Robust mixture prior and borrowing from an existing drug
A robust mixture prior combines an informative component from existing data with a vague component, weighted by a borrowing weight. It is meant to reduce borrowing when new data conflict, but with small samples it may hardly do so.
Definition
In Bayesian trials, data from an earlier study (here, an existing drug) can be borrowed as a prior. A robust mixture prior is a weighted mixture of
- an informative component based on the existing data, with weight ω, and
- a vague component, with weight 1 − ω.
ω is the borrowing weight: 0 means no borrowing and 1 means full borrowing. When the new data disagree with the existing data, the posterior weight of the informative component should fall, so the existing data influence the result less (Schmidli et al., 2014).
What the site's analysis found
The analysis of borrowing an existing drug's dose comparison used published comparisons of sotorasib and ifinatamab deruxtecan as priors for a hypothetical new drug with 40 patients per arm.
- If the new drug behaved like the existing drug, borrowing improved correct dose selection by 6 to 10 points, but expected regret by only about 1 point.
- If the new drug favored the opposite dose, full borrowing lowered correct selection to 28 to 42%.
- With 40 patients per arm the conflict was hardly detected, so the mixture prior did not protect automatically. A borrowing weight of 0.25 to 0.5 kept the loss within 1 point.
The existing drug's sotorasib dose comparison is one of the two inputs.
Frequently asked questions
What does the borrowing weight ω mean?
It is the weight of the informative (existing-data) component in the mixture prior: 0 means no borrowing and 1 means full borrowing of the existing drug's data.
Does a robust mixture prior protect against a new drug that behaves differently?
It is designed to, by lowering the weight of the existing-data component when the data conflict. In the simulation with 40 patients per arm, a difference of about 15 points was hardly detected, so it cannot be assumed to protect automatically.