Scholion.ai

A Bayesian Test of Some Classical Hypotheses—With Applications to Sequential Clinical Trials

Jerome Cornfield

Journal of the American Statistical Association · 1966 · Volume 61 · Issue 315

DOI: 10.1080/01621459.1966.10480890

Abstract

Consider someone who sets out to collect data sequentially in such a way as to disprove a hypothesis, H 0 , about the value of θ, the mean of a normal distribution. Observation is continued as long as the posterior probability of H 0 exceeds α 1 and stopped when it falls below it. It is shown that if a non-zero prior probability, p , is assigned to H 0 (0≤α 1 < p ) and the remaining prior probability is spread over alternate values of θ that the probability that H 0 will eventually be rejected when true is With non-zero assignment of p it is therefore not possible to sample to a foregone conclusion. The Wald 3-decision scheme is shown to be equivalent to setting and assigning prior probability of to the alternatives θ = θ 0 + Δ and θ = θ 0 − Δ. Anomalous features of this scheme, particularly for the case of unknown σ, are shown to be a consequence of the unrealistic nature of the alternatives assumed. It is suggested that meaningful assignments of p are possible in practice by considering the minimum number of cases that would lead one to accept H 0 under the most favorable possible evidence for it. Rejection of H 0 is shown to be insensitive to choice of p. Closed schemes are obtained by sampling as long as P (θ ≤ Δ) > α 1 and P (θ > 0) < α 2 and stopping when either inequality is infringed, where Δ is the smallest positive departure from H 0 of substantive interest and α 1 < α 2 . Approximating the likelihood function for the logit of a binomial parameter with a normal likelihood function it is possible to extend these methods to the comparison of two binomial parameters, without the usual sequential assumption of pairing and equal number of cases.

Enable JavaScript for citation tools and additional paper interactions.