Interim Analysis and DSMB Operations

Definition

An interim analysis is a pre-specified statistical evaluation of accumulating trial data conducted during the course of a blinded, ongoing trial. Unlike exploratory interim reviews, interim analyses with pre-specified stopping rules protect the Type I error rate while allowing the trial to continue to its planned endpoint.

"Interim analyses can evaluate whether a trial should be stopped early for efficacy, futility, or safety. Properly designed interim analyses with appropriate statistical methods allow the trial to continue efficiently to its planned target while maintaining the overall Type I error rate." — FDA Adaptive Designs Guidance (November 2019, Final)

"Interim analyses should be justified based on the anticipated required persuasiveness, the number of participants available at interim, and the duration of follow-up at interim. Interim decision rules must be pre-specified and based on pre-defined statistical thresholds, not clinical judgment." — ICH E20 Adaptive Designs (2025, Draft)

Key distinction: An interim analysis with a pre-specified stopping rule preserves the Type I error rate; an unblinded exploratory interim review without pre-specified stopping rules does not.

Regulatory Position

FDA requires (Adaptive Designs 2019, Final):

• Pre-specification: All interim analyses, stopping rules, and decision criteria must be fully specified in the Statistical Analysis Plan (SAP) before unblinding. Post hoc interim analyses are exploratory and cannot support primary efficacy claims.
• Statistical rigor: Interim analyses must employ recognized methods (alpha spending functions, group sequential designs) that maintain strong control of the family-wise error rate (FWER) at the specified significance level.
• DSMB independence: For trials with interim stopping rules, an Independent Data Safety and Monitoring Board (DSMB) must monitor unblinded data and make stopping recommendations. Sponsor access to unblinded comparativel data should be restricted or absent.
• Binding vs. non-binding: Efficacy stopping rules are typically binding (if met, trial must stop); futility stopping rules may be binding or non-binding (recommendation to consider stopping).

ICH E20 (Draft, 2025) adds:

• Interim timing justification: Sponsors must justify the timing of interim analyses based on the information fraction (events/expected events) or event count, not calendar time alone.
• Estimation adequacy: At interim, sample size and follow-up duration should be sufficient to estimate the treatment effect reliably for regulatory decision-making.
• Sensitivity analyses: When interim OS analyses are conducted in trials with anticipated crossover, pre-specified sensitivity analyses (e.g., RPSFT, IPCW) must be included to assess robustness.

Source: FDA Adaptive Designs Guidance (November 2019) = Final; ICH E20 Adaptive Designs Draft (2025) = Draft

Types of Interim Analyses

1. Efficacy Interim Analysis

Assess whether the treatment effect at interim is sufficiently large to declare the trial positive early. The efficacy boundary (or "stopping rule") specifies the threshold p-value or test statistic at each interim and final analysis that permits early stopping.

Use cases:

• When overwhelming benefit is expected early (e.g., large effect size, high event rate)
• To enable early approval if treatment effect is substantially larger than anticipated
• To reduce exposure of control arm patients to untreated disease

Oncology example:

Primary endpoint: PFS (progression-free survival, HR = 0.65)
Expected events at final analysis: 350

Interim 1 at 50% events (175 events):
  Z-boundary (one-sided) = 2.77  (p ~ 0.003)
  If p < 0.003: can stop early for efficacy

Interim 2 at 75% events (262 events):
  Z-boundary = 2.28  (p ~ 0.011)
  If p < 0.011: can stop early for efficacy

Final analysis at 100% events (350 events):
  Z-boundary = 2.04  (p ~ 0.021)

Properties:

• Strongly controls Type I error rate at the nominal level (e.g., α = 0.025 one-sided)
• Binding: if efficacy boundary is crossed, trial must stop
• More restrictive at interim (harder to declare significant) than final analysis
• Appropriate when treatment benefit is expected to appear early

2. Futility Interim Analysis

Assess whether the trial is unlikely to demonstrate efficacy at final analysis. If futility boundaries are crossed, continuing enrollment may be wasteful. Futility rules protect power and reduce exposure of patients to an ineffective treatment.

