Randomization, Masking, and Control Groups

“Randomized, double-blind, controlled trial” compresses several design features into one phrase. Each feature addresses a different source of bias or a different question. A study can randomize well but mask poorly, conceal allocation well but lose participants after assignment, or use a control group that cannot support the conclusion in its headline.

Reading the components separately makes the design interpretable.

What randomization is designed to do

Randomization assigns experimental units, such as participants, clusters, sites, lesions, or periods, to comparison conditions through an unpredictable chance process. A valid random sequence prevents assignments from following a deterministic pattern; allocation concealment separately prevents foreknowledge of the next assignment from influencing enrollment. Together, when properly implemented, they tend to balance known and unknown prognostic factors in expectation and provide a probability basis for inference.[1][2][3][7]

Randomization does not guarantee exact baseline balance. Baseline characteristics should be described by their magnitude and prognostic importance, not tested to decide whether randomization “worked.” Under an uncompromised random process, observed imbalances arise by chance, although unusual patterns can still prompt checks for data or implementation problems. Covariate adjustment should follow the prespecified analysis.[1][9]

The word “random” is not enough. Alternation, date of birth, admission date, or record number are predictable allocation methods, not random sequence generation. Assignment by investigator judgment is nonrandom rather than a chance process. CONSORT 2025 asks authors to name who generated the sequence and the method used.[2][5]

Sequence generation and allocation concealment are separate

Sequence generation creates the unpredictable order of assignments. Allocation concealment prevents people enrolling participants from knowing the next assignment before enrollment is irrevocable.

CONSORT emphasizes that both are needed.[2][3] A computer-generated sequence can still be subverted if the next assignment is visible. Conversely, opaque envelopes cannot turn a predictable alternating sequence into randomization.

Allocation concealment protects the trial before and until assignment. Common mechanisms include a secure central system or properly implemented sequential containers. The report should describe the mechanism, not merely claim that allocation was concealed.[3]

If recruiters can foresee the next group, they may consciously or unconsciously delay, accelerate, include, or exclude a participant. That can create systematic group differences before the intervention begins.

Masking acts after assignment

Masking, also called blinding, withholds assigned-intervention information from people whose behavior or judgments could be influenced after assignment. These roles may include:

• participants;
• care providers or intervention staff;
• investigators or data collectors;
• outcome assessors, including endpoint adjudicators[10]; and
• data analysts.

CONSORT recommends naming the roles that were masked instead of relying on “single blind” or “double blind,” because those labels have been used for different combinations.[4] ClinicalTrials.gov’s protocol definitions likewise record masking by role, including participant, care provider, investigator, and outcomes assessor.[5]

A report should state not only who was masked, but how masking was achieved and maintained, how similar the interventions appeared, who could access treatment codes, and whether distinctive effects or emergency unmasking revealed assignments.[10]

Masking can limit knowledge-driven differences in participant behavior, care, co-interventions, outcome assessment, and analytic decisions. Its importance depends partly on the outcome. Knowledge of assignment may have more room to influence a subjective rating than an objectively verified death, although even objective outcomes can be affected by follow-up and analysis decisions.[4][10]

Masking is not always feasible. A surgical procedure, device, behavioral intervention, or distinctive effect may reveal assignment. An open-label trial is not automatically invalid, but it needs other protections: prespecified outcomes, standardized follow-up, masked outcome assessment where possible, and transparent discussion of the remaining risk of bias. Masked outcome assessment can reduce ascertainment bias, but it does not remove differences in behavior, co-interventions, adherence, or dropout caused by participants or care providers knowing the assignment.[1][4][8][10]

Concealment is not masking

The timing distinction is the simplest memory aid:

A paper can report that participants were masked while saying nothing about whether recruiters could see the next assignment. That leaves allocation concealment unresolved.

The control group defines the contrast

Control descriptions have two separate dimensions: what the comparator group receives, such as placebo or sham, no assigned study treatment, another dose or regimen, or an active treatment, and how the comparator group was formed, such as randomized concurrent, nonrandom concurrent, or external or historical selection.[6][11] “Concurrent” does not itself mean randomized, and a control-group label does not guarantee a valid counterfactual.

