Preface

This volume is the clinical-trials course of the rgtlab graduate curriculum. It exists because the two clinical-trial chapters of the companion volume Applied Statistical Methods for Health Sciences Research are a survey, and a survey is not enough for a student who will soon be sitting on a study team. Trial work is unusually consequential: the decisions are made before the data exist, they are written into a regulated document, and they cannot be revised once enrollment begins.

Who this book is for

The intended reader is a first- or second-year graduate student in biostatistics, epidemiology, or public health, or a medical student with an interest in research. That audience is deliberately mixed, and it shapes the book in three ways.

First, every method is motivated by a decision someone has to make, not by the mathematics that describes it. A reader who will never derive a group-sequential boundary still has to understand what a data monitoring committee is deciding when it looks at one, and a reader who will derive it needs the same motivation before the algebra is worth anything.

Second, the mathematics is kept at the level of a first graduate methods sequence. The book uses normal-theory sample-size formulas, likelihood ratios, the log-rank statistic, and linear and generalized linear models. It does not use measure theory, martingale limit arguments, or asymptotic theory beyond a stated result. Where a derivation would be instructive but long, the book states the result, shows what it implies, and points to the reference that proves it.

Third, notation is defined where it is used, and the Conventions page collects the recurring symbols in one place. A medical student who has had one statistics course should be able to read every chapter; a biostatistics student should find enough technical content in the advanced chapters to be worth their time.

The advanced-topic chapters (11 through 28) are harder than the core ten. They are written so that the first half of each is accessible to the whole audience, with the simulation detail concentrated in the second half.

What this book covers

Ten chapters, one per week of a quarter:

  1. The trial as an experiment. Why randomized experiments answer causal questions that observational designs answer only under assumptions. The historical trials that established the method, the ethical framework (Declaration of Helsinki, Belmont, Common Rule), the regulatory architecture (FDA, EMA, ICH), and the phase structure of drug development.
  2. The protocol, the estimand, and the endpoint. The protocol as the trial’s contract. Endpoint selection, surrogate endpoints, composites. The ICH E9(R1) estimand framework and the five intercurrent-event strategies. The statistical analysis plan.
  3. Randomization. Simple, permuted-block, stratified, urn, minimization, and response-adaptive allocation. Allocation concealment as distinct from blinding. Randomization-based inference. What each procedure obliges the analysis to do.
  4. Blinding, bias, and conduct. The taxonomy of bias in trials. Blinding levels and their failure modes. Site management, protocol deviations, data quality, database lock, and the operational machinery that protects the design.
  5. Sample size and power. Continuous, binary, and time-to-event endpoints. Non-inferiority and equivalence margins. Cluster randomization and the design effect. Simulation-based sizing. Sensitivity of the calculation to each input.
  6. Phase I dose-finding. The 3+3 design and why it persists despite poor operating characteristics. Continual reassessment, escalation with overdose control, BOIN and other model-assisted designs. Simulation as the required evidence for a dose-finding design.
  7. Phase II screening designs. Single-arm designs with historical controls and their hazards. Simon’s two-stage design. Randomized phase II selection designs. Bayesian predictive-probability monitoring. Phase II/III seamless designs.
  8. Phase III analysis. The intention-to-treat principle and its relationship to the estimand. Covariate adjustment and stratified analysis. Repeated measures, time-to-event, and count endpoints. Multiplicity across endpoints, arms, and looks. Subgroups and interaction tests.
  9. Interim monitoring. Data monitoring committees, group-sequential boundaries and alpha spending, conditional power and futility, sample-size re-estimation, and the broader family of adaptive designs including platform trials.
  10. Missing data, sensitivity, and reporting. Missing data mechanisms in the estimand framework, MMRM, multiple imputation, reference-based imputation and tipping-point analysis, CONSORT reporting, registration, and the regulatory submission.

What this book does not cover

  • The R language and workflow foundation (see R for Biostatistics: A One-Week Boot Camp and Biostatistics Practicum).
  • Model classes as such: linear models, GLM, mixed models, survival, Bayesian computation (see Statistical Computing in the Age of AI).
  • Observational causal inference, epidemiologic design, mediation, and meta-analysis (see Applied Statistical Methods for Health Sciences Research).
  • Trial operations at the level of a clinical research associate: monitoring visits, site contracts, drug supply. The book covers only what the statistician must understand to protect the design.
  • Device and diagnostic trials, which follow a distinct regulatory pathway, and vaccine trials, which merit their own treatment.

Prerequisites

One graduate course in statistical methods (estimation, hypothesis testing, confidence intervals, linear regression) and basic familiarity with R. Survival analysis and mixed models are introduced where they are needed rather than assumed. The book uses pwr, gsDesign, rpact, randomizeR, BOIN, clinfun, mmrm, mice, and survival; code blocks are short enough to read even by someone who does not write R routinely. No chapter assumes prior exposure to trials.

Chapter template

Each content chapter follows the sequence-wide structure: Learning objectives, Orientation, The statistician’s contribution, content sections with collapsible Check-your-understanding callouts, a Worked example, Collaborating with an LLM, Principle in use, Exercises, and Further reading. The template puts human judgment and verification at the center of every chapter rather than treating them as afterthoughts.

A note on tone

The book is opinionated where the literature is settled and explicit about disagreement where it is not. The 3+3 design is poor and remains in use; single-arm phase II trials with historical controls mislead more often than they inform; covariate-adjusted analyses are under-used in confirmatory trials. Each of these claims is argued, not asserted, and each is followed by what a statistician should actually do when the study team wants otherwise.