Binding vs. Non-Binding:

Aspect	Binding Futility	Non-Binding Futility
Decision if boundary crossed	Trial must stop (mandatory)	DSMB recommends stopping; sponsor can choose to continue
Type I error control	Maintained (boundary properly adjusted)	Maintained (conditional on decision to stop)
Type II error	Increased (trial cannot recover if boundary crossed incorrectly)	Protected (can continue if DSMB allows)
Regulatory acceptance	FDA generally comfortable; requires strong justification	More common; allows flexibility to continue if clinical context supports
When to use	High confidence in effect size; want to avoid unethical continuation	Uncertainty in effect size; want option to continue despite interim futility signal

Common futility thresholds in oncology:

• Conditional Power (CP) < 20%: "If we continue to final analysis, what is the probability we will achieve significance?" If CP < 0.20 at interim, futility boundary is crossed.
• P-value > (non-adaptive threshold): e.g., "If PFS HR upper confidence interval includes 1.0 at interim, stop for futility."
• Posterior probability: Bayesian approach — "What is the posterior probability that true HR < target HR?" If too low, stop.

Oncology example (Conditional Power):

Primary endpoint: OS (HR target = 0.75, 80% power)
Expected events at final: 220

Interim at 110 events (50% information):
  Observed HR = 0.85 (favoring treatment, but modest)
  Current p-value = 0.08 (not significant)

  Conditional Power calculation:
    "If true HR = 0.85 (observed) and we continue to 220 events,
     probability of achieving p < 0.025 at final = 12%"

  Decision: CP = 12% < 20% threshold -> Futility boundary crossed
  Recommendation: Consider stopping trial (non-binding); may continue with DSMB approval

R implementation:

library(gsDesign)

# Define group sequential design with futility
design <- gsDesign(
  k = 2,           # 1 interim + final
  test.type = 2,   # Two-sided
  alpha = 0.025,   # One-sided
  beta = 0.20,     # 80% power
  sfu = sfLDOF,    # O'Brien-Fleming spending
  sfupar = 0,      # No futility spending (non-binding)
  sflpar = 0       # Or sfLDPower for binding futility
)

# Extract futility boundaries
design$lower$bound  # Futility (lower) boundaries
design$upper$bound  # Efficacy (upper) boundaries

3. Safety Interim Analysis

Evaluate accumulating adverse event data to identify signals of serious, unexpected toxicity. Safety interims are typically non-binding recommendations to the DSMB/DMC.

Use cases:

• Trials with novel mechanism or new drug class (unknown safety profile)
• Combination therapies with potential for drug-drug interactions or synergistic toxicity
• Patient populations at higher risk (elderly, hepatic/renal impairment)
• When serious adverse events are anticipated but need monitoring (e.g., immunotherapy with autoimmune toxicity)

Oncology metrics monitored:

• Grade 3+ adverse events (CTCAE v5.0)
• Treatment-related deaths
• Dose reductions or treatment discontinuations
• Serious adverse events (SAEs): any event requiring hospitalization, organ dysfunction, or death
• Safety signal threshold: e.g., "If incidence of severe (Grade 3+) hepatotoxicity > 10% in treatment arm vs. < 2% in control, pause enrollment pending safety review"

Regulatory perspective:

• FDA expects proactive safety monitoring and mid-trial protocol amendments if safety signals emerge
• Pre-specified safety stopping rules are not required but encouraged for novel mechanisms
• DSMB charter should specify safety monitoring procedures

4. Administrative Interim Analysis

A planned "look" at trial data for non-comparative purposes: checking for data quality, enrollment status, event accrual rate, and protocol deviations — without unblinding treatment assignments or statistical testing of primary endpoint.

Use cases:

• Monitor enrollment trajectory vs. projected timelines
• Assess event accrual rate and forecast final analysis date
• Verify data completeness and quality assurance
• Check protocol adherence (deviations, inclusion/exclusion violations)
• Plan resource allocation (DSMB member schedules, statistical programming)

Regulatory perspective:

• Administrative interims do not require pre-specification in SAP (though documentation is prudent)
• No alpha spending; Type I error not affected
• Blinded interim (no treatment comparison) is fully transparent and acceptable

---

Timing of Interim Analyses

Information Fraction (Preferred)

Define interim timing based on the proportion of the target information (events or sample size) accrued, not calendar time.

For time-to-event endpoints (OS, PFS, DFS):

Information = (actual events) / (planned events)

Interim 1: 50% information = 175/350 events
Interim 2: 75% information = 262/350 events
Final:    100% information = 350/350 events

Advantages:
  - Timing is predictable and adaptive to event accrual rate
  - Appropriate follow-up accumulates (not biased toward fast-enrolling sites)
  - Statistical power and alpha spending are valid

For binary endpoints (ORR, safety):

Information = (actual patients with events) / (planned patients with events)

Interim 1: 50% information = 50/100 response events
Interim 2: 75% information = 75/100 response events
Final:    100% information = 100/100 response events

Event-Count Triggers (Alternative)

Define interim analyses to occur at specific cumulative event counts rather than proportions.

Example (OS primary, n=400 per arm):
  Interim 1: Schedule when 150 OS events have occurred
  Interim 2: Schedule when 225 OS events have occurred
  Final:     Schedule when 300 OS events have occurred

Note: Calendar-driven interim timing (e.g., "every 6 months") is
discouraged for efficacy analyses because event timing is unpredictable
and can lead to imbalanced information fractions.

Calendar-Time Triggers (Limited Use)

Calendar-based interim scheduling (e.g., "review at 6, 12, 18 months") is acceptable for administrative and safety monitoring but not recommended for efficacy analyses because event accrual is unpredictable.

---

Alpha Spending Functions

Alpha spending functions govern how the overall Type I error rate (e.g., α = 0.05 two-sided, or 0.025 one-sided) is distributed across interim and final analyses while maintaining strong FWER control.

O'Brien-Fleming (OBF) Spending Function

The most conservative spending function. Allocates very little alpha at early interims and most alpha at the final analysis. Results in efficacy boundaries that are very restrictive at interim (high bar for early stopping) but permissive at final (close to nominal).

Formula (one-sided):

α(t) = 2 * [1 - Φ(z_α * √(1/t))]

where:
  t = information fraction (0 < t ≤ 1)
  z_α = critical value for overall alpha (z_0.025 = 1.96 for 0.025 one-sided)
  Φ = standard normal CDF

Interpretation:
  At t = 0.5 (50% info): α_interim ~ 0.003  (very stringent)
  At t = 0.75 (75% info): α_interim ~ 0.011 (stringent)
  At t = 1.0 (100% info): α_interim ~ 0.025 (nominal)

When to use:

• Trials where early stopping for efficacy is unlikely but acceptable if treatment effect is very large
• When you want to preserve power for final analysis (minimize power loss from interim spending)
• Most common in oncology (conservative, unlikely to cause false early positives)

Boundary example (two interims, α_total = 0.025 one-sided):

Information  Spending   Z-boundary  p-value threshold
Interim 1:   50%        0.003       Z > 2.75  (p < 0.003)
Interim 2:   75%        0.011       Z > 2.29  (p < 0.011)
Final:      100%        0.025       Z > 1.96  (p < 0.025)

Pocock Spending Function

More liberal than O'Brien-Fleming. Allocates approximately equal alpha to each analysis stage. Results in more permissive efficacy boundaries at interim but slightly higher boundary at final (p < 0.029 instead of 0.025 for two stages).

