A New Bayesian Test to test for the Intractability-Countering Hypothesis

TL;DR

This paper introduces a novel Bayesian hypothesis test designed to evaluate the intractability-countering null hypothesis by comparing marginalised posteriors, avoiding traditional Bayes factor limitations, and employing MCMC methods.

Contribution

The paper proposes a new Bayesian test that circumvents Bayes factor computation by using generated data and marginalised posteriors, improving inference for intractable models.

Findings

Supports hypothesis testing without Bayes factors

Uses generated data to assess null hypothesis support

Employs MCMC for posterior marginalisation

Abstract

We present a new test of hypothesis in which we seek the probability of the null conditioned on the data, where the null is a simplification undertaken to counter the intractability of the more complex model, that the simpler null model is nested within. With the more complex model rendered intractable, the null model uses a simplifying assumption that capacitates the learning of an unknown parameter vector given the data. Bayes factors are shown to be known only up to a ratio of unknown data-dependent constants--a problem that cannot be cured using prescriptions similar to those suggested to solve the problem caused to Bayes factor computation, by non-informative priors. Thus, a new test is needed in which we can circumvent Bayes factor computation. In this test, we undertake generation of data from the model in which the null hypothesis is true and can achieve support in the measured data for the null by comparing the marginalised posterior of the model parameter given the measured data, to that given such generated data. However, such a ratio of marginalised posteriors can confound interpretation of comparison of support in one measured data for a null, with that in another data set for a different null. Given an application in which such comparison is undertaken, we alternatively define support in a measured data set for a null by identifying the model parameters that are less consistent with the measured data than is minimally possible given the generated data, and realising that the higher the number of such parameter values, less is the support in the measured data for the null. Then, the probability of the null conditional on the data is given within an MCMC-based scheme, by marginalising the posterior given the measured data, over parameter values that are as, or more consistent with the measured data, than with the generated data.