Clinical Trial Design & Analysis

Randomization and allocation concealment

Why randomize. Random allocation gives each participant a known chance of each assignment, independent of any characteristic. On average this spreads both known and unknown confounders evenly between arms, so a difference in outcomes can be attributed to the intervention. It also removes investigator choice about who gets what, preventing selection bias. Readers verify that randomization worked by checking the baseline characteristics table (usually Table 1): the arms should look alike. Randomization protects internal validity; strict eligibility criteria can still limit external validity (generalizability).

Methods. Simple randomization, the equivalent of a coin flip for each patient, works well in large trials but can leave small trials unbalanced in size or key traits. Block randomization assigns within small blocks (for example, blocks of 4 containing 2 of each arm), keeping arm sizes equal throughout enrollment; if the block size is known, the last assignments become predictable. Stratified randomization runs a separate randomization within strata of a strong prognostic factor, such as disease stage or study site, guaranteeing balance on that factor; too many strata become unworkable.

Allocation concealment. Concealment keeps the upcoming assignment hidden from whoever enrolls patients, using central web or telephone randomization or sequentially numbered, sealed, opaque envelopes. If recruiters can foresee the next assignment, they can steer sicker or healthier patients into a particular arm, reintroducing selection bias despite a perfectly random sequence. Concealment acts before allocation and is always achievable, whereas blinding acts after allocation and is not always possible.

Blinding

Levels. In a single-blind trial participants do not know their assignment; in a double-blind trial neither participants nor the clinicians and investigators do; in a triple-blind trial the outcome adjudicators or data analysts are also unaware.

What it prevents. Blinding limits placebo effects, unequal co-interventions or care between arms (performance bias), and biased outcome measurement (observer or detection bias). These problems matter most for subjective outcomes such as pain or quality of life.

When blinding is impossible. Participants and surgeons cannot be blinded when surgery is compared with medication or when an exercise program is compared with usual care. The remedies are blinded outcome assessors, objective endpoints such as death, and, where ethical, sham procedures or a double-dummy design in which each arm receives one active product and one placebo that mimics the other.

Analysis populations

Intention-to-treat (ITT). Every participant is analyzed in the arm to which they were randomized, regardless of adherence, crossover, or withdrawal. ITT preserves the balance created by randomization, estimates real-world effectiveness, and is the standard primary analysis for superiority trials.

Per-protocol and as-treated. Per-protocol analysis includes only participants who completed their assigned treatment as specified; as-treated analysis groups patients by the treatment they actually received. Both break randomization, because adherent patients differ systematically from nonadherent ones. In trials, even patients who faithfully take placebo have lower mortality than those who do not (the healthy-adherer effect). These analyses therefore estimate efficacy under ideal conditions and can overstate benefit.

Bias toward the null. When control patients cross over to the active therapy, or treated patients stop it or drop out, the arms become more alike and the observed difference shrinks. ITT then underestimates the biological effect. This is conservative in a superiority trial but anti-conservative in a noninferiority trial.

Superiority, noninferiority, and equivalence

Superiority. The null hypothesis is no difference. Superiority is shown when p is below α (usually 0.05) and the confidence interval excludes the no-effect value, 0 for a difference or 1 for a ratio such as a hazard ratio. A nonsignificant superiority trial does not prove two treatments equivalent; showing similarity requires a trial designed around a margin.

Noninferiority. Noninferiority trials ask whether a new therapy that is cheaper, safer, or easier is not unacceptably worse than standard care. A margin (Δ), the largest loss of efficacy considered acceptable, is prespecified. Noninferiority is established only if the confidence interval limit on the unfavorable side stays within Δ. Because nonadherence and crossover pull ITT toward the null and make arms look alike, both ITT and per-protocol results are required, and they should agree.

Equivalence. Equivalence requires the entire confidence interval to lie between −Δ and +Δ, showing the new treatment is neither meaningfully better nor worse. It is used for biosimilars and generic drugs; generic bioequivalence requires the 90% CI for the test-to-reference ratio of AUC and Cmax to fall within 80–125%.

Trial designs

• Parallel: each arm receives one intervention throughout; the most common design.
• Crossover: each participant receives both treatments in sequence, separated by a washout, and serves as their own control; suited to stable chronic conditions and short-acting treatments, but vulnerable to carryover effects.
• Factorial (2×2): participants are randomized to A vs no A and, independently, to B vs no B, creating four groups (A+B, A only, B only, neither). Two questions are answered for about the price of one, provided the interventions do not interact; the Physicians' Health Study tested aspirin and β-carotene this way.
• Cluster-randomized: whole units such as clinics, schools, villages, or ICUs are randomized when the intervention works at group level or when individuals sharing a setting would share it (contamination). Outcomes within a cluster are correlated (intracluster correlation), so more participants are needed and the analysis must account for clustering.
• Adaptive: accumulating data modify the trial according to rules written in advance, such as dropping futile arms, re-estimating sample size, shifting randomization toward better-performing arms, or adding arms in platform trials such as RECOVERY for COVID-19.

Endpoints, multiplicity, and monitoring

Primary and secondary endpoints. The primary endpoint is fixed before enrollment and determines the sample size needed for adequate power. Secondary endpoints and post hoc findings are hypothesis-generating rather than confirmatory. Switching outcomes after seeing the data inflates false-positive findings; prospective registration on sites such as ClinicalTrials.gov exposes outcome switching and unpublished negative trials.

Surrogate endpoints. Biomarkers such as HbA1c, LDL, blood pressure, and tumor shrinkage are faster and cheaper to measure than clinical outcomes (death, MI, stroke) but may not translate into fewer events. In the Cardiac Arrhythmia Suppression Trial (CAST), encainide and flecainide suppressed ventricular premature beats after MI yet increased mortality.

Composite endpoints. Combining outcomes (for example death, MI, stroke, or revascularization) raises event rates and power, but the overall result is often driven by the most frequent and least serious component. Each component should be examined, and composites whose components differ greatly in importance deserve skepticism.

Multiple comparisons. Every additional test at α = 0.05 adds false-positive risk. The Bonferroni correction divides α by the number of tests, so 10 subgroup comparisons require p < 0.005. Credible subgroup claims are prespecified, few in number, supported by a formal test of interaction, and treated as hypothesis-generating.

Interim analyses and the DSMB. An independent data safety monitoring board reviews unblinded interim results at prespecified points and can recommend stopping for clear benefit, harm, or futility while investigators remain blinded. Stopping boundaries such as O'Brien-Fleming demand very small p-values at early looks, preserving the overall α of 0.05. Trials stopped early for benefit tend to overestimate the true effect.

Loss to follow-up. Participants lost to follow-up rarely disappear at random; they may have had adverse outcomes. Attrition bias grows with the fraction lost: under 5% causes little concern, while more than 20% seriously threatens validity, and unequal losses between arms are especially worrying. Worst-case sensitivity analyses test whether conclusions survive plausible assumptions about the missing patients.