Internally, adoptr is built around the
joint distribution of a test statistic and the unknown location
parameter of interest given a sample size, i.e. ℒ[(Xi, θ) | ni]
where Xi
is the stage-i test statistic
and ni the
corresponding sample size. The distribution class for Xi is defined by
specifying a DataDistribution object, e.g., a normal
distribution
To completely specify the marginal distribution of Xi, the distribution of θ must also be specified. The classical case where θ is considered fixed, emerges as special case when a single parameter value has probability mass 1.
The simplest supported prior class are discrete
PointMassPrior priors. To specify a discrete prior, one
simply specifies the vector of pivot points with positive mass and the
vector of corresponding probability masses. E.g., consider an example
where the point δ = 0.1 has
probability mass 0.4 and the point
δ = 0.25 has mass 1 − 0.4 = 0.6.
For details on the provided methods, see
?DiscretePrior.
adoptr also supports arbitrary continuous priors with support on compact intervals. For instance, we could consider a prior based on a truncated normal via:
cont_prior <- ContinuousPrior(
pdf = function(x) dnorm(x, mean = 0.3, sd = 0.2),
support = c(-2, 3)
)For details on the provided methods, see
?ContinuousPrior.
In practice, the most important operation will be conditioning. This is important to implement type one and type two error rate constraints. Consider, e.g., the case of power. Typically, a power constraint is imposed on a single point in the alternative, e.g. using the constraint
If uncertainty about the true response rate should be incorporated in the design, it makes sense to assume a continuous prior on θ. In this case, the prior should be conditioned for the power constraint to avoid integrating over the null hypothesis: