Clinical Trial Sample Size: Power Analysis Theory and Practice
Determining the correct sample size is one of the most consequential decisions in clinical research design. An underpowered study wastes resources, exposes participants to interventions without a reasonable chance of producing actionable knowledge, and may provide misleadingly negative results. An overpowered study exposes more participants than necessary to experimental risks. Regulatory agencies — including the U.S. Food and Drug Administration (FDA) and the European Medicines Agency (EMA) — require prospective sample size justification in all pivotal trial protocols.
The Foundations of Power Analysis
A sample size calculation requires five inputs: the primary outcome measure and its expected distribution, the effect size (the difference you want to detect), the significance level (α), the desired power (1−β), and the allocation ratio between arms. Each of these is a deliberate research decision, not a statistical artifact.
Effect size is the most important and most frequently misspecified input. It should represent the minimum clinically important difference (MCID) — the smallest effect that would change clinical practice or patient management — not the largest effect that might plausibly occur. Inflating the assumed effect size to achieve a smaller sample size leads to a trial that is underpowered for the true treatment effect, even if statistically significant findings were observed in the pilot work.
Significance level controls the Type I error rate. The conventional α=0.05 (two-sided) means you accept a 5% probability of concluding an ineffective treatment works. In trials with multiple comparisons (multiple primary outcomes, multiple interim analyses, or multiple arms), the familywise error rate must be controlled — through Bonferroni correction, hierarchical testing, or group sequential boundaries — to maintain the overall Type I error at α.
Power controls the Type II error rate. The 80% convention (introduced by Jacob Cohen) remains standard, though 90% is increasingly expected for confirmatory Phase III trials by regulatory bodies, particularly for trials where the treatment addresses a serious condition and missing a real effect would be costly.
Two Proportions: The Most Common RCT Design
In a superiority randomised controlled trial with a binary primary endpoint — responder rate, mortality, complication rate — the sample size formula using the normal approximation is:
n = (z_α/2 + z_β)² × [p1(1−p1) + p2(1−p2)] / (p1−p2)²
Where p1 and p2 are the expected proportions in the control and treatment groups, z_α/2 is the critical value for the desired α (1.96 for α=0.05, two-sided), and z_β is the critical value for the desired power (0.842 for 80% power). The formula calculates n per arm for a 1:1 allocation; unequal allocation requires adjustment.
Cohen's h (the arcsine transformation effect size for proportions) provides a standardised measure: h = 2 × arcsin(√p1) − 2 × arcsin(√p2). This transformation stabilises the variance across different baseline proportions, making h more directly comparable across studies than the raw difference p1−p2.
Two Means: Continuous Primary Outcomes
For trials where the primary outcome is continuous (blood pressure, HbA1c, pain score, cognitive test score), the standard formula for equal-sized groups is:
n per group = 2(z_α/2 + z_β)² × σ² / (μ1−μ2)²
This simplifies to n = 2(z_α/2 + z_β)² / d², where d is Cohen's d (the standardised effect size: mean difference divided by the pooled standard deviation). The critical challenge in applying this formula is obtaining a reliable estimate of σ. Sources include: prior studies in the same population, systematic reviews and meta-analyses, pilot data from your own institution, or the literature's reported standard deviations for the same outcome measure. Using an underestimate of σ leads to an underpowered study; using an overestimate leads to a larger-than-necessary sample.
Survival Analysis: Number of Events, Not Participants
For time-to-event outcomes (overall survival, progression-free survival, time to readmission), sample size determination differs fundamentally from other designs: what matters is the number of events (deaths, progressions, etc.), not the number of participants enrolled. The required number of events for a log-rank test is:
E = 4(z_α/2 + z_β)² / [ln(HR)]²
Where HR is the target hazard ratio (the ratio of median survival times, approximately). Converting from events to participants requires assumptions about event rates during accrual and follow-up — this is why our survival mode asks for accrual and follow-up periods. Shorter follow-up relative to accrual means lower event rates during accrual, requiring more participants to achieve the target events.
Common Mistakes in Sample Size Calculation
Using observed pilot data as the assumed effect size. Pilot studies are inherently imprecise; observed differences in a 30-person pilot can vary widely around the true population difference. Using the pilot's observed difference as the target in a power calculation leads to optimistic (too-small) sample sizes and underpowered trials. Instead, use the MCID and treat the pilot's SD as the estimate of variability.
Failing to account for multiple comparisons. If a trial has three co-primary endpoints tested at α=0.05 each, the overall false-positive rate exceeds 5%. Each additional endpoint requires Bonferroni or other correction. The corrected α for each test then feeds into a higher required sample size.
Ignoring correlation in crossover designs. Crossover trials, where each participant receives both treatments, benefit from within-subject correlation — subjects serve as their own controls. The sample size formula for crossover trials includes a (1−ρ) term where ρ is the intra-subject correlation coefficient. Not accounting for this leads to gross overestimation of required participants.
Underestimating dropout. In oncology trials and long-duration trials, dropout rates of 20–30% are common. Failing to inflate the sample size for expected dropout leads to an evaluable population that is smaller than the powered estimate.
Regulatory Expectations: FDA and EMA
Both FDA and EMA guidelines for clinical trial design require sample size justification in the statistical analysis plan (SAP). FDA guidance (FDA E9, E10, and E9(R1) addendum) emphasises the importance of pre-specifying the estimand — the precise treatment effect being estimated, accounting for intercurrent events (treatment discontinuation, rescue medication use) — and linking the sample size to the power for detecting that estimand.
EMA's guideline on the investigation of subgroups in confirmatory clinical trials notes that subgroup analyses must be powered separately if they are to be confirmatory; otherwise, they are exploratory and hypothesis-generating only. This has direct implications for sample size when the label claim is intended for a specific population subset.
For adaptive trials, both agencies require that the adaptation rule be pre-specified in the protocol, that appropriate type I error control be demonstrated through simulation or analytical methods, and that an independent data monitoring committee (IDMC) oversee interim analyses to preserve the trial's integrity and the blind of the sponsor's analysis team.