4.5 Governance
The gap between the plan and the decision
A pre-specified alpha-spending function, a set of pre-computed stopping boundaries, and a table of operating characteristics are the statistical components of an interim analysis plan. They are necessary. They are not sufficient.
The gap between the statistical plan and the actual interim decisions is filled by governance: the organizational structure that determines who sees the interim data, in what form, with what authority, under what constraints, and with what documentation requirements. When governance is well-designed, the statistical plan is implemented faithfully, the interim decisions are made by the right people with the right information, and the final result retains interpretive integrity. When governance fails—when unblinded data reach parties who should not see them, when stopping decisions are made outside the pre-specified rules, when documentation is incomplete—the statistical plan is irrelevant. The trial has been compromised in ways that no post-hoc analysis can repair.
Governance is not the DSMB’s charter alone, though the charter is its most visible document. It is the complete system of information access rights, decision authority, documentation requirements, and communication protocols that surrounds every interim analysis from the moment the data are extracted to the moment the decision is documented. Designing this system is as important as designing the statistical boundaries, and it requires as much deliberate attention.
The DSMB: composition and independence
The Data Safety Monitoring Board is the independent committee that reviews interim data and makes recommendations about trial continuation. Its composition and its independence are the first governance questions.
Composition requires expertise across the domains relevant to the interim decision: clinical expertise in the indication, statistical expertise in sequential trial monitoring, and—when safety signals are plausible—expertise in the specific safety domains at risk. A DSMB composed entirely of clinicians with no statistical monitoring expertise will have difficulty interpreting the alpha-spending context of an interim efficacy signal. A DSMB composed entirely of statisticians may lack the clinical judgment to evaluate whether a safety finding meets the threshold for stopping. The composition should reflect the full range of decisions the DSMB may be asked to make.
Independence requires that DSMB members have no financial or professional relationship with the sponsor that would compromise their ability to make decisions contrary to the sponsor’s interest. Advisory board membership, consulting relationships, equity positions, and institutional research funding all create potential conflicts that should be evaluated at the time of appointment. The standard is not the absence of any relationship—that standard would exclude many of the most experienced clinical trialists—but the absence of relationships significant enough to bias judgment in the direction of sponsor interests.
Independence also requires procedural protection: the DSMB’s deliberations should be confidential, the interim data should be seen only by the DSMB and the independent statistician, and the DSMB’s recommendation should be transmitted to the sponsor without disclosure of the arm-specific interim data that formed its basis. A recommendation that says “the DSMB recommends continuation” tells the sponsor what they need to know. A recommendation that reveals the interim hazard ratio, arm-specific event rates, or subgroup results tells the sponsor much more than they need to know for the continuation decision—and what they learn may influence their conduct of the trial in ways that were not planned.
The charter: what it must specify
The DSMB charter is the governance document that specifies, in advance of any interim data review, how the DSMB will operate. A charter that is specific enough to be useful is a charter that could be handed to a new DSMB member with no prior knowledge of the trial and would tell them everything they need to know to participate in an interim meeting.
The charter must specify the composition of the DSMB and the criteria for recusal. It must specify the information that will be presented at each interim analysis—what the DSMB will see, at what level of aggregation, with what annotation. It must specify the statistical boundaries that the DSMB will apply and the operating characteristics those boundaries produce. It must specify the decision rules—stopping for efficacy, stopping for futility, stopping for safety—with their respective evidentiary standards and decision authorities. It must specify the communication protocol between the DSMB and the sponsor: what the DSMB will communicate, in what form, on what timeline, and through what channel.
It must also specify what the DSMB is not authorized to do. This is as important as specifying what it is authorized to do. A DSMB that is not explicitly authorized to recommend eligibility changes, endpoint modifications, or analysis modifications should not make such recommendations—and if it does, the sponsor should not act on them without the full protocol amendment process that such changes require. The charter is the boundary of the DSMB’s authority, and that boundary should be explicit.
Information flow: who sees what, when
The most consequential governance design decision is the structure of information flow at interim analyses. This decision determines what the sponsor knows about the interim results, and what they can do with that knowledge.
The cleanest governance structure is full firewall: the independent statistician prepares the interim analysis report for the DSMB, the DSMB reviews the arm-specific interim data and makes a recommendation, and the sponsor receives only the recommendation—not the data on which it was based. Under this structure, the sponsor cannot use knowledge of the interim trend to make decisions about the trial—commercial, regulatory, or operational—that are influenced by the interim result. The firewall protects the trial’s scientific integrity by preventing the sponsor’s response to interim data from becoming a source of bias.
This structure has costs. Sponsors have legitimate operational interests that may depend on interim trial characteristics: site management, supply chain planning, regulatory submission timelines, co-development partnerships. Knowing only “the DSMB recommends continuation” does not give the sponsor the information it needs to manage these operational decisions optimally. The cost of the firewall is operational inefficiency; the benefit is scientific integrity.
The intermediate structure—where the sponsor receives summary interim data without arm-specific comparisons—attempts to preserve operational utility while limiting the information that could bias the trial. This structure requires careful specification of what constitutes “arm-specific comparison” information. A report that shows overall event rates without arm attribution tells the sponsor something about the event rate assumption accuracy; a report that shows time trends in event accumulation may implicitly convey arm-specific information through pattern recognition. The intermediate structure requires active enforcement of the information boundary, which requires a statistician who understands what information is sensitive and what is not, and a process that systematically prevents sensitive information from crossing the firewall.