Formula (one-sided):

α(t) = α_total * ln(1 + (e-1)*t)

where t = information fraction

At t = 0.5: α_interim ~ 0.0108  (symmetric allocation)
At t = 1.0: α_interim = α_total

When to use:

• When early stopping is clinically valuable and likely (e.g., trials with very large anticipated effect sizes)
• Trials where cost/time savings from early stopping justify slightly higher alpha spending at interim
• Acceptable but less common in oncology than O'Brien-Fleming

Boundary example (two interims, α_total = 0.025 one-sided):

Information  Spending   Z-boundary  p-value threshold
Interim 1:   50%        0.0108      Z > 2.31  (p < 0.010)
Interim 2:   75%        0.0191      Z > 2.06  (p < 0.020)
Final:      100%        0.025       Z > 1.96  (p < 0.025)

Lan-DeMets (Alpha Adaptive) Spending Function

Flexible framework that mimics a desired spending function (OBF, Pocock, linear, etc.) but adapts to unequally spaced information fractions. Allows flexible interim scheduling while maintaining the spending rate of the reference function.

Use case:

Planned: Interim at 50%, final at 100%
Actual: Interim 1 at 45%, Interim 2 at 72%, Final at 100%
        (enrollment was slightly faster than expected)

Lan-DeMets adapter recalculates spending boundaries to match
the intended spending curve (e.g., OBF) at the actual information fractions.

Formula:

α(t_actual) = reference_spending_function(t_actual)

This allows adaptive interim scheduling without pre-defining
exact information fractions, while maintaining the statistical properties
of the reference function.

When to use:

• Trials where interim information fractions are uncertain (variable event accrual rates)
• Adaptive designs with sample size re-estimation
• Responsive interim scheduling (e.g., "review when 50-60% events accrue")

Kim-DeMets (Power Family) Spending Function

A parametric family of spending functions indexed by a shape parameter ρ (rho). Offers a continuum between Pocock (ρ = 0) and O'Brien-Fleming (ρ = 2).

Formula (one-sided):

α(t) = α_total * t^ρ

ρ = 0   → Linear spending (uniform)
ρ = 1   → Intermediate
ρ = 2   → O'Brien-Fleming-like (conservative)

Example: ρ = 1.5 (compromise between Pocock and OBF)
  At t = 0.5: α_interim ~ 0.0088
  At t = 0.75: α_interim ~ 0.0174
  At t = 1.0: α_interim = 0.025

When to use:

• Fine-tuning the balance between power preservation (OBF) and early stopping opportunity (Pocock)
• Trials with clear clinical rationale for a specific "aggressiveness" level
• Sensitivity analyses showing robustness to spending function choice

---

Futility Spending (Beta-Spending)

Futility boundaries consume beta (Type II error rate), not alpha. The framework mirrors alpha spending but governs how much beta is "available" at each stage for futility assessment.

Pampallona-Tsiatis Beta-Spending Function

The most common approach in oncology. Allocates beta across stages using a parametric function (typically linear or inverse normal).

Formula (one-sided):

β(t) = β_total * t^ρ_beta

ρ_beta = 1  → Linear beta spending
ρ_beta = 2  → Conservative (inverse normal-like)

Common in oncology: Linear beta spending (equal allocation across stages)
  At t = 0.5: β_interim ~ 0.10 (of 0.20 total)
  At t = 1.0: β_interim = 0.20

Conditional Power at interim t:
  CP(t) = P(reject H_0 at final | observed data at interim)

  If CP < (1 - β(t)) = 0.80, futility boundary is crossed.

