Blog Post

Statistical Design Considerations for Dose-Finding Oncology Trials

February 2, 2023

Since the introduction of the original 3+3 method, the available design options for dose escalation in early phase studies have expanded beyond traditional algorithm-based designs (e.g., 3+3, rolling 6, accelerated titration) to include model-based (e.g., CRM, EWOC, BLRM) and model-assisted (e.g., mTPI, keyboard, BOIN) designs. The emergence of new designs has been motivated by the following needs:

  • Improve the operating characteristics of existing designs.
  • Increase efficiency in balancing the goals of finding the MTD while safely making decisions for individual patients.
  • Incorporate the flexibility needed to handle emerging therapies where toxicities may be delayed and the optimal dose rather than the MTD is of interest.

With many newly emerging design options, it can be difficult to select the best option for a dose-finding trial. Literature detailing pros and cons of various dose escalation designs is extensive and the choice of design depends on several factors. However, the following factors often drive the design decision: accuracy, flexibility, and complexity.

First, it is important to understand which designs have better accuracy in determining the MTD/RP2D. Data have shown that, compared to rule-based designs, model-based designs generally are more accurate in identifying the MTD and assign more patients to this dose level. Model-assisted designs have been shown to outperform the 3+3 but have comparable accuracy in identifying the MTD relative to model-based designs. This makes them attractive given that they combine the benefits of modeling with rules that can be implemented similarly to 3+3 designs. The newer designs also tend to assign fewer patients to subtherapeutic doses and are safer than rule-based designs in not assigning patients to overly toxic doses.

Second, one must consider flexibility in design needed to address the mechanism and dose-toxicity profile of the agent of interest. For example, traditional rule-based designs were built based on the assumption of monotonically increasing toxicity (and efficacy) with dose. This assumption was well suited for cytotoxic agents; however, molecularly targeted agents and more recently immunotherapeutic agents may not conform to this assumption. Model-based and model-assisted designs allow for greater flexibility to manage cases where interest primarily lies in the optimal biologic dose and delayed toxicities may be more common. More recent designs can also efficiently manage dose-finding for combination regimens.

Third, one must consider statistical complexity and challenges of executing these designs properly. One of the main reasons for the longstanding popularity of the 3+3 design is simplicity. Indeed, the lack of statistical expertise to support model-based and model-assisted designs has contributed to their limited use. However, as noted, more recent designs (like the BOIN) can offer increased efficiency and flexibility in identifying the MTD/RP2D while retaining the ease of implementation of rule-based designs. In fact, the difference between model-based and model-assisted designs is that rules can often be pre-determined then implemented proactively similar to rule-based designs. Software to support these designs are now more readily available making it easier to generate side-to-side comparisons to inform choice of design.

Nuances of selecting the most appropriate model for a dose finding study can introduce challenges when designing a clinical trial. The thoughtful selection of a CRO that is familiar with the benefits and challenges of potential options can be a critical aspect of ensuring the success of early phase oncology trial execution.

patricia stephensonPatricia Stephenson, Sc.D., Associate Director, Biostatistics is a Harvard graduate with over 10 years of experience working in oncology, including working with researchers at the Dana-Farber Cancer Institute. Dr. Stephenson has served as the lead statistician for multiple oncology studies, including a Phase 1 study pivotal for an NDA submission for accelerated approval. As a result, she was involved in several submission activities including supporting the statistical preparation and review for the Summary of Clinical Efficacy (SCE) and Summary of Clinical Safety (SCS) as well as preparation for a FDA Advisory Committee Meeting.  Previously, Dr. Stephenson also served as the lead statistician for multiple Phase 1 and 2 studies in ovarian cancer, renal cell carcinoma, gastrointestinal stromal tumors, and non-small cell lung cancer.