optimal range for safety and pharmacokinetics (PK). In small cohorts—often 3 to 6 subjects per group—traditional statistical power does not apply. Instead, developers use adaptive designs, Bayesian modeling, and real-time decision algorithms to ensure efficient dose escalation with minimal risk. This article outlines the key statistical frameworks used in small-cohort Phase 1 studies and how they influence decision-making, stopping rules, and maximum tolerated dose (MTD) estimation.

Why Small Cohorts Require Special Attention

• Limited data per dose level makes conclusions less precise
• Outlier effects can skew interpretation
• Sequential nature of dose escalation introduces time-dependency bias

Common Statistical Models in Dose Escalation

1. 3+3 Rule-Based Design

• Simplest method: 3 subjects per cohort
• If 0/3 experience dose-limiting toxicity (DLT), escalate
• If 1/3 experience DLT, add 3 more subjects (3+3)
• If 2+ DLTs, de-escalate or stop
• Limitations: Rigid, statistically inefficient, may underestimate MTD

2. Continual Reassessment Method (CRM)

• Bayesian model that updates probability of DLT at each dose level
• Each cohort’s data updates the likelihood curve
• More precise MTD estimation with fewer subjects
• Requires prior dose-toxicity assumptions

3. Bayesian Optimal Interval (BOIN) Design

• Improves on 3+3 by using probability intervals
• Simple algorithm: if DLT rate falls within a predefined range, stay at dose
• Outside the range? Escalate or de-escalate accordingly
• Efficient for monoclonal antibodies, ADCs, and immunotherapies

4. Modified Toxicity Probability Interval (mTPI)

• Uses beta-binomial models to guide dose decisions
• Acceptable, unacceptable, and excessive toxicity zones pre-defined
• Used in oncology and cell therapy programs

Key Parameters and Definitions

• DLT (Dose-Limiting Toxicity): Serious AE occurring within predefined time window
• MTD (Maximum Tolerated Dose): Highest dose with ≤33% DLT incidence
• RDE (Recommended Dose for Expansion): May be below MTD if efficacy plateau observed

Choosing the Right Model

• Use 3+3 if the molecule has unknown or high-risk toxicity
• Use CRM or BOIN for more efficient designs with early data
• Use mTPI when preclinical data supports multiple dose steps

Stopping Rules and Cohort Decisions

• Predefine when to stop for:
  • Excessive toxicity
  • Pharmacokinetic plateau
  • Lack of PD effect
• Use simulations during protocol design to estimate escalation probabilities

Global Regulatory Considerations

FDA

• Supports model-based approaches (CRM, BOIN) with justification
• Recommends DLT window of 21–28 days for most oncology studies

EMA

• Encourages adaptive escalation designs
• Requires simulation plan and model operating characteristics

CDSCO

• Prefers 3+3 or BOIN design for Indian FIH studies
• Mandates explicit stopping rules in protocol

Best Practices

• Use simulations to evaluate design performance before trial start
• Involve a biostatistician with experience in Bayesian methods
• Incorporate real-time safety review and decision-making tools
• Document escalation logic in the protocol and safety charter
• Maintain flexibility to pause, add cohorts, or redefine RDE if needed