Implementation:

library(gsDesign)

# Design with binding futility (sequential conditional power)
design_futility <- gsDesign(
  k = 2,
  test.type = 2,
  alpha = 0.025,
  beta = 0.20,
  sfu = sfLDOF,  # O'Brien-Fleming for efficacy
  sfl = sfLDOF   # O'Brien-Fleming for futility (binding)
)

design_futility$lower$bound  # Futility boundaries
design_futility$upper$bound  # Efficacy boundaries

Oncology example (Conditional Power with Linear Beta Spending):

Primary endpoint: PFS, HR target = 0.75, 80% power
Expected events: 250

Interim 1 at 125 events (50% info):
  Observed HR = 0.92 (favoring treatment, modest)
  Current p-value = 0.15

  Beta allocation at 50% info: 0.10 (of 0.20 total)
  Remaining beta for final: 0.10

  Conditional Power = 18% (calculated from observed HR and variance)
  Futility threshold = 1 - 0.10 = 90% CP required to continue

  Decision: CP (18%) < threshold (90%) -> Futility boundary crossed
            DSMB recommends considering trial termination

---

DSMB Charter: Composition, Independence, and Operations

Composition

An Independent Data Safety and Monitoring Board (DSMB) typically consists of 3–5 members with diverse expertise:

Role	Expertise	Number
Chair	Oncology/biostatistics	1
Biostatistician	Group sequential design, interim analysis, data security	1
Oncologist	Clinical judgment, safety, benefit-risk assessment	1
Additional members (optional)	Industry experience, patient advocacy, other relevant specialty	0–2

Independence criteria (FDA/ICH E20 requirement):

• No financial interest in trial sponsor or treatment
• No prior/current affiliation with sponsor (excluding contract service)
• No direct involvement in trial conduct (not site investigator or steering committee member)
• Able to maintain confidentiality of unblinded data
• Published track record in relevant field (preferred)

DSMB Charter Requirements

A written charter must specify:

1. Membership & roles: Names, affiliations, expertise, terms of service (e.g., 2-year appointments)

2. Confidentiality & firewall:

   • DSMB members are unblinded to treatment assignment
   • Sponsor staff with access to DSMB recommendations: only statistical team (usually biostatistics) and trial leadership (PI, sponsor medical monitor)
   • Site investigators, enrollment staff, data management staff remain blinded
   • DSMB meeting materials are confidential; members cannot discuss unblinded results with anyone outside DSMB

3. Meeting frequency:

   • Interim analyses: typically 1 meeting per interim look
   • Safety monitoring: may be quarterly or ongoing depending on trial risk
   • Escalated safety: emergency meetings if serious adverse event signals emerge

4. Decision rules & voting:

   • Stopping recommendations: DSMB recommends (not mandates) stopping for efficacy, futility, or safety
   • Voting: typically unanimous consensus; if 1–2 dissent, document reasons
   • Binding vs. non-binding efficacy: If efficacy boundary crossed, DSMB must recommend (and sponsor should accept) stopping. Futility may be non-binding (DSMB recommends, sponsor can continue with justification).

5. Safety monitoring procedures:

   • Review serious adverse events at each meeting
   • Assess causality: treatment-related vs. unrelated
   • Define safety stopping rules (e.g., "If grade 4+ hepatotoxicity in >5% treatment arm, pause enrollment")
   • Monitor protocol deviations that could affect safety

6. Closed vs. open sessions:

   • Closed session: DSMB members + trial biostatistician (unblinded). Discuss unblinded data, efficacy, futility, safety signals.
   • Open session: DSMB + sponsor team (blinded). Discuss administrative/operational issues, enrollment, data quality (no interim analysis results).

Firewall Model (FDA Requirement for Adaptive Designs)

UNBLINDED DATA ACCESS (DSMB + Statistical Team)
├─ DSMB members (all members)
├─ Statistician conducting interim analyses (designated person)
└─ Medical monitor/trial leadership (restricted; summary-level only)

BLINDED DATA ACCESS (Sponsor & Investigators)
├─ Study sites (blinded to treatment assignment)
├─ Data management team (blinded)
├─ Enrollment/patient care staff (blinded)
└─ Biostatistics team preparing final analysis (blinded until final DB lock)

Practical implementation:

