Description
Bayesian predictor selection in which each term carries a binary inclusion indicator sampled by MCMC, giving inclusion probabilities for predictors and for predictor combinations.
Stepwise selection picks one model and reports its coefficients as if no choice had been made. Stochastic search attaches an inclusion indicator to each candidate predictor and samples the indicators along with the coefficients. In the Kuo-Mallick form the coefficient is multiplied by its indicator; in George and McCulloch's form the indicator switches the coefficient between a narrow spike prior and a wide slab prior, the spike set from the largest effect regarded as negligible and the slab between about 10 and 100 times wider. The share of iterations in which each indicator is 1 is that predictor's posterior inclusion probability, and the frequency of each pattern estimates model probabilities. Congdon extends the device to random-effect variances and to choosing autoregressive lags; dependent priors can keep interactions only alongside their main effects.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Goal
- Prioritise, Understand why
- Stage
- Discovery, Evaluation
- Data
- Quantitative
- Effort
- Days
Steps
- Centre or standardise the candidate predictors and list any constraints, such as interactions only with their main effects.
- Give each coefficient an inclusion indicator with a prior probability, typically 0.5.
- Choose the form: multiply coefficients by indicators (Kuo-Mallick), or set a spike variance from the largest negligible effect and a slab 10 to 100 times wider (George-McCulloch).
- Run MCMC, recording every indicator at every iteration, and check mixing.
- Report each predictor's posterior inclusion probability and the most frequent predictor combinations with their frequencies.
- Refit the leading models and repeat with other prior inclusion probabilities and spike widths to check sensitivity.
Sources
- Applied Bayesian Modelling (Congdon, 2003)
- Bayesian Statistical Modelling (Congdon, 2006)
Also known as
- SSVS