The weakest governance structure—where the sponsor has access to the full interim dataset, subject to a prohibition on using the information to modify the trial—relies on behavioral controls rather than structural ones. The prohibition is unenforceable in practice: the sponsor’s decisions about the trial, from site management to communication strategy to regulatory interactions, will inevitably be influenced by knowledge of the interim trend, even if the influence is unconscious. This structure is appropriate only in situations where the sponsor’s operational involvement in the trial requires real-time data access—certain adaptive designs, platform trials—and where compensating governance controls exist. For standard fixed designs with pre-specified interim analyses, it should not be the default.
The open session, closed session, and executive session
A well-designed interim meeting typically has three phases, and the governance implications of each are distinct.
The open session is attended by representatives of the sponsor, the clinical operations team, and the regulatory team, along with the DSMB. The open session presents data that do not reveal the treatment arm comparison: overall enrollment progress, site performance, protocol deviation rates, overall safety events without arm attribution, and any operational issues that require joint discussion. The DSMB can ask questions and the sponsor can present information, but the arm-specific interim efficacy and safety data are not discussed.
The closed session is attended only by DSMB members and the independent statistician. The arm-specific interim data—efficacy results, safety results, subgroup results—are presented and discussed in closed session. The DSMB deliberates, applies the pre-specified stopping rules, and formulates its recommendation.
The executive session is a DSMB-only session, without the independent statistician, in which the DSMB may discuss the recommendation before transmitting it to the sponsor. Not all interim analysis structures include this phase, but it is valuable when the DSMB’s deliberations involve judgments about the charter’s interpretation—what the stopping rules require—that the DSMB may wish to discuss without the constraint of external presence.
This three-phase structure is not bureaucratic excess. It is the mechanism by which the arm-specific interim data remain within the closed session while the non-arm-specific operational data are shared openly. When the session structure is not maintained—when arm-specific data are discussed in the open session, when sponsor representatives attend the closed session, when the DSMB’s deliberations are communicated to the sponsor in more detail than the recommendation—the governance has failed, regardless of whether the stopping decision itself was correct.
Documentation: the governance record
Every interim analysis should produce a documentation record that specifies what was reviewed, who attended which sessions, what data were presented, what the DSMB deliberated on, what recommendation was made, and—critically—what the stopping rule status was at the time of the recommendation. This record is not primarily for regulatory inspection, though it will be examined during regulatory review. It is for the scientific record: the basis on which the interim decision was made should be preserved and interpretable independent of the memory of the participants.
The documentation should be specific enough to allow reconstruction of the interim analysis. It should identify the data snapshot date, the data completeness rate, the number of events in each arm (in the closed session record), the test statistic value relative to the boundary, and the DSMB’s reasoning if the recommendation involved judgment beyond mechanical application of the stopping rules.
When the stopping rules were not mechanically controlling—when the DSMB chose to continue despite approaching a futility boundary, or chose to continue despite crossing a non-binding efficacy boundary, or raised a safety concern without crossing a statistical threshold—the reasoning must be documented. An undocumented deviation from the pre-specified rules is a governance failure regardless of whether the deviation was scientifically justified. The regulatory reviewer who later asks “why did the DSMB continue after the interim test statistic was 1.89 against a boundary of 2.01?” must have a documented answer. “The DSMB exercised clinical judgment” is not a documented answer. The clinical judgment must be recorded.
When governance fails
Governance failures in interim analysis have a characteristic signature in the regulatory review: a timeline of events that is hard to reconcile with the premise that the trial was conducted without knowledge of interim results. Site activation patterns that change after the interim analysis. Eligibility criteria that are clarified in ways that affect enrollment composition after the interim. Communication strategy shifts that coincide with interim meetings. These patterns are not proof of misconduct—they may reflect legitimate operational decisions—but they raise questions that the governance documentation should answer.
When the governance documentation is incomplete—when the interim analysis records do not specify who attended which sessions, when the DSMB charter does not specify the information flow protocol, when the sponsor’s response to the DSMB recommendation is not documented—the patterns cannot be explained, and the integrity of the trial is uncertain. This uncertainty does not necessarily prevent regulatory approval, but it complicates the review and can result in conditions or labeling language that the sponsor did not anticipate.
Governance design is not an afterthought. It is the mechanism by which the pre-specified interim analysis plan is actually implemented—faithfully, under conditions that may be very different from what the design team anticipated. A well-governed trial is one that can be explained from the inside: every decision at every interim analysis can be traced to a documented basis, and that basis is the pre-specified plan. A poorly governed trial is one that can only be explained from the outside: the decisions are defensible in retrospect, but the record that would establish their defensibility at the time they were made does not exist.
References: Ellenberg, Fleming, and DeMets, Data Monitoring Committees in Clinical Trials: A Practical Perspective (2002); ICH E6(R2) Good Clinical Practice (2016); FDA Guidance for Industry: Oversight of Clinical Investigations—A Risk-Based Approach to Monitoring (2013); Herson, Data and Safety Monitoring Committees in Clinical Trials (2009).