• Unblinded statistician works in separate "locked room" or secure virtual environment
• Interim analysis code/data kept in restricted repository (not shared with blinded team)
• DSMB minutes do not disclose specific p-values or effect sizes to sponsor; only recommendation (e.g., "Continue trial as planned" or "Recommend stopping for efficacy")
• Blinded team continues enrollment and follow-up based on DSMB recommendation without knowledge of results

---

SAP Template: Interim Analysis Section (Two-Look Design)

7. INTERIM ANALYSES

7.1 Overview

This trial will include one interim analysis conducted at approximately 50% 
of the target information (events/patients), with the option for an 
unscheduled safety interim if adverse events warrant. The interim analysis 
will assess efficacy and futility using pre-specified statistical stopping 
rules while maintaining strong control of the family-wise error rate (FWER) 
at alpha = 0.025 (one-sided).

The trial design is a group sequential design with two planned analyses 
(one interim, one final) using O'Brien-Fleming alpha spending for efficacy 
and linear beta spending for futility.

7.2 Interim Analysis Timing

The interim analysis will be scheduled when approximately 50% of the target 
event count has been accrued:

  Target primary endpoint events (final): 350 PFS events
  Interim analysis trigger: 175 PFS events

Expected timing: Month 24 (range 22–26 months depending on accrual rate)

Interim scheduling is based on events accrued, not calendar time. The Data 
Safety and Monitoring Board (DSMB) will be notified when the event count 
reaches 165–180 events so that the interim analysis can be scheduled 
approximately 2–4 weeks after reaching 175 events (to allow for latency in 
event reporting and data lock).

7.3 Efficacy Stopping Rule (Binding)

A pre-specified boundary for early stopping due to overwhelming efficacy will 
be evaluated. If the interim efficacy boundary is crossed, the trial MUST be 
stopped and results submitted for regulatory approval.

The efficacy boundary is based on O'Brien-Fleming alpha spending for one 
interim and one final analysis:

| Analysis  | Info Fraction | Target Events | Cumulative α Spent | Z-Boundary | p-Value Threshold |
|-----------|---------------|----------------|--------------------|------------|-------------------|
| Interim 1 | 50%           | 175 PFS events | 0.0035             | 2.75       | < 0.003           |
| Final     | 100%          | 350 PFS events | 0.0250             | 1.96       | < 0.025           |

Efficacy test statistic: Log-rank test comparing PFS between treatment arms
(stratified by [entry criteria]; two-sided test, then applied one-sidedly 
for comparison to boundary).

Decision rule:
  - If Z_interim ≥ 2.75 (p < 0.003) at interim: STOP FOR EFFICACY
    - DSMB recommends stopping to sponsor
    - Sponsor submits BLA/NDA with interim analysis data
    - Assume final analysis at higher p-value threshold (0.025) is 
      unachievable due to early stopping; no additional analysis planned

  - If Z_interim < 2.75 at interim: CONTINUE TO FINAL
    - No interim result is communicated to blinded study team
    - Enrollment and follow-up continue
    - Final analysis proceeds when 350 events accrued

7.4 Futility Stopping Rule (Non-Binding)

A pre-specified conditional power boundary will be evaluated at the interim 
analysis to assess whether the trial has a reasonable chance of success at 
the final analysis. If conditional power is low at interim, the DSMB may 
recommend consideration of stopping, but the sponsor retains the option to 
continue if new clinical or scientific information warrants.

Futility is assessed using conditional power with linear beta spending 
(β_total = 0.20 for 80% power):

| Analysis  | Info Fraction | Beta Spent at Stage | CP Threshold for Futility |
|-----------|---------------|--------------------|--------------------------|
| Interim 1 | 50%           | 0.10               | 0.80 (i.e., CP < 80%)     |
| Final     | 100%          | 0.20               | N/A                       |

Conditional Power calculation:
  CP = P(reject H_0 at final | observed data at interim)

  The observed hazard ratio and its variance at interim are used to project 
  the probability of achieving p < 0.025 (one-sided) at the final analysis 
  when 350 total events are reached.

