Designing a Clinical Trial: A Practical Guide

Deciding whether a study is observational or interventional, and where it sits among adaptive or decentralized variants, is the first branch of clinical study design (see Clinical Study Design: The Major Types and How They Relate). Once that structural choice is made, a second, more operational set of decisions determines whether the trial can actually answer its scientific question and stand up to regulatory and statistical scrutiny: what exactly the trial is measuring, how many participants it needs, how those participants get allocated to arms, and how the resulting data will be analyzed. This guide walks through those four practical mechanics — endpoint selection, sample-size and power calculation, randomization scheme choice, and the statistical analysis plan (SAP) — for research administrators, coordinators, and investigators involved in protocol development.

Defining primary and secondary endpoints

An endpoint (also called an outcome measure) is the specific, pre-specified variable used to assess whether an intervention worked. Endpoint selection happens early in protocol development because it drives nearly everything downstream, including the sample-size calculation and the statistical analysis plan.

Primary endpoint — the single outcome (or, less commonly, small co-primary set) the trial is powered to detect and that defines whether the trial succeeded or failed. A trial should generally have one primary endpoint; stacking multiple primary endpoints inflates the sample size needed and complicates the trial’s conclusions.
Secondary endpoints — additional outcomes measured to build a fuller picture of an intervention’s effect (supporting efficacy signals, safety, quality of life). Secondary endpoints are hypothesis-generating unless the trial’s statistical analysis plan explicitly allocates them a share of the overall error rate through a pre-specified multiplicity strategy.
Exploratory endpoints — outcomes collected for hypothesis generation or mechanistic insight, not for confirmatory claims.

Endpoints are also classified by data type, which determines which statistical test and sample-size formula apply: binary/categorical (event occurred or not, e.g. response vs. no response), continuous (a measured value, e.g. change in blood pressure), and time-to-event (how long until an event occurs, e.g. overall survival, analyzed with methods like the log-rank test that account for censoring in participants who haven’t yet had the event when the trial ends). A further useful distinction is between hard endpoints (objective, difficult to misclassify — death, hospitalization) and surrogate endpoints (a biomarker or intermediate measure used as a substitute for a clinical outcome that would take longer or more participants to observe directly, e.g. tumor shrinkage as a surrogate for survival) — a surrogate is only useful to the extent it has been validated as a genuine predictor of the clinical outcome it stands in for, not merely correlated with it.

A composite endpoint combines several individual events into one outcome (e.g. “cardiovascular death, myocardial infarction, or stroke, whichever occurs first”) to increase the total event count and improve statistical power, at the cost of interpretability if the components don’t move together or don’t carry equal clinical importance.

Sample-size and power calculation basics

A sample-size calculation answers one question: how many participants does the trial need to have a reasonable chance of detecting a real treatment effect, if one exists, while keeping the risk of a false-positive conclusion acceptably low? ICH E9, “Statistical Principles for Clinical Trials,” is the foundational international guideline covering this and the broader statistical substance of trial design, alongside its estimands-focused addendum ICH E9(R1). Every standard sample-size calculation rests on four ingredients:

Significance level (alpha) — the accepted risk of a Type I error (concluding the intervention works when it doesn’t), conventionally set at 0.05 (two-sided) for a confirmatory trial.
Power (1 minus beta) — the accepted risk of a Type II error (failing to detect a real effect), with power conventionally set at 80% or 90%. An underpowered trial risks a false-negative result even when the intervention genuinely works — a wasted trial, and an ethically weighty one, since participants were exposed to trial procedures for a study unlikely to yield a conclusive answer.
Effect size — the minimum difference between groups the trial is designed to detect as clinically meaningful, typically estimated from pilot data, prior published trials, or a minimal clinically important difference (MCID) established in the literature.
Variability — the expected spread of the outcome measure (standard deviation for a continuous endpoint, event rate for a binary or time-to-event endpoint), also usually estimated from prior data.

