applies to cleaning. The file material is built on the HF-HOME raw export, so that the naive read, the specified read, and the differences between them are computed rather than asserted; the import checklist (Table 10.2) is the chapter’s reusable artifact. The byte-order-mark discussion is deliberately qualified by reader and locale, because base R strips the mark in a UTF-8 locale and keeps it in the C locale; students who meet the common claim that the mark always breaks the first column name should be shown the three reads side by side. This edition split the former chapter in two: this chapter keeps the three routes and the file half, and Chapter 11 takes the interface, scraping, and freezing material with its exercises. The former ‘Collaborating with an LLM’ section is reduced to the column-specification callout (from the codebook, not the data); the scraping prompt moved with the second half.
E.3.10 APIs, Scraping, and Freezing What You Receive
Sources. As for ‘Getting Data In: Files’, from which this edition split it.
Rationale. An interface or page is a query against a living system, so the chapter’s argument is that what is received must be frozen to a dated raw file with a provenance note before any analysis touches it; the freezing section, which separates a deliberately run acquisition script from a cleaning script that touches no network, is the chapter’s center. The LLM callout is the scraping prompt, which asks the model to argue against scraping and treats the page as untrusted input to an agent; the former JSON prompt is absorbed into the JSON Verification. The network exercise remains open (cumulative practice); the closed exercises use a fabricated REDCap-shaped payload and a local stand-in for a paginated API, so that the chapter renders offline.
E.3.11 Data Wrangling Essentials
Sources. Stat 545, Chapters 5 to 9 (Jenny Bryan, University of British Columbia), and the author’s blog posts 04-lowercasingdataframes, 15-piping, and 43-dynamic-column-names.
Rationale. Every cleaning decision (what a code means, which rows are excluded, which of two conflicting values survives) is a decision about what the analysis is of, made before any model and rarely questioned afterward. The chapter therefore teaches the decisions rather than the verbs: it opens on a missing-value code that parses as a blood pressure, and it works entirely on the HF-HOME CRF export, whose per-site sex codes, date formats, and missing-value codes are the pathologies students meet in real multisite data. The generic dplyr tour, the pandas side-by-side, and the stringsAsFactors and %>% digressions are cut in favor of a footnote to R for Data Science (Wickham et al., 2023). The worked example reports the rows removed at each filter and distinguishes that subset from the as-randomized analysis population, so that the flow table cannot be mistaken for an exclusion from the primary analysis. Closed exercises have keys computed from the running study, which ships the clean tables as an answer key; the chapter says explicitly that a real export has none. The verbs-and-shapes card (Table 12.2) is meant for the cheat-sheet appendix.
E.3.12 Factors, Strings, and Dates
Sources. Stat 545, Chapters 10 to 13 (Jenny Bryan, University of British Columbia); the forcats, stringr, and lubridate packages and their documentation; Grolemund and Wickham (2011) for lubridate.
Rationale. Numeric columns fail loudly; text, categories, and dates fail by producing plausible values, which is why they have a chapter of their own. The generic API tours of the three packages are left to R for Data Science (Wickham et al., 2023, chs. 14 to 17); the full verb list survives as the cheat sheet, and the pandas side-by-side is cut. In their place the chapter works problem-then-verb on the running study: the HF-HOME sex and site codes (factors, declared levels, and the reference level of arm), the free-text dosing strings in medications.csv (drug class, dose with decimals, units and combination tablets, frequency codes), and the per-site date formats in crf_export.csv. It deliberately builds on Chapter 12, which recodes the same export with case_when(), and answers the question that chapter defers: why a single multi-order parse_date_time() call misreads some dates without a warning, and how to count the values at risk. The ambiguity demonstration shows that the reading of a value such as 03/04/2026 depends on the order of the orders list and on the other values in the vector, which corrects the earlier wording that the first-listed order always wins. The earlier corrections (the daylight-saving example with dhours(), the as.Date() origin, factor joins matching labels) are retained. Every closed exercise has an adjacent key computed from the running study, several checked against its clean tables as answer keys.
E.3.13 Joining and Reshaping
Sources. Stat 545, Chapters 14 to 16 (Jenny Bryan, University of British Columbia); the dplyr two-table verbs reference.
Rationale. The join call is trivial; the errors are not. The chapter is organized around the two silent failure modes (dropped rows and multiplied rows) and the arithmetic that catches them: predicted row counts, declared relationships, and anti-join audits. The generic verb tour is left to R for Data Science (Wickham et al., 2023), as it is throughout the data chapters, the pandas comparison is reduced to a footnote, and rows_* is an aside. The join checklist (Table 14.1) is meant to be reused in Chapter 15 and Chapter 25; students can use it as a review rubric.
E.3.14 Databases and SQL for Biostatisticians
Sources. Authored for this book. The survey of peer programs (Appendix B) found relational databases and SQL taught as a core skill in several comparison courses, among them Berkeley STAT 243 (University of California, Berkeley Department of Statistics, 2024), UNC BIOS 735 (University of North Carolina at Chapel Hill, 2024), Michigan (University of Michigan School of Public Health, 2024b), Minnesota (University of Minnesota School of Public Health, 2024), and the Johns Hopkins statistical-computing sequence (Johns Hopkins Bloomberg School of Public Health, 2024), as well as in both Software and Data Carpentry.
