Bias induced by adaptive dose-finding designs

Area of Science:

• Biostatistics
• Clinical Trial Design
• Pharmacometrics

Background:

• Maximum likelihood estimators are known to exhibit bias in statistical literature.
• Adaptive dose-finding designs are increasingly used in clinical trials for efficiency.

Purpose of the Study:

• To demonstrate that adaptive dose-finding procedures inherently induce bias.
• To provide an explicit formula for the bias of observed response rates in such designs.
• To investigate the implications of this bias and propose mitigation strategies.

Main Methods:

• Analysis of adaptive dose-finding procedures with Bernoulli responses.
• Derivation of an explicit mathematical formula for bias in observed response rates.
• Simulation and illustration of bias patterns for designs targeting a specific quantile.

Main Results:

• Adaptive dose-finding procedures (Continual Reassessment Methods, Up-and-Down, Interval Designs) were shown to induce bias.
• Bias tends to be positive above the target dose and negative below it for designs concentrating dose allocations.
• This bias property in dose-finding designs appears to be a novel finding.

Conclusions:

• Adaptive dose-finding designs introduce a previously unrecognized bias in response rate estimation.
• Understanding and quantifying this bias is crucial for accurate dose selection in clinical trials.
• A simple shrinkage formula is proposed to mitigate bias and improve estimation accuracy, especially away from the target dose.