The specific formula used depends on the primary endpoint’s data type: comparing two means uses a formula built around the standardized effect size and standard deviation; comparing two proportions uses a formula built around the expected event rates in each arm; time-to-event endpoints are powered around the number of events needed (via the log-rank test and related approaches such as the Schoenfeld formula) rather than the number of participants directly, since what drives statistical information in a survival analysis is events observed, not enrollment alone. Non-inferiority and equivalence trials add a further parameter — a pre-specified margin defining how much worse the new intervention is allowed to be while still being considered acceptably comparable to the control — which materially changes both the calculation and the sample size required relative to a standard superiority design.

Because every input is an estimate, sample-size calculations are typically reported with the specific assumptions stated (the effect size and variability used, and their source), and often supplemented with a sensitivity analysis showing how the required sample size shifts under more conservative assumptions. Overestimating the effect size, a common failure mode, produces a trial that is enrolled to a size too small to detect the real, more modest effect once it’s run.

Choosing a randomization scheme

Randomization is what allows a randomized controlled trial (RCT) to support a causal claim: random allocation balances both known and unknown confounding factors between arms before the intervention is applied, addressing selection bias specifically (a separate concern from blinding, which addresses performance and detection bias after allocation). How that random allocation is actually generated is itself a design choice with several standard schemes:

Simple randomization — each participant’s allocation is generated independently (e.g. a coin flip or random-number equivalent) with no constraint on the running balance between arms. Simple randomization is straightforward and fully unpredictable, but in small trials it can produce a meaningfully unbalanced split between arms by chance alone.
Block randomization — participants are randomized in fixed-size blocks (e.g. blocks of 4 or 6) constructed so that each block contains a balanced number of allocations to each arm, keeping the overall arm sizes closely balanced throughout enrollment rather than only at the end. A known trade-off: if block size is fixed and knowable, it can become predictable toward the end of a block, which is why many protocols use randomly varying block sizes to preserve allocation concealment.
Stratified randomization — randomization (commonly using a block scheme within each stratum) is performed separately within subgroups defined by one or more baseline factors expected to influence outcome or response (e.g. disease severity, study site, age group), ensuring those factors are balanced across arms rather than left to chance. Stratification is most useful when a factor is a strong, known predictor of outcome and the trial is not large enough for simple randomization to balance it reliably on its own; stratifying on too many factors relative to sample size fragments the strata and can undermine the balance it’s meant to achieve.

Allocation ratios don’t have to be 1:1 — unequal ratios (e.g. 2:1 favoring the investigational arm) are sometimes used to improve recruitment or gather more safety data on the new intervention, at a modest statistical-power cost relative to equal allocation for a fixed total sample size. The randomization scheme, including any stratification factors and the method used to generate and conceal the allocation sequence, is specified in the protocol and detailed further in the statistical analysis plan.

The role of the statistical analysis plan (SAP)

The statistical analysis plan is the document that translates the protocol’s endpoints and design into the exact statistical methods that will be used to analyze the trial’s data, written and finalized in detail before database lock and, for blinded trials, before unblinding. ICH E9 establishes the general expectation that a trial’s statistical methodology be pre-specified rather than determined after seeing the data; ICH E9(R1) formalizes this further through the estimand framework, which requires explicitly defining, for each key question, the population, the variable, how post-baseline “intercurrent events” (like treatment discontinuation) are handled, and the summary measure used, before analysis begins. A typical SAP covers:

Analysis populations — which participants are included in which analysis, most commonly the intent-to-treat (ITT) or modified intent-to-treat (mITT) population versus the per-protocol (PP) population; see Intent-to-Treat vs. Per-Protocol Analysis for how these differ and why ICH E9 recommends the full analysis set as the primary analysis for superiority trials.
Statistical methods for each endpoint — the specific test or model for the primary and each secondary endpoint, consistent with the endpoint’s data type (e.g. a t-test or mixed model for a continuous endpoint, a chi-square or logistic-regression approach for a binary endpoint, a log-rank test or Cox model for time-to-event).
Handling of missing data — the pre-specified approach (e.g. multiple imputation, a specific sensitivity analysis) for participants with incomplete follow-up, decided in advance rather than chosen after seeing which approach produces a more favorable result.
Multiplicity strategy — how the overall Type I error rate is controlled when testing multiple secondary endpoints or performing interim looks (e.g. hierarchical/gatekeeping testing procedures, alpha-spending), preventing the false-positive inflation that comes from repeated or multiple testing.
Interim analysis rules, where applicable — the timing, statistical stopping boundaries, and decision rules for any planned interim look, including who reviews unblinded interim results (typically a Data Safety Monitoring Board; see Data Safety Monitoring Board (DSMB)) and under what conditions the trial could stop early for efficacy, futility, or safety.