Rationale. Clinical data usually reach a statistician through a query against a warehouse, REDCap, or an EHR extract rather than as a file, and the failures that matter are the ones Chapter 14 already teaches, without dplyr‘s guard rails: SQL has no relationship argument, no many-to-many warning, and aggregates that skip NULL silently. The chapter therefore loads the HF-HOME tables into an in-memory SQLite database and reuses the two planted labs.csv defects (the duplicated visit-test rows and the orphan patient 2433), which in the discharge NT-proBNP join nearly cancel and leave a count off by one. Two database-specific habits are added: parameterized queries (with an executed injection example on synthetic data) and a dated extract with its query and checksum as the raw data. The generic DBI and dbplyr tour is left to R for Data Science (Wickham et al., 2023, ch. 21); DuckDB and Parquet are a short, conditionally executed subsection because they are the current default for larger-than-memory files. The ’parity’ framing that opened the earlier draft justified the chapter to a curriculum committee rather than to a reader, and is retained here only.
E.3.15 Plotting with ggplot2 and purrr
Sources. Stat 545, Part VII (Jenny Bryan, University of British Columbia); the author’s blog posts 16-plotsfrompurrr and 38-tableplacementrmarkdown; zzlongplot for longitudinal examples (Chapter 30).
Rationale. A figure is the part of a paper most readers examine closely, and its defaults are decisions made by someone who never saw the data. The chapter opens with a before-and-after: a bar of means redrawn as points with a robust summary, on the running study’s NT-proBNP. It then builds one longitudinal trajectory figure layer by layer (individual patients, summary and interval, the number measured at each visit), so that the figure is the chapter’s spine rather than one example among many. The former ‘statistician’s contribution’ material is distributed to where each judgment arises and collected as the six-item figure checklist (Table 16.3), which students can use as a review rubric. The generic ggplot2 tour, the penguins examples, the regression-diagnostic figure, and the matplotlib side-by-side are cut in favor of R for Data Science (Wickham et al., 2023); DPI guidance is reduced to a reference table. zzlongplot is presented as an optional in-house tool beside the portable ggplot2 function, consistent with the book’s policy on in-house packages. Closed exercises check the plotted numbers against independent dplyr summaries with layer_data().
E.3.16 Missing Data: Diagnosis, Imputation, Reporting
Sources. The survey of peer US biostatistics MS programs (Appendix B) found missing-data handling taught as a core or near-core topic at a majority of them, including dedicated courses at Michigan (University of Michigan School of Public Health, 2024a) and the University of Washington (Sadinle, 2019), and coverage within the curricula at Emory (Emory University Rollins School of Public Health, 2024), Yale (Yale School of Public Health, 2024), Iowa (University of Iowa College of Public Health, 2024), UT Health Houston (UT Health Houston School of Public Health, 2024), and Florida (University of Florida College of Public Health and Health Professions, 2025).
Rationale. Every real clinical dataset has missing values, and the decisions around them often move point estimates more than the model choice does. The chapter opens with the running study’s unascertained outcomes, computed from the HF-HOME files, and introduces the three mechanisms by story before notation. Its central claim, corrected in the revision, is that the MCAR/MAR/MNAR label does not decide whether complete-case analysis is biased; what decides it is whether missingness depends on the outcome given the model’s covariates. The simulation that shows this (bias, standard error, and coverage for complete case, mean imputation, and multiple imputation under three deletion processes) now sits in the body beside the mechanism table rather than in a solution, and a Verification callout fails the build if its sign pattern changes. The section on missing outcomes in trials, added in this edition, starts from the estimand, cites the National Research Council (2010) report, runs an MMRM on the mmrm package’s fev_data, shows reference-based imputation with rbmi as display-only code (the package is not in the book’s library), and works a two-dimensional tipping point on the HF-HOME files. That worked example also shows that the SAP’s one-directional tipping rule has nothing to tip when the primary result is inconclusive, which instructors can use to discuss pre-specifying sensitivity analyses that do not presume the direction of the result. The limit-of-detection aside and the reporting-standards detail moved to ‘Going further’; the sample reporting paragraph and the pre-specification subsection were cut in favor of Chapter 25. The former Further reading items not kept in the chapter remain useful for instructors: the mice documentation at amices.org/mice; ICH E9(R1) on estimands; Rubin (1976); Harrell’s author checklist; and Harrell’s R Workflow, chapter 6, on describing missingness before choosing an imputation model. The former open exercise (write the missing-data section of a SAP for a hypothetical trial with 15% dropout) is folded into the cumulative exercise against HF-HOME.
E.3.17 De-identification in Practice
Sources. As for Chapter 4, from which this edition split it: the HHS de-identification guidance (US Department of Health and Human Services, 2012) and Sweeney (Sweeney, 2002) for k-anonymity. The epigraph is from Paul Ohm, ‘Broken Promises of Privacy’ (UCLA Law Review, 2010); its wording should be checked against the article before the chapter ships.