A comparator defines the intended treatment contrast. Its credibility depends on assignment, baseline comparability, common follow-up and co-interventions, outcome ascertainment, missing-data handling, and analysis. Randomized concurrent controls generally require fewer comparability assumptions than nonrandom concurrent or external controls.[6]

Placebo and sham controls

A placebo or sham is intended to resemble the test strategy while omitting the component hypothesized to be active. Its composition or procedure must be reported; the label does not establish that it is physiologically inert or risk-free.[6][10][11]

In a randomized comparison where masking is adequately achieved, the between-group estimate contrasts assignment to the test intervention with assignment to the placebo or sham under the shared care and follow-up conditions. It does not separately quantify expectations, natural history, regression to the mean, or study participation, and it does not compare the intervention with an effective alternative unless an active-control group is included.[6]

No-treatment, usual-care, and add-on comparisons

A no-study-treatment group receives no assigned study intervention but may still receive background or supportive care. “Usual care” is not a standardized ICH E10 control type: it may include active treatment, heterogeneous clinician-directed care, or no additional study intervention, and it can vary across sites, patients, and time. Its classification here is an editorial inference from the treatment actually delivered, not an ICH E10 label; reports should describe what each group actually received and what differed between them.[6][11]

In an add-on design, both groups receive common background therapy while one receives the test intervention and the other receives placebo or no added study treatment.[6][11]

Active control

An active control means that the comparator group receives another treatment; the label alone does not establish that the comparator is effective in the present population, outcome, dose, or setting. A superiority trial estimates the relative difference between the assigned strategies.[6]

For active-control non-inferiority to support efficacy, the margin must be prespecified and justified from clinical and historical evidence, the active control must be expected to retain its established effect under the current trial conditions, and trial conduct must preserve assay sensitivity. The relevant confidence bound must exclude a loss as large as the margin. Equivalence requires the entire two-sided confidence interval to lie within both prespecified margins. Failure to reject superiority is not evidence of non-inferiority or equivalence.[1][6]

Dose-comparison control

Comparing doses can characterize relative effects, but without placebo or another benchmark it may be difficult to determine whether any tested dose had an effect. The inference depends on the shape of the dose-response relationship and the chosen range.[1][6]

Historical or external control

External controls may be historical or contemporaneous but outside the trial. Because treatment assignment is not randomized, measured and unmeasured differences can bias the comparison, and statistical adjustment cannot guarantee their removal. FDA’s ICH E10 guidance says such designs are generally most persuasive only in unusual settings with a highly predictable disease course, objective endpoints, well-characterized prognostic factors, closely aligned observation methods, and an outcome markedly different from the external control. Even then, erroneous conclusions remain possible.[6]

A simple treated-group change from baseline has no concurrent internal control and ordinarily cannot isolate treatment effects from natural history, regression to the mean, or time trends. Large numerical differences or small p values do not erase these design limitations.[6]

Trial framework matters as much as the comparator

CONSORT asks authors to state whether a trial is designed for superiority, non-inferiority, equivalence, or another framework, along with its structure and allocation ratio.[7] Those labels describe different hypotheses:

• Superiority: whether the prespecified analysis supports a favorable treatment effect relative to the comparator.
• Non-inferiority: whether the relevant confidence bound excludes inferiority as large as or larger than the prespecified margin.
• Equivalence: whether the entire two-sided confidence interval lies within both prespecified margins.

The margin and analysis cannot be invented after the results are known. A point estimate inside a margin, or a non-significant superiority result, is insufficient to establish non-inferiority or equivalence.[1][6]

Randomization does not solve everything

Proper sequence generation and allocation concealment protect assignment at baseline; an analysis consistent with the randomized design and stated estimand is needed to preserve that inferential basis. Randomization does not automatically prevent:

• differential dropout after assignment;
• missing outcome data;
• unmasking;
• selective outcome reporting;
• measurement error;
• deviations from the assigned interventions;
• inappropriate statistical models; or
• poor generalizability.

FDA’s current ICH E6(R3) Good Clinical Practice guidance emphasizes quality by design, participant protection, and reliable results throughout the trial lifecycle.[8] A design label should therefore lead to a check of conduct and reporting, not end the appraisal.

