A Bounded Derivative Model for Prior Ignorance about a Real‐valued Parameter (Q4376536)

The paper describes a method for drawing statistical inferences about a real-valued parameter \(\theta \in \Theta\), where \(\Theta\) is an interval (possibly infinite) of real numbers, using the set \({\mathcal M}\) of all probability density functions \(f\) satisfying the following conditions: (i) \(f\) is continuous everywhere in \(\Theta\); (ii) \(f(\theta)>0\) for all \(\theta \in {\Theta}\); (iii) \(f\) is differentiable and \(| f'(\theta)| \leq cf(\theta)\), for almost all \(\theta \in \Theta\). The set \({\mathcal M}\) depends on the positive constant \(c\) and as \(c\) gets larger one obtains a nested sequence of larger and larger sets \({\mathcal M}\). As \(c \to \infty\) the inferences using the bounded derivative model become vacuous. Detailed results are given which enable one to calculate the posterior upper and lower previsions (expectations) for two special types of functions -- monotonic functions and indicator functions of finite intervals. For arbitrary functions \(Y(\theta)\), results from optimal control theory are used to characterize the prior densitites \(f\) that achieve \(\overline P(Y | x)\) and \(\underline P(Y | x)\). It is shown that, in general, posterior upper and lower expectations are achieved by prior density functions that are piecewise exponential. Normal and binomial examples are used to illustrate the techniques.