Whenever feasible, randomized trials should incorporate mechanisms for concealing study group assignment from investigators, study staff, and participants. In traditional, individually randomized trials, study group assignment is concealed from the investigators before randomization to protect against treatment selection bias. For example, a clinician enrolling patients in an individually randomized trial may believe the study intervention would not be good for older patients. If that clinician is able to predict the next study group assignment, the clinician might hold back older patients from being assigned to the intervention group or might only “randomize” older patients when assignment to the control condition is likely. Concealment of the allocation of participants to study groups helps to prevent selection bias. In other words, “Allocation concealment is preventing the subversion of the randomization process” (Morris et al 2007).

After randomization, masking (also called "blinding") is used to guard against a placebo effect or biased outcome assessment. Masking is important for enabling the study team to measure and record the primary outcome objectively, without knowledge of the actual treatment assignment. A trial may be single-masked, in which either the participants do not know whether they are receiving the study intervention or the researchers and clinical staff do not know but he participants do; double-masked, in which both the participants and the researchers and clinical staff do not have access to this information but the study statisticians do; or triple-masked, in which the statisticians are also unaware of which participants were assigned to which study groups.

In cluster randomized trials, concealment is not usually a concern because the clusters are identified in advance and are randomized at the same time. Thus, unlike individually randomized trials, in which participants enter the study over time, there is no opportunity in a cluster randomized trial to “predict” the assignment and alter behavior in response to that knowledge. However, it is important to obtain written assurances before randomization that each cluster will comply with the assigned strategy for the duration of the study.

Masking, on the other hand, is usually either impossible or impractical in cluster randomized trials. Most interventions that call for cluster randomization need to be disclosed to those who are implementing them. However, it is important to maintain as much objectivity as possible in recording the outcome assessments. Objective measures such as hospitalization and mortality should be collected the same way in all arms of the trial. For quality-of-life outcomes, care should be taken to ensure the measures do not refer to the conditions of the trial (for example, beginning a questionnaire with "Did you find that your dialysis time influenced how you feel on a daily basis?") Likewise, for a handwashing campaign, the assessment of nosocomial infection should be conducted according to objective criteria applied in the same way in all study arms.

Masking in Pragmatic Clinical Trials

The individually randomized, double-masked, placebo-controlled study continues to be the gold standard for generating evidence for biomedical interventions. However, for pragmatic clinical trials to estimate real-world effectiveness, it is generally necessary to allow greater flexibility in certain standards of research design and conduct, such as by broadening the participant eligibility criteria, relaxing the specificity of the intervention, and choosing relevant comparison conditions. To ensure the integrity of the study findings and reduce bias, it is crucial to prespecify and justify all such choices. Routine study monitoring may allow the investigators to carefully modify certain aspects of a pragmatic trial while it is in progress to ensure the clinical relevance and rigorous conduct of the trial.

Masking in pragmatic trials was not well described until recently (Christian et al 2020; Meyers and Yu 2020). In this section, we offer guidance for prespecifying the masking status of different categories of study personnel, with adequate justification, and to maintain such status throughout the conduct of the study. See Table 1 below for a template of masking status information.

In a pragmatic clinical trial of a behavioral intervention, the individuals delivering the interventions often know whether they are administering the study intervention or the control condition. Although they are unmasked to treatment assignment, it is important to keep them masked to the outcome data of the trial to minimize any unintentional influence they might have on the outcomes. On the other hand, outcome assessors in the trial should remain masked and should be uninvolved in intervention delivery if the trial design allows this. An example of when an outcome assessor may be unmasked is a trial with cluster randomization at the clinic level in which outcome data are collected from the electronic health record in the course of routine medical practice. In this example, if the outcome assessor is a practice nurse, the outcome assessor may be aware of the intervention status. This awareness could lead to bias, and the reporting of the study should be transparent about the limitations arising from unmasked outcome data collection.

Investigators should also consider masking status with respect to adherence and fidelity assessments to ensure these outcomes are measured objectively and uniformly. For example, an investigator may track overall adherence but still be masked to these assessments by study arm (that is, by not reviewing summary data by study arm). In addition, fidelity raters may be masked to intervention assignment. Reliable and valid fidelity instruments could also reduce bias in the assessment and help improve study quality.