Finalizing the SAP before unblinding is the mechanism that keeps a trial’s analysis confirmatory rather than exploratory after the fact: it forecloses the possibility of choosing, post hoc, whichever analytical approach produces the most favorable result for a given endpoint. Trials that use an adaptive design depend even more heavily on this discipline, since the SAP must also pre-specify exactly which design elements can change mid-trial and under what statistically controlled rules.

How these decisions fit together

Endpoint selection, sample-size calculation, randomization scheme, and the SAP are not four independent decisions made in sequence — they are interdependent, and changing one usually forces revisiting the others. The primary endpoint’s data type determines the sample-size formula; the sample size achievable given feasible recruitment can, in turn, force a change in endpoint (e.g. moving from a hard clinical endpoint to a validated surrogate, or from superiority to non-inferiority) or in randomization ratio; and every one of these choices must be locked into the protocol and SAP before enrollment begins, consistent with ICH E8(R1)’s general framework for what constitutes a well-designed trial and ICH E9’s expectation of statistical pre-specification. See Clinical Study Design: The Major Types and How They Relate for how these practical mechanics fit within the broader observational-vs-interventional and adaptive/decentralized taxonomy, and Clinical Trial Phases for how endpoint and sample-size expectations shift across Phase 1 through Phase 4.

Frequently asked questions

What is the difference between a primary and secondary endpoint?

The primary endpoint is the single outcome the trial is statistically powered to detect and that determines whether the trial met its objective. Secondary endpoints are additional outcomes measured to build a fuller picture of the intervention’s effect; they are generally hypothesis-generating unless the statistical analysis plan explicitly allocates them a controlled share of the trial’s overall error rate.

How do you calculate sample size for a clinical trial?

A sample-size calculation combines four inputs — the significance level (alpha, conventionally 0.05), the desired statistical power (conventionally 80–90%), the minimum clinically meaningful effect size, and the expected variability of the outcome — using a formula specific to the primary endpoint’s data type (continuous, binary, or time-to-event). All four inputs are estimates, usually drawn from pilot data or prior published trials, which is why sample-size assumptions are typically reported alongside the resulting number.

What’s the difference between simple, block, and stratified randomization?

Simple randomization allocates each participant independently with no balancing constraint. Block randomization allocates participants in fixed-size balanced blocks to keep arm sizes closely matched throughout enrollment. Stratified randomization performs randomization separately within subgroups defined by a baseline factor expected to affect outcome (such as site or disease severity), ensuring that factor is balanced across arms rather than left to chance.

What is a statistical analysis plan (SAP) and why does it matter?

The SAP is the document that specifies, in detail and before database lock or unblinding, the exact statistical methods that will be used to analyze the trial’s endpoints — including analysis populations, the specific test or model per endpoint, how missing data is handled, and the multiplicity strategy for secondary endpoints. Finalizing it before the data is unblinded is what keeps the trial’s primary analysis confirmatory rather than an after-the-fact, results-driven choice among methods.

Does a bigger sample size always mean a better trial?

No. An underpowered trial risks failing to detect a real effect, but an unnecessarily oversized trial exposes more participants to trial procedures and risk than the scientific question requires, and consumes resources and time that could otherwise support a second trial or an earlier answer. The goal of a sample-size calculation is the smallest number of participants that gives an adequate, pre-specified chance of a conclusive answer — not the largest number affordable.