From: Practical recommendations for implementing a Bayesian adaptive phase I design during a pandemic
Parameters | Motivation and method for choosing | Considerations | Chosen Value |
---|---|---|---|
Target (additional) toxicity rate, γ | Motivated by safety considerations. Clinical decision | The DLT definition should be consciously and extensively specified, with acceptable toxicity determined by the clinical context (e.g., availability of other treatments, severity of disease) | 20% |
Tolerance around target (additional) toxicity rate, δ | Motivated by safety considerations. Clinical decision | Any given set of doses is unlikely to contain a dose with DLT rate exactly equal to the target, so flexibility around the target should be considered. As above, this should consider the clinical context | 5% |
Upper (additional) toxicity bound, γtoxic | Motivated by safety considerations. Clinical decision | Level of unacceptable toxicity (here, toxicity above the control) | 30% |
Number of Doses, m | Motivated by knowledge of treatment. Clinical decision | Number of doses should ensure that the dose-toxicity curve is adequately explored | 4 (300 mg bd, 400 mg bd, 600 mg bd, 800 mg bd) |
Cohort size | Safety and practical considerations. Options can be evaluated using simulations based on discussion with clinicians | The number of participants the clinicians are comfortable dosing between decisions on dose escalation; how often the model will be updated. Results for various cohort sizes can be shown to clinicians | 6 (4 on treatment, 2 control) |
Sample Size, N | Practical considerations. Options can be evaluated using simulations based on discussion with clinicians | Sample size should ensure an accurate selection of target doses with high probability | 30 |
Threshold controlling overdosing, coverdose | Simulations; reference in the literature | The value from the literature for the 2-parameter logistic dose-toxicity model was chosen (to speed up simulations) | 25% |
Hyperparameter μ1 – the mean of the prior distribution for θ1 | Historical information | Modelled as random variable to account for the uncertainty early in the pandemic | μ1 = logit(0.1) (fixed by 10% DLT rate on control) |
Hyperparameters μ2, σ1, σ2 – mean of prior for θ2 and standard deviations of priors for θ1 and θ2 | Simulations | Calibrated over a set of feasible dose-toxicity scenarios | μ2 = − 0.05, σ1 = 1.10, σ2 = 0.30 |
Prior estimates of DLT risk on each dose (also known as the dose-toxicity skeleton), \({p}_j^{(0)}\) | May solely reflect existing knowledge of treatment doses, or may be evaluated via simulations | If determined by simulation, the DLT risks on each dose should still align with existing knowledge | 17.5, 25, 32.5, 40% |