Bayesian Evaluation of Informative (Inequality-Constrained) Hypotheses

Description

Compares competing hypotheses stated as orderings of means or effects, such as C > B > A, using Bayes factors from the share of prior and posterior satisfying each ordering.

Researchers often hold ordered expectations, but null hypothesis tests only ask whether any difference exists. Informative hypothesis testing states each expectation as inequality constraints (for example mu_C > mu_B > mu_A) and compares them directly. The marginal likelihood acts as an automatic Occam's razor: an order-constrained model that agrees with the data beats the unconstrained one because it occupies less of the parameter space. Vague priors are unsafe for Bayes factors (the Bartlett-Lindley paradox), so the encompassing approach specifies only one prior, for the unconstrained model, with the same prior on every mean so no ordering is favoured; each constrained model's prior is its truncation. The Bayes factor of a constrained hypothesis against the encompassing model is then the proportion of posterior draws satisfying its constraints (fit) divided by the proportion of prior draws satisfying them (complexity). Posterior model probabilities follow when hypotheses are given equal prior probability.

Attributes

Loss type
Unmigrated
Stage
Evaluation
Effort
Days

Steps

  1. State the competing informative hypotheses as inequality constraints among parameters, plus an unconstrained or null model for reference.
  2. Specify one encompassing prior for the unconstrained model, identical for every mean so that no ordering is favoured a priori.
  3. Sample the encompassing prior and posterior (by MCMC or with software such as bain or BIEMS).
  4. For each hypothesis compute the proportion of prior draws (complexity) and posterior draws (fit) that satisfy its constraints.
  5. Form Bayes factors (fit divided by complexity) and posterior model probabilities with equal prior model probabilities.
  6. Report the best-supported ordering, its posterior probability and the probability of being wrong, and check sensitivity to the prior.

Sources

  • The SAGE Handbook of Quantitative Methods in Psychology (Millsap & Maydeu-Olivares, 2009)

Also known as

  • Informative hypothesis testing
  • Order-constrained Bayes factors
  • Encompassing prior approach
  • Bayesian model selection for ordered means
  • Bayesian evaluation of informative (inequality-constrained) hypotheses with encompassing priors

Categories

  • Analytics & Behaviour
  • Statistics & Modelling