  If the observed HR at interim is consistent with the target (HR = 0.65), 
  CP is expected to be > 95%. If CP < 80%, the futility boundary is crossed.

Futility decision rule (non-binding):
  - If CP ≥ 80% at interim: CONTINUE TO FINAL (expected case)

  - If CP < 80% at interim (futility boundary crossed):
    DSMB may recommend considering stopping for futility. However, sponsor 
    may choose to continue the trial based on:
      * Clinical context (e.g., safety or tolerability improvements)
      * Secondary endpoint trends (e.g., modest PFS effect with strong OS trend)
      * Unblinded efficacy boundary not yet crossed

    If sponsor continues despite futility recommendation, rationale must be 
    documented in trial file and communicated to FDA.

7.5 Safety Review (Non-Binding, Continuous)

Safety data will be reviewed at each interim analysis and continuously 
between analyses. Safety stopping rules are non-binding recommendations:

Pre-specified safety stopping rules:
  1. Any treatment-related death (grade 5 adverse event) unless clearly 
     unrelated to study drug; trigger: ≥2 deaths → emergency DSMB meeting

  2. Serious adverse events (grade 3+) with plausible causal relationship 
     to study drug:
     * Hepatotoxicity (grade 3+): > 5% in treatment arm vs. < 2% in control
     * Pneumonitis (grade 3+): > 3% in treatment arm vs. < 1% in control
     * Other [specified organ toxicity]: [threshold]

     Trigger: Safety signal threshold met → DSMB reviews causality and 
     recommends continuing, modifying protocol (e.g., dose reduction), or 
     stopping enrollment

Safety decision framework (DSMB):
  - CONTINUE: Observed safety profile consistent with prior data; 
    benefit-risk favorable
  - MODIFY: Protocol amendment (e.g., dose reduction, increased monitoring) 
    to mitigate signal; continue with safeguards
  - PAUSE: Temporarily halt enrollment pending additional safety data review
  - STOP: If signal meets pre-defined stopping criteria and is deemed 
    serious/unexpected; recommend trial termination

7.6 Statistical Test and Analysis Population

The primary efficacy analysis will use:
  - Test statistic: Log-rank test (two-sided) for PFS in the intent-to-treat (ITT) population
  - Stratification: By [enrollment variables, e.g., ECOG status, prior therapy status]
  - Multiplicity adjustment: Group sequential design (O'Brien-Fleming spending)

Interim and final p-values will be reported without rounding; they will be 
compared to the pre-specified boundaries above.

7.7 Timing of DSMB Meetings

  - Baseline DSMB meeting (before first enrollment): Charter review, 
    procedures, interim plan review

  - Interim DSMB meeting (scheduled ~2 weeks after 175 PFS events accrued): 
    Review unblinded efficacy, futility (CP), and safety data. Provide 
    stopping recommendation to sponsor.

  - Safety check meetings (every 6 months or as needed): If serious safety 
    signals emerge between planned interim and final, DSMB meets in-person 
    or via teleconference to assess (emergency meeting).

  - Final DSMB meeting (after final analysis completed): Confirm trial 
    success/failure and oversee unblinding to site investigators.

7.8 Interim Analysis Conduct and Statistical Methods

The interim analysis will be conducted by the Statistical Analysis Center 
(SAC) under the direction of the trial biostatistician, blinded from the 
interim efficacy data until analysis is complete.