Rationale. The concepts of Chapter 4 can be taught in the first week, but building a release requires the single-table verbs, dates, and joins of Part III, so the executed half of the former chapter now follows the data chapters. It opens on a computed uniqueness count (2,422 of 2,432 HF-HOME patients are unique on ZIP code, date of birth, and sex) and carries the two HF-HOME releases from the former chapter: a Safe Harbor release, and a linkable release with per-patient date shifts whose need for an Expert Determination is shown rather than asserted, each with assertions that fail if an identifier survives or an interval is not preserved. The residual-risk section computes k-anonymity and finds the records that set it, so that the cost of a generalization is weighed against the analytic resolution it removes. The exercises and the code cheat sheet moved with this material. The restricted ZIP prefixes in the code (036, 059, 821) are only those the HF-HOME generator plants; the chapter deliberately does not print the full HHS list, which students must take from the current guidance. The review by a second reader with privacy or compliance expertise that Chapter 4 asks for applies here too.
E.3.18 Reproducible Reports with Quarto
Sources. Stat 545, Chapter 4 (Jenny Bryan, University of British Columbia); the author’s blog posts 29-setupquarto, 07-multilanguagequartodemo, 17-rapidconversionRtoRmd, and 38-tableplacementrmarkdown; for the float-placement and troubleshooting sections, the R Markdown Cookbook (Xie, Dervieux, and Riederer, 2020).
Rationale. An analysis is finished not when the code runs but when a collaborator can regenerate the plots, the tables, and the paragraphs of the paper from source with one command, and the gap between those states is usually a number typed into a sentence. The chapter is organized around that gap: inline computation first, then the two mechanisms (freeze and cache) that can reopen it by serving stored results after the data change. The worked example runs the real quarto command on a temporary HF-HOME project and shows a project render under freeze: auto returning interim numbers after the full extract arrived, which is the chapter’s silent failure made visible; it replaces the earlier display-only simulated report. The full front-matter listing and the website section are reduced to a reference section, the chunk-option, cross-reference, and output-format lists are tables, and the three-prompt LLM section is one callout whose check is the chapter’s own: change the data, render again, and every number must change. A document is rendered once when it is written and perhaps forty times before it is submitted, and the frictions that are negligible on the first render (a float pages from its citation, a chunk left at eval: false) dominate by the fortieth. This edition therefore merged the troubleshooting half of the former ‘Migrating and Troubleshooting Documents’ chapter into this one: ‘Float placement in PDF’ (Section 19.7) and ‘Troubleshooting a render’ (Section 19.8), with the error, cause, and fix table and the executed debugging story (the real quarto command on a document with two planted defects, one that stops the render and one that ships), are now its last two sections. The debugging story reuses the worked example’s trial table rather than rebuilding it, and the old section ids were kept so that existing references resolve. Exercises are renumbered: 6 (floats) and 7 (reading a render log) are new, and the cumulative exercises are now 8 and 9. Exercises 2, 6, and 9 render on the student’s own machine, so their keys describe a correct result rather than giving output.
E.3.19 renv for Package Management
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics.
Rationale. Package drift is the most common cause of an analysis that re-runs without error and returns a different number, and renv is the community-standard remedy. The chapter teaches the three verbs through the judgments they require (when to snapshot, which direction to resolve a status() report, which packages the static scan misses) and defers everything beneath the R layer to Chapter 21. The full Dockerfile that previously appeared here now lives only in Section 21.3; the chapter keeps the four lines that belong to renv. zzrenvcheck is presented as one way to automate an audit that renv::status() and renv::dependencies() perform first. Its exercises modify a project library, so their keys describe what a correct result looks like rather than giving output.
E.3.20 Docker for Reproducibility
Sources. The author’s blog posts 32-sharermdcodeviadocker and 33-shareshinycodeviadocker; the rocker project (Boettiger and Eddelbuettel 2017); the author’s Docker and renv demonstration materials; Boettiger (2015) for the case for containers in research.
Rationale. The chapter opens with ‘same code, different number’ and makes it concrete with an executed example on the HF-HOME data: the same bootstrap with the same seed gives different interval bounds under the integer sampler of R before and after 3.6.0, a change renv cannot see because no package changed. Apple Silicon guidance stays at the first command. The canonical Dockerfile (Section 21.3), the digest pin, the .dockerignore, the localhost binding for RStudio Server, and the other earlier corrections are unchanged; the Dockerfile is byte-identical to the copy in Section 22.5. Each recipe now ends with a Verification, and the Docker ones were run against the author’s test image on 2026-10-09: R --version and the OS check, docker inspect for the digest, the CACHED restore step after editing a .qmd, docker port for a 127.0.0.1 binding, and a docker save and docker load round trip with matching image IDs. Running renv::status() inside that test image also showed the Pitfall’s failure in practice: that image’s lockfile records R 4.6.1 while the base is 4.4.0, and the image builds and renders, noting the mismatch in one log line. Docker Compose is cut to a pointer paragraph that keeps the corrected secret handling (a .env file excluded from Git and the build context; docker compose, not docker-compose), and the Compose exercise was dropped. The former ‘Orientation’ and ‘statistician’s contribution’ material is folded into the sections where each judgment arises (base-image choice and pinning into the base-image section, layer order and data placement into the Dockerfile section). Of the three LLM prompts, the Dockerfile prompt is kept with digest pinning and the .dockerignore in its criteria, and the error-diagnosis prompt is rewritten around the build log; the base-image prompt was cut. The base-image sizes in the rocker table were observed for local images (r-ver 4.4.0, tidyverse and verse 4.6.0) and are approximate.