A design-reading checklist

1. Identify the unit randomized: person, site, cluster, lesion, or time period.
2. Find the exact sequence-generation method and allocation ratio.
3. Determine how the next assignment was concealed from recruiters.
4. List who was masked and how masking was maintained.
5. Note who remained aware of assignment and which outcomes they could influence.
6. Identify the comparator and the trial framework.
7. Check whether groups received comparable follow-up and co-interventions.
8. Review participant flow, crossovers, exclusions, and missing data by group.
9. Confirm that the analysis follows the stated estimand and retains participants in their assigned groups when appropriate.
10. Check that the sample-size calculation and analysis account for the unit randomized and for any clustering, repeated measures, pairing, or crossover.
11. For multiple arms or confirmatory comparisons, identify the prespecified contrasts, multiplicity strategy, effect estimates, and confidence intervals.

For broader appraisal, use the research-study reading framework, trial-estimands and missing-data guide, and confidence intervals, p values, and multiplicity guide.

Important limitations

• Randomization is a process, not a word in the abstract.
• Exact baseline balance is not guaranteed, and baseline significance tests do not validate the allocation process.
• Allocation concealment and masking address different stages of a trial.
• “Double blind” is ambiguous unless the masked roles are named.
• Open-label trials can answer useful questions, but knowledge of assignment may introduce additional bias.
• The control label does not substitute for reporting what each group actually received or how the group was formed.
• A placebo or sham is not guaranteed to be physiologically inert or risk-free.
• Non-inferiority and equivalence require justified prespecified margins, suitable conduct, and confidence-interval evaluation; absence of superiority is insufficient.
• External controls and uncontrolled change-from-baseline analyses require stronger assumptions than randomized concurrent comparisons.
• The control group, estimand, and trial framework determine which conclusion the trial can support.
• Randomized evidence about a defined intervention does not establish comparability with another material sharing a name; see the FDA-approved drug versus research-material guide.

Frequently asked questions

Does randomization guarantee identical groups at baseline?

No. It creates a chance-based allocation process that balances prognostic factors in expectation, not necessarily exactly in one realized trial.[1][9]

Is allocation concealment the same as masking?

No. Allocation concealment prevents foreknowledge of the next assignment from influencing enrollment. Masking limits assigned-intervention knowledge after assignment.[3][4]

What does “double blind” mean?

The term is ambiguous. A clear report names the masked roles and explains how masking was achieved and maintained.[4][10]

Does an open-label trial have no value?

No. It can answer important questions, but readers should examine prespecification, standardized procedures, masked outcome assessment where possible, co-interventions, dropout, and the remaining knowledge-driven risks.[4][8][10]

Are placebo, sham, no treatment, and usual care interchangeable?

No. They can differ in what participants receive, masking feasibility, co-interventions, and the contrast estimated. The report must describe the actual intervention and comparator rather than rely on a label.[6][11]

Does similarity to an active control prove non-inferiority or equivalence?

No. Those conclusions require justified prespecified margins, a suitable design and conduct, and confidence intervals that satisfy the relevant decision rule.[1][6]

Is an external control equivalent to a randomized concurrent control?

No. External comparisons lack randomized assignment to the control and therefore require stronger comparability assumptions, including assumptions about unmeasured differences.[6]

Primary and official sources

1. FDA: ICH E9 Statistical Principles for Clinical Trials
2. CONSORT 2025 Explanation and Elaboration: Sequence Generation
3. CONSORT 2025 Explanation and Elaboration: Allocation Concealment
4. CONSORT 2025 Explanation and Elaboration: Who Was Blinded
5. ClinicalTrials.gov: Protocol Registration Data Element Definitions
6. FDA: ICH E10 Choice of Control Group and Related Issues
7. CONSORT 2025 Explanation and Elaboration: Trial Design
8. FDA: ICH E6(R3) Good Clinical Practice
9. CONSORT 2025 Explanation and Elaboration: Baseline Data
10. CONSORT 2025 Explanation and Elaboration: How Blinding Was Achieved
11. CONSORT 2025 Explanation and Elaboration: Intervention and Comparator

Review and corrections

Sources checked: September 1, 2026. Questions or corrections can be sent through the Serum Research contact page.

Editorial responsibility: Serum Research prepared and maintains this educational page. No independent scientific, statistical, regulatory, or clinical review is claimed. See how Serum Research reviews evidence or return to the Research Peptide and Compound Library.