Bayesian Analysis Choice of Priors for Low-dimensional Parameters推荐.pdfVIP

Choice of Prior s for Low-dimensiona l Parameter s Given data , a Bayesian will need a likelihood function p{x\0) and a prior p{0). For many standar d problems , th e likelihood is known either from past experience or by convention . To drive th e Bayesian engine, one would still need an appropriate prior . In thi s chapter , we consider only low-dimensional parameters . Admittedly , low dimension is not easy t o define, but we expect the dimension d t o be much smaller than th e sample size n t o qualify as low. In most of th e examples in thi s chapter , d = 1 or 2 and is rarely bigger than 5. Ideally, one want s t o choose a prior or a class of priors reflecting ones prior knowledge and belief about th e unknown parameter s or about different hypotheses . Thi s is a subjective choice . If one has a class of priors , it would be necessary t o study robustnes s of various aspect s of th e resulting Bayesian inference. Choice of subjective priors , usually called elicitation of priors , is still rather difficult. For some systematic method s of elicitation , see Kadan e et al . (1980), Garthwaite and Dickey (1988, 1992). A recent review is Garthwaite et al . (2005) . Empirical studies have shown experience and maturity help a person in quantifying uncertainty about an event in th e form of a probability . However, assigning a fully specified probability distribution t o an unknown parameter is difficult eve