E.3.21 Assembling the Compendium
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics. The concept is Gentleman and Temple Lang (2007); the working definition, the Five Pillars framing, the compendium layout, and the rrtools package are Marwick, Boettiger, and Mullen (2018). The optional section draws on the author’s zzcollab framework (github.com/rgt47/zzcollab; version 0.2.0 when the chapter was first written, checked against the repository as of October 2026) and the blog posts 14-penguins1zzcollab and 42-zzedcindependence. Harrell’s R Workflow and Biostatistics for Biomedical Research (Chapter 21, ‘Reproducible Research’), and the Vanderbilt Biostatistics department’s public page on reproducible reporting, are the peer treatments; the last two left Further reading when it was cut to three items, and remain good instructor reading for how these practices look as a unit’s standing policy.
Rationale. This edition merged the former ‘Research Compendia with rrtools’ chapter into this one, which keeps the section id sec-zzcollab because other chapters reference it; the rrtools section ids (Section 22.1 and its subsections) were kept for the same reason. The merged chapter is an anatomy of a compendium rather than a tour of one package. It opens with a collaborator whose rerun gives a different p-value and traces each cause (newer packages, a stale derived file, a hand exclusion, and an outcomes file copied from a shared drive that now holds a later extract) to a missing pillar: each of the preceding chapters (lockfile, container) pins one layer and leaves the wiring to the reader, and the failure that results is silent. The definition and the layout come first; the layout section builds a toy compendium in a temporary directory from two HF-HOME raw files and records a checksum manifest for the raw data, which the Pitfall then trips by saving an edit over the raw export. The Five Pillars follow as a tool-neutral checklist (a table of what each pins, the failure without it, and the check, with the R tool and its equivalents in other languages), so that zzcollab is one way to automate them rather than their definition. ‘Assembling the compendium by hand’ (Section 22.4) wires the pillars from the rrtools scaffold, renv, the canonical Dockerfile, and the manifest; the paper inside the compendium, the build, and the worked example follow, and the audit section turns the checklist into an executable function run on a temporary HF-HOME compendium (labeled as a structural check); the fresh-clone rebuild is presented as the real test. The former ‘Orientation’ and ‘statistician’s contribution’ material is folded into the sections where each judgment arises (convention into the definition, file placement and the README into the layout section, scaffolding choice into the conventions table and the section on how much to automate). The canonical Dockerfile is byte-identical to Section 21.3; the sample paper.qmd and the ‘rticles’ correction are unchanged from the earlier revision. The community tools come first, and zzcollab is optional: it appears in one clearly marked section beside rrtools and workflowr. Its walkthrough, make targets, and profiles were retained as corrected earlier, and the bundle and extension mechanics were condensed to a short subsection at the end. Of the LLM prompts, the directory-restructuring prompt is kept as the single callout, requiring a clean commit first and forbidding any change to raw data, with a rename-only diff and the checksum manifest as its check. One exercise audits the book’s own repository, so its computed key shows which pillars the book itself lacks; the exercise that needs Docker and the zzc command has a key that describes a correct result.
E.3.22 Reproducible Pipelines with targets
Sources. The targets user manual (Landau, 2021) and the paper introducing the package in the Journal of Open Source Software (Landau, 2021); Karl Broman’s reproducibility course materials (Broman, 2019) and the Software Carpentry lesson on GNU Make.
Rationale. A research compendium holds the code, the data, and the environment, but it does not by itself make the analysis re-runnable end to end; that requires a tool that knows which step depends on which. The peer survey (Appendix B) found dependency-aware pipeline tools taught in two generations, GNU Make in Broman’s reproducibility course and in Software Carpentry, and targets in more recent collaborative-practice courses. For a book that already teaches compendia and Docker, a build tool is the missing piece. The chapter opens with a Table 2 built from a stale intermediate and makes the pipeline real: a _targets.R on the HF-HOME randomization and outcomes files is built in a temporary directory during the render, with real tar_make() output before and after an edit, tar_outdated() between them, and computed keys. The edit repairs a planted derivation bug (unascertained outcomes counted as event-free), so the chapter’s silent failure is the one the tool prevents. These chunks run only when the targets package is installed (requireNamespace()); otherwise the chapter prints a note and shows the code alone. The figure’s nodes are the pipeline’s target names, tar_quarto() is shown for the report target, and the earlier corrections (when CMD runs, tar_option_set(packages = )) are kept. Exercises 5 to 7 need a Quarto report, Docker, or a second clone, so their keys describe a correct result.