A trial’s lead study statistician is often masked to both treatment assignment and outcome data until all data collection is complete and the study’s database is locked. This allows the masked statistician to make any necessary changes to the statistical analysis plan due to issues with implementation of the trial while an unmasked statistician is able to prepare analysis and reports by treatment arm for the data and safety monitoring board (DSMB). (For example, slow recruitment may require an interim reassessment of sample size or decisions about protocol violations that could affect enrollment status.) In a cluster randomized trial, the unmasked statistician should also handle randomization, especially if there is a small number of clusters, unless the trial will be fully unmasked to treatment assignment. (For cluster randomized trials, see also the “timeline cluster tool” proposed by Caille and colleagues [2016] for assessing bias, including that in relation to masking). Sometimes, stepped-wedge, cluster randomized trials are unmasked to treatment assignment, but statisticians can remain masked to outcomes (or only be informed of aggregate outcome missing rates) depending on the setting. All these decisions should be clearly stated in advance and be transparently described in reports of the trial’s results.

Table 1. Template of Masking Status Information

Single-Masked Trials

A double-masked design may not be feasible in some trials of behavioral interventions, such as trials of cognitive behavioral therapy, yoga, and mindfulness. Although some studies may include a sham intervention as the control condition, other studies use a single-masked design in which study participants are not masked to intervention assignment but the investigators and statisticians are masked. In single-masked trials, it is crucial that the interventionists do not see any outcome data from the trial before the data analysis—including summary statistics, and especially data summarized by arm—and that the assessors be masked to treatment assignment. The trial design should be structured to minimize bias at various stages, from participant selection to outcome assessment, ensuring that the conclusions drawn from the trial are as reliable and valid as possible.

Case Study of a Single-Masked Trial: BackInAction

BackInAction, an NIH Collaboratory Trial, was an individually randomized group treatment trial of the effectiveness of acupuncture for older adults with chronic low back pain (DeBar et al 2023; DeBar et al 2025). Due to the nature of the intervention, study participants and the acupuncturists delivering the intervention were not masked to treatment assignment. The investigators, the lead statistician, and the statistician responsible for conducting the primary analysis were masked to both treatment assignment and outcome data until data collection was complete. Masking to treatment assignment meant not knowing (and not having access to data that would make it possible to know) to which treatment arm a participant had been assigned. Masking to outcome data meant not having access to follow-up study data for individual participants, in aggregate, or by treatment arm.

Early in the BackInAction trial, a member of the study team who was not a statistician for the project but had analysis experience served as an unmasked statistician for the trial so they could attend closed meetings of the DSMB and respond to questions. Toward the end of the trial, another statistician was unmasked to treatment assignment to help address questions about data missingness. (Note: Because there was no planned interim monitoring of outcome effectiveness, the only outcome data the unmasked statisticians were able to access were the rates of primary outcome missingness by intervention arm, but not comparisons of intervention effectiveness estimates between arms.)

The study team did not prespecify the masking requirement in the initial grant proposal for the trial, but they were able to be creative to respond to requests about masking from the funder and the DSMB. In hindsight, it would have been ideal to have planned for the staff needed to support this team structure in the proposed budget—specifically, a masked statistician assigned to conduct all analysis, or modify the statistical analysis plan, during the course of the study and an unmasked statistician to work with programmers to prepare the closed DSMB reports and to participate in related closed meetings.

Assigning both masked and unmasked statisticians in BackInAction had benefits. For example, during the study, the DSMB noted that the rate of data missingness was high at the 6-month time point (that is, the primary endpoint) and differed between 2 of the 3 study arms. The DSMB encouraged the study team to take additional steps to minimize data missingness. The unmasked statistician was able to have in-depth discussions in closed session with the DSMB meeting and to consult with the unmasked statistician at the NIH. The lead statistician, who remained masked, provided guidance about the analysis to address the differential missingness. The study team also updated the statistical analysis plan to provide greater specificity about when to classify the primary analysis to be a pattern-mixture imputation approach that assumes nonignorable missing data assumptions vs the simpler, yet stronger, assumptions of complete-case data analysis with covariate adjustment. This consisted of more explicitly defining what imbalance by treatment arm meant (for example, 10% difference between arms in missing the primary outcome at the primary time point) and the overall approach to specifying the imputation modeling. This process allowed updating of the statistical analysis plan by a masked statistician with oversight by the DSMB and the funder, who were able to approve those updates according to best practices.

Table 2. Masking Status Information for BackInAction

*  Had masked treatment assignment, but did not know which group was A, B, or C.

Recommendations

The study team should specify in advance which members of the team will be masked and unmasked to all, or aspects of, treatment assignment and outcome data and should keep study personnel masked unless unmasking is necessary for the conduct of their roles. Defining each team member’s masking status early in the study, ideally during preparation of the initial grant proposal and budget, allows for the optimal personnel structure for the monitoring and conduct of the trial through all stages, transparency of data collection, and the ability to address potential bias that may arise as a limitation in reporting the outcomes of the study.