Steps:
  1. SAC locks the interim database when 175 PFS events are confirmed.
  2. DSMB biostatistician (or delegated statistician) unblind treatment 
     assignments and performs the interim log-rank test and conditional 
     power calculation.
  3. Results are compared to the pre-specified efficacy and futility 
     boundaries (Section 7.3–7.4).
  4. DSMB meets in closed session to discuss results and make a stopping 
     recommendation (Section 7.5–7.6).
  5. DSMB communicates recommendation to sponsor in writing (without 
     disclosing specific p-values or HRs to blinded staff; only: 
     "Continue trial as planned" or "Recommend stopping for efficacy/futility").
  6. Sponsor either accepts recommendation or, in the case of futility, 
     documents justification for continuation.
  7. Interim analysis is not disclosed to site investigators or data 
     management staff unless trial is stopped for efficacy.

7.9 Reporting of Interim Results

  - If trial is stopped early for efficacy: Full interim results will be 
    submitted in the BLA/NDA (primary analysis of efficacy, secondary 
    endpoints, safety). All DSMB meeting minutes are included in the 
    Common Technical Document (CTD) Module 2.

  - If trial continues to final: Interim results are NOT disclosed to 
    blinded site staff. DSMB recommendation (stopping/continuing) may be 
    mentioned in regulatory submissions if relevant to trial integrity 
    (e.g., "DSMB reviewed interim efficacy/futility data and recommended 
    continuing trial as planned").

7.10 Type I Error Preservation and Power

The group sequential design preserves the overall one-sided Type I error rate 
at α = 0.025 despite the interim analysis. The final efficacy boundary 
(Z ≥ 1.96, p < 0.025) is adjusted for the interim look via alpha spending.

Power calculation:
  - Single-stage design power: 80% (β = 0.20) for HR = 0.65 at final analysis
  - Two-stage design (with interim) power: 79% (minor power loss due to 
    interim alpha spending)
  - Sample size: n = 400 per arm (800 total) to achieve 350 PFS events

---

Limitations and Pitfalls

1. Interim alpha spending not coordinated with multiplicity adjustment: If the trial has multiple primary endpoints (e.g., PFS and OS co-primary), each endpoint has its own alpha budget. Interim analyses for both must coordinate spending: total alpha at interim cannot exceed the sum of endpoint-specific budgets. Failure to account for this inflates Type I error.

2. Futility boundary set too tight: If conditional power threshold is set at > 95% CP, almost any modest treatment effect will pass futility, reducing efficiency gains. A 70–80% CP threshold is common in oncology.

3. Unequal information fractions at interim: If interim is planned at 50% events but actually occurs at 40% or 65%, alpha spending boundaries must be recalculated using Lan-DeMets adapters; using pre-calculated 50% boundaries is incorrect.

4. DSMB firewall breached: If interim results are communicated to blinded staff or site investigators (even unintentionally), Type I error protection is lost and trial credibility is compromised. Strict confidentiality protocols are essential.

5. Safety stopping rules too permissive: Defining safety thresholds post hoc (after seeing data) invalidates statistical interpretation. Pre-specify thresholds in SAP (e.g., "> 5% grade 3+ hepatotoxicity").

6. Interim analysis for OS before mature follow-up: Testing OS at interim with < 100 events and < 60% follow-up for censoring risks immature claims. FDA 2025 draft guidance recommends caution with premature interim OS analyses. Sensitivity analyses for non-proportional hazards should be included.

7. Non-binding futility ignored: If DSMB recommends stopping due to futility but sponsor continues without documented justification, FDA may scrutinize final results with skepticism, especially if trial eventually fails. Document rationale if overriding futility recommendation.

---

Backlinks

• Overall Survival (OS)
• Progression-Free Survival (PFS)
• Multiplicity Control in Oncology Trials
• ICH E9(R1) Estimand Framework
• Statistical Analysis Methods in Oncology Trials

---

Source: FDA Guidance for Industry — Adaptive Design Clinical Trials for Drugs and Biologics (November 2019, Final); ICH E20 Adaptive Designs for Clinical Trials (2025, Draft) Status: FDA Adaptive Designs = Final; ICH E20 = Draft Compiled from FDA/ICH guidance and literature on group sequential designs in oncology