E.3.23 Cloud Compute and Remote Servers
Sources. The author’s blog posts 22-serversetupawscli, 23-serversetupawsconsole, 48-ttyd_setup, and 20-researchbackupsystem; the SLURM sbatch and job-array documentation; Yoo, Jette, and Grondona (2003) for SLURM and Kurtzer, Sochat, and Bauer (2017) for Apptainer.
Rationale. Most students meet remote compute through a university cluster before they ever rent a cloud machine, so the chapter now opens with a job array killed for memory and teaches SLURM (a minimal sbatch script, a job array for a simulation study, sacct as the check) before AWS. The executed part is the R side of the job array: a per-task simulation of the HF-HOME design’s power with L’Ecuyer random-number streams, and a combine step that refuses to summarize an incomplete set; the Pitfall deletes one task’s output to show the silent version. The AWS service catalog is reduced to one glossary table, and every price, instance type, and image path lives in one dated table built from a data frame, so that the worked example’s costs are computed rather than typed. A budget alert now precedes any provisioning, and the idle-stop alarm and the trap-based job script are the two stopping mechanisms; this edition also adds security basics (SSO, instance roles, no long-lived keys). The former ‘Orientation’ and ‘statistician’s contribution’ material is folded into the sections where each judgment arises (instance choice into the shortage section, cost discipline into the cost section, data hygiene into ‘Where the data may go’, reproducibility into the container paragraph). The ‘Alternatives to AWS’ prose became a table. Of the three former LLM prompts, the failure-handling one is kept as the chapter’s single LLM callout, with the deliberate-failure test as its check; the instance-choice and error-diagnosis prompts were cut. The prices in the dated table are carried from the previous text (c6i.4xlarge, S3, and EBS) or were added during the revision from 2025 list prices (the other instance types and egress), and none was re-verified against the pricing pages in the 2026-10 revision; they should be checked against the AWS pricing pages before each offering. The SLURM and AWS commands are display-only because they need accounts; the sacct state names, the job-array filename patterns, and the budgets and cloudwatch argument shapes follow the vendors’ documentation but were neither run nor re-checked against it in this revision.
E.3.24 Statistical Analysis Plans
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics.
Rationale. The SAP is the professional biostatistician’s most important non-code document, and the chapter makes the narrow argument for it: not a defense against dishonesty, but the only dated record that separates a planned analysis from a long series of small, defensible, data-prompted choices. The opening uses the COMPare audit (Goldacre et al., 2019) as the silent failure; the former Orientation and ‘statistician’s contribution’ sections, which repeated each other, are folded into the sections where each judgment arises (the narrow argument into ‘Why pre-specify’, the specificity and HC3 judgments into ‘What a SAP contains’, tagging into its own section). A small simulation of researcher degrees of freedom replaces the bare citation of Simmons et al. (2011); it reproduces the direction of their result, not their 60% figure, because its four forks are different and correlated. Estimands enter as a section, worked with HF-HOME and consistent with section 2 of the canonical SAP, including a computed demonstration that comparing patients by the visit they received answers no estimand. The thirteen-item section list became a table mapping each section to its Gamble et al. (2017) items and ICH E9 sections. The power curve marking 750 and 1,094 per arm turns an earlier correction into a lesson about typed numbers, and a Verification checks that the running study’s randomization file matches the SAP’s computed sample size. zzpower remains as an option beside base R, with its registry list and power_table() example removed. In this edition the chapter absorbed the data management and sharing plan section from Chapter 33 (the six-element table, the HF-HOME DMSP, and the identifier Verification), now Section 25.16, placed after the worked SAP because both are written before the work begins. A downloadable SAP skeleton is planned. The former Further reading items not kept in the chapter remain useful for instructors: SPIRIT (Chan et al., 2013); CONSORT (Schulz et al., 2010) and STROBE (von Elm et al., 2007); Simmons et al. (2011); Gelman and Loken (2013); Regression Modeling Strategies (Harrell, 2015); and Biostatistics for Biomedical Research, chapter 22, on controlling the error rate of a decision rather than of a test.
E.3.25 Testing and Continuous Integration
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics. The continuous-integration sections draw on the GitHub Actions documentation, including ‘Security hardening for GitHub Actions’; the r-lib/actions repository; the book’s own workflow, .github/workflows/render-book.yml; and Humble and Farley’s Continuous Delivery (2010), which supplies the ‘fail loudly and early’ argument the chapter applies to research code.
