Description
Weight several plausible models by their posterior probabilities and average their estimates and predictions, so conclusions carry model uncertainty rather than resting on one chosen model.
Picking a single best model and then reporting its estimates understates uncertainty, because the choice itself was uncertain. Bayesian model averaging assigns each candidate model a prior probability, converts marginal likelihoods into posterior model probabilities and averages parameters and predictions with those weights. Marginal likelihoods can be estimated model by model (Laplace or BIC approximations, importance sampling, Chib's method, bridge sampling) or by sampling a model indicator alongside the parameters (Carlin and Chib's product space, reversible jump MCMC). Lighter approximations use AIC or BIC weights. Congdon stresses that harmonic-mean estimates are unstable and that diffuse priors on extra parameters push the weights towards simpler models, so the prior for each parameter needs as much care as the model list.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Strategy, Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- List the candidate models and give them prior probabilities, often equal.
- Estimate each model's marginal likelihood, or sample a model indicator jointly with the parameters by product-space or reversible-jump MCMC.
- Convert to posterior model probabilities and report them.
- Average parameter estimates and predictions across models with these weights, keeping the between-model spread in the intervals.
- Check stability of the marginal likelihood estimates across batches of iterations and their sensitivity to the parameter priors.
- With many candidate predictors, switch to inclusion indicators (stochastic search variable selection) instead of enumerating models.
Sources
- Bayesian Statistical Modelling (Congdon, 2006)
Also known as
- BMA
- Model averaging
- Posterior model probabilities
- AIC weights
Categories
- Analytics & Behaviour
- Statistics & Modelling