Adaptive dose-finding designs induce a bias on observed rates,
away from the target dose. This is well-known in other adaptive-design fields,
but has been overlooked by the dose-finding research community.
Flournoy and Oron (2020) examine the bias in the dose-finding context,
and suggest a simple shrinkage fix that reduces both bias and variance.
The fix is analogous to the empirical-logit fix to binary data,
but instead of adding 0.5 to each cell, target is added to the 1's at each dose,
and 1-target to the 0's.
The shrinkage is applied to the raw observation, so CIR or IR are carried out
on the shrunk data.
DRshrink(y, x = NULL, wt0 = NULL, target, swt = 1, nmin = 2, ...)can be either of the following: y values (response rates), a 2-column matrix with positive/negative response counts by dose, a DRtrace object or a doseResponse object.
dose levels (if not included in y).
weights (if not included in y).
the balance point (between 0 and 1) around which the design concentrates allocations.
the weight of the shrinkage. Default 1 (a single observation)
the minimum n at each dose, for the shrinkage to be applied. Default 1 (all doses with any observation).
parameters passed on to doseResponse()
Flournoy N and Oron AP, 2020. Bias Induced by Adaptive Dose-Finding Designs. Journal of Applied Statistics 47, 2431-2442.