Rationale. Analysis code is rarely tested because it appears to run once on known data, and the chapter answers that argument with the failure it misses: upstream drift that leaves every number plausible and wrong. Continuous integration is what runs those checks, with the dependency pinning of Chapter 20 and the reports of Chapter 19, on every change, so that a broken pipeline is caught within minutes rather than at submission; the peer survey (Appendix B) found the practice taught at almost none of the comparison courses, because their canonical curricula predate it. This edition merged the former ‘Continuous Integration with GitHub Actions’ chapter into this one: its four sections (what CI proves, the anatomy of a workflow, reading a failing run, and securing the workflow) follow Section 26.6, and the exit-status demonstration appears once, there. The chapter teaches tinytest for its lack of dependencies and small vocabulary, and moves the testthat comparison to a short translation table at the end; a fuller translation appendix is planned. The organizing distinction is between structural checks and tests of correctness, and the worked example puts both layers on the HF-HOME CRF export: a hand-built fixture with one row per site and an integration test against two independent records of the same patients. A planted bug (Site D’s sex codes swapped) is run against the suite in the chapter, which introduces mutation testing by name. The CI section keeps the exit-status correction made earlier (the test step’s exit status through tinytest::any_fail()) and demonstrates it by running a failing suite with and without the quit() line. The red CI log is real output: an executed chunk runs the CI test step in a fresh R process whose library holds only tinytest, a package installed locally and never snapshotted, and a table annotates the log line by line, with the Verification asserting that the first error is the cause. The former ‘statistician’s contribution’ and the three ‘what to check’ paragraphs became the CI checklist table. The other earlier corrections are kept: the Quarto setup step, setup-renv versus setup-r-dependencies (a two-row table), and the workflow’s correct name, render-book.yml. Supply-chain basics, added in this edition, form the last CI section: pinning actions by commit SHA (motivated by the March 2025 tj-actions/changed-files compromise, CVE-2025-30066), a permissions: block, OpenID Connect instead of stored cloud keys, and Dependabot for actions; the SHA shown for actions/checkout v4.4.0 was read with git ls-remote in October 2026. Two Pitfalls name real silent failures, continue-on-error: true and fixing a red restore with an install step in the workflow. The LLM callout is agentic and states the rule that an agent may not edit or weaken tests. The testing exercises on the HF-HOME files have executed keys; the CI exercises are offline with keys (predicting which step fails, a scheduled trigger, reading an executed render log, diagnosing a system-library failure, and auditing the book’s own workflow with an executed audit function, whose output currently reports every action pinned by tag and no permissions: block, a finding for the author). The cumulative exercise needs the student’s own project.
E.3.26 Clinical Data Standards: CDISC, SDTM, and ADaM
Sources. Authored directly for this book rather than adapted from earlier course materials. Target audience: students heading to pharmaceutical or CRO statistical programming and biostatistics roles who have no prior exposure to CDISC. Written in response to reports from alumni that a gap in this area was felt on the first day of their industry positions.
Rationale. The chapter builds SDTM and ADaM from the running HF-HOME trial (hfhome_sdtm() for DM, EX, and DS, and an HO domain built from the outcome file) instead of a separately fabricated study, so that its ADSL and ADTTE are the datasets Chapter 28 reuses and every count can be checked against the raw CSV files. The time-to-event analysis is presented as supportive of the SAP’s primary risk-difference analysis, runs on the ITT set from randomization, and is reported as inconclusive, consistent with the seeded draw. The CNSR convention, previously restated five times, is stated once as a Pitfall and once at the point of analysis; the HF-HOME ADTTE uses two censoring codes so that CNSR == 0 is shown to matter rather than asserted. Industry practice (TLF packages, the R submission pilots, package validation, double programming) is one section; packages not in the book’s environment (rtables, tern, Tplyr, diffdf, riskmetric) are shown display-only, and double programming is executed as an independent re-derivation of ADTTE from the raw files. The LLM callout frames an independent re-derivation by a model as the analogue of double programming. Two items deserve a second reader with regulatory or clinical-programming expertise: the status of the R Consortium pilots (dated 2026-10 in a footnote) and the choice of CNSR = 2 and last contact for subjects lost to follow-up.
E.3.27 SAS for R Programmers
Sources. A survey of the published curricula of peer US biostatistics MS programs (Appendix B, which lists every source) found that most require or strongly recommend SAS in their core or analysis-track coursework. Industry pharmaceutical, CRO, and regulatory-filing positions in the United States frequently filter screening on SAS experience, and R-only graduates from this program have reported being excluded at the screening stage.
Rationale. The chapter is the minimum credible response: enough SAS to read code and a log, move data between SAS and R, and reproduce the procedures that appear in most clinical study reports. It is organized around the defaults that differ between the two languages (Table 28.2), because those produce plausible, disagreeing numbers, and it reuses the HF-HOME ADSL and ADTTE from Chapter 27 for the survival and transport examples. The ODA registration click-path was replaced by a link and a short summary, with limits in a dated footnote. The log exercises and the default-prediction exercises need no SAS installation, so every student can attempt them, and their keys are checked in R where a check is possible. No SAS code in the chapter has been run for this edition (SAS OnDemand access was deferred by the author); every SAS block is labeled display-only, and the annotated logs are constructed, not captured. Running the SAS blocks once on ODA, and capturing a real log for the annotated example, is the outstanding verification for this chapter.
E.3.28 Case Study: Palmer Penguins
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics. The dataset and its description are from Horst, Hill, and Gorman (2022); the causal-language subsection draws on Hernán and Robins (2016) and Hernán (2018).
Rationale. The chapter is the worked model for the capstone in Chapter 34: it runs the whole arc once on data that impose no load of their own, and Table 29.1 labels every step with the chapter that taught it, so that students can reuse the table as their capstone checklist for HF-HOME. Penguins stays here (and in the early graphics chapter) as the deliberate non-clinical warm-up before HF-HOME. The former ‘Orientation’ and ‘statistician’s contribution’ material is kept and folded into the steps where each judgment arises: complete-case reasoning in the cleaning section, the reference level and the adjusted-versus-raw species difference in the model section, ‘resist over-modeling’ in the interaction subsection, and ‘treat the case study as a template’ in the arc section. This edition makes three additions. The diagnostics are now interpreted panel by panel from computed statistics instead of ‘should be nearly featureless’, and a residuals-by-sex plot shows structure the four standard plots miss, which is the chapter’s silent failure. Testing and CI steps were absent; the chapter now executes a tinytest file whose slope test re-derives the estimate from pooled within-species moments, and plants the na.omit() bug to show the suite failing. A one-page causal-language subsection (a DAG, the target trial, and the ‘C-word’ argument) was added in this edition. The literal paper.qmd listing was replaced by a description and pointers to Section 19.4 and Section 22.4, and gtsummary now carries Table 1, with zztable1 shown as one way to automate it; the community tools come first and the author’s packages are optional. The scaffold command is zzc tidyverse; zzc analysis is a deprecated alias. Former exercise 6 (deposit to Zenodo) is the chapter’s one ‘on your own work’ exercise; former exercise 4 (reproduce the compendium) is the cumulative exercise.
E.3.29 Case Study: ADNI MCI Prediction
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics. ADNI background is from Weiner et al. (2017); the prediction framing follows Grassi et al. (2019); the reporting map follows the TRIPOD+AI statement (Collins et al. 2024). The modeling details belong to the companion book, Statistical Computing in the Age of AI.
Rationale. ADNI is the book’s real-data case study and the one whose data students cannot see, so the chapter is built around what a reviewer without a data-use agreement can still check. It opens with the clinical stakes, a family asking whether MCI will progress, and makes every quiet decision behind that probability explicit. The worked example was display-only; it now executes end to end against simulate_adnimerge(), a seeded generator with the ADNIMERGE schema, and only the lines that load the restricted package remain display-only. The former ‘statistician’s contribution’ judgments are kept and placed where each arises: cohort definition in the cohort section, pre-registration in the SAP section, version pinning and asymmetric sharing in the DUA section, participant-level cross-validation in the modeling section. A Pitfall on sending participant-level rows anywhere, including to a language model, was added. gtsummary and ggplot2 are taught first, with zztable1 and zzlongplot shown beside them as optional alternatives. The pre-specified versus exploratory lists became a side-by-side table, and a TRIPOD+AI table maps reporting domains to the compendium. One statistical correction needs a second reader with biostatistical expertise: excluding participants who leave early without converting, which the chapter previously called ‘conservative’ and ‘cleanest’, biases the three-year risk upward, because early converters who leave are kept; a Kaplan-Meier estimate with censoring is close to the target when dropout is uninformative. Exercise 1 demonstrates this by simulation over 20 seeds, and the chapter now presents censoring and exclusion with that bias stated. The cohort code itself, as corrected earlier, is unchanged apart from being wrapped in build_cohort() so that the cumulative exercise can test it. The ‘requires ADNI access’ exercises were merged into a single ‘on your own work’ exercise; the former synthetic-generator exercise became the chapter’s executed stand-in. vanbuuren2018flexible was dropped from Further reading (it is cited in Chapter 17) to keep the list at three.
E.3.30 Ethics Beyond Compliance
Sources. The author’s lecture notes and supporting materials for a graduate practicum in biostatistics; the ASA Ethical Guidelines for Statistical Practice (2022); Obermeyer et al. (2019); Rothwell (2005); Hernan (2018) and Hernan and Robins (2016).
Rationale. The book’s treatment of ethics is otherwise regulatory: Chapter 33 covers what the funders require, Chapter 4 what the HIPAA Privacy Rule requires, and Chapter 27 what the FDA requires. All three are necessary and none is sufficient, because an analysis can satisfy every one and still mislead every reader it reaches. This chapter is about the failures compliance does not prevent, for which the statistician, not the IRB, is the last line of defense. The ‘where each failure is caught’ table opens the chapter as an advance organizer. It also covers the ASA guidelines, the federal definition of research misconduct, questionable research practices including HARKing, and a page of causal foundations (a DAG, target-trial emulation, and Hernan’s ‘C-word’ argument) where the chapter already uses causal language. Each of the five failures belongs to an earlier chapter (Chapter 4, Chapter 10, Chapter 16, and Chapter 32), and each of those chapters carries a short Ethics callout pointing here, so that students meet the failure where it arises and the full argument later.
E.3.31 Communicating a Finished Analysis
Sources. The author’s lecture notes and supporting materials for a graduate practicum in biostatistics; Altman and Bland (1995) on absence of evidence; Harrell’s author checklist; the CONSORT statement.
Rationale. The book teaches the reader to produce a defensible analysis, and Chapter 3 covers the register in which a statistician speaks to a clinician, but neither addresses the artifacts a working biostatistician is judged on: the talk, the memo, and the response to reviewers. A book that gives a chapter to SAS on grounds of employability cannot be silent about the three documents that constitute the job. The chapter is built around HF-HOME’s real result, which is inconclusive: the memo, the narrative arc, the talk, and the reviewer responses all read the running study’s files, and the null-results section is the center rather than an aside. The sensitivity analysis is the two-dimensional tipping point that SAP section 11 now specifies, with the same code and seed as Section 17.6, so the two chapters report the same grid. A hidden assertion stops the render if a data correction changes the pattern the memo’s words describe; this is Habit 10 extended from numbers to the words that interpret them. The ‘Collaborating with an LLM’ section is reduced to two callouts (memo drafting; reviewer responses); the adversarial self-review prompt survives as a paragraph in the peer-review section.
E.3.32 Federal Requirements and Deposition
Sources. Adapted from the author’s lecture notes and supporting materials for a graduate practicum in biostatistics; the 2022 OSTP memorandum; the NIH data management and sharing and public access policy pages.
Rationale. Chapter 2 argues that public-access compliance has become a biostatistical responsibility, because what must be shared moved from the paper to the things the statistician owns; this chapter is the operational half of that argument. It sits near the end because almost everything it asks for (compendium, lockfile, container, pipeline) exists by then, so deposition is a checklist against work already done. The DMSP is the exception, since it is written before the work begins; this edition moved it, with its id sec-federal-dmsp, to Chapter 25, as a section after the worked SAP. This chapter now opens at publication, when the plan is carried out, and points back to it. Perishable policy facts live in one dated table with a ‘Verified’ column; rows the author has not checked are left blank on purpose. The renv versus Imports: material, which repeated Chapter 20, is reduced to a pointer. The worked DMSP is HF-HOME’s, with de-identified participant-level data under controlled access in a clinical-trial repository, correcting the earlier model plan’s PHI deposit in dbGaP and TCIA. The compliance statements should be read by a second reader with regulatory or privacy expertise before the chapter is assigned.
E.3.33 Capstone: HF-HOME from Export to Deposit
Sources. Written for this edition as the second-quarter capstone; no earlier lecture source. The milestone table is modeled on Table 29.1.
Rationale. The former ‘Course Synthesis and Review’ was a recap of the table of contents. It is replaced by the second-quarter capstone: twelve milestones from crf_export.csv to a sandbox deposit, a rubric indexed to the twelve habits (0, 2, or 3 per habit, maximum 36), and a ten-week schedule aligned with the second-quarter reading. The long reading list and the ‘not covered’ list were cut; the latter duplicates Chapter 1. The seven-habit mermaid figure (fig-habits-map) was removed because its part labels no longer matched the book and the rubric supersedes it.
Teaching notes. The HF-HOME README reports the observed arm rates, so the outcome ‘seal’ before the SAP tag is an honor-system discipline; the chapter says so in a Pitfall. Students deposit to the Zenodo sandbox, not production Zenodo. With a short term, drop milestone 11 (peer review) and have students respond to the instructor’s review. Grade milestone 11 with the same rubric.
Outstanding verification. The schedule and rubric have not been used in a live course; the 0/2/3 weighting is a proposal.
E.3.34 Tooling as of October 2026 (Appendix)
Sources. As for Chapter 9, from which this edition split it: the ellmer package from Posit, the btw package documentation, and the author’s experience with coding agents.
Rationale. Package interfaces, product names, and model identifiers change on a schedule of months, while the practice in Chapter 9 (the four Cs, the failure modes, governance, and provenance) should hold for years. The appendix therefore holds the perishable half, dated and meant to be replaced: the ellmer interface with the model identifier in one variable and an ‘as of’ comment, the coding agents in common use, and the btw and MCP setup. Its main section keeps the id sec-ai-tooling (Section F.1), so that references written before the move still resolve. The btw and MCP commands are display-only because btw is not installed in the build library, and are hedged with a pointer to the package README. The ellmer exercise moved here with its section. The appendix should be rewritten against the current documentation before each offering; the chapter should need no change when it is.
E.3.35 Migrating from R Markdown (Appendix)
Sources. The author’s blog posts 17-rapidconversionRtoRmd, 38-tableplacementrmarkdown, and 07-multilanguagequartodemo; the R Markdown Cookbook (Xie, Dervieux, and Riederer, 2020) and bookdown (Xie, 2016).
Rationale. The Quarto chapter argues for the literate document; this appendix is about the work around it: converting between formats, deciding whether to migrate a project and doing so without breaking its cross-references, and running other languages beside R. A converted cross-reference that resolves to nothing is its silent failure, and the conversions run for real on a short HF-HOME script in a temporary directory. The material began as the chapter ‘Rmd Workflow: Conversions and Tables’, later retitled ‘Migrating and Troubleshooting Documents’; this edition dissolved that chapter. Its troubleshooting and float-placement sections moved into Chapter 19; the conversion, migration, and engines sections became this appendix (Appendix G), whose opening section carries the old chapter id sec-rmd-workflow. As reference material the appendix has no pipeline diagram, chapter card, or quiz. Exercises 3, 5, and 6 work on the student’s own projects, so their keys describe a correct result.