A General Framework for Updating Belief Distributions

Abstract

SummaryWe propose a framework for general Bayesian inference. We argue that a valid update of a prior belief distribution to a posterior can be made for parameters which are connected to observations through a loss function rather than the traditional likelihood function, which is recovered as a special case. Modern application areas make it increasingly challenging for Bayesians to attempt to model the true data‐generating mechanism. For instance, when the object of interest is low dimensional, such as a mean or median, it is cumbersome to have to achieve this via a complete model for the whole data distribution. More importantly, there are settings where the parameter of interest does not directly index a family of density functions and thus the Bayesian approach to learning about such parameters is currently regarded as problematic. Our framework uses loss functions to connect information in the data to functionals of interest. The updating of beliefs then follows from a decision theoretic approach involving cumulative loss functions. Importantly, the procedure coincides with Bayesian updating when a true likelihood is known yet provides coherent subjective inference in much more general settings. Connections to other inference frameworks are highlighted.

Pseudo-Bayesian quantum tomography with rank-adaptation

“…Using the same proof of Lemma 4 for inequality (8) in Lemma 3, we obtain with probability at least 1 − /2, ∈ (0, 1), for any distributionπ that…”

“…Here, for computational reasons, we replace the likelihood by a pseudo-likelihood. This is an increasingly popular method in Bayesian statistics [8] and in machine learning [16,2,7]. We define the pseudo-posterior bỹ…”

Bayesian inference via projections

“…Using an off-the-shelf method, poor mixing may follow, particularly if latent variables are sampled explicitly [16]. Alternatives include estimating the likelihood function within a MCMC step [1]; using only summary statistics implied by draws from the model [4]; using surrogate likelihood functions [6,25]. While much of the motivation for the latter is to avoid a complete specification of a model, which calls for explicit assumptions on nuisance parameters, computational considerations are also invoked.…”

“…We generate synthetic models with 10 latent variables and three variables per cluster, and 5 ordinal levels for each observed variable (a total of 30 observed variables) 6 .…”

A Definition of Conditional Probability with Non-Stochastic Information

“…In this paper, instead, the definition of conditional probability is solely based on a set of axioms. This idea was developed by reference [2] who defined a framework for general Bayesian inference and discussed its application to important statistical problems. The present paper also addresses the issue of calibrating the conditional probability with non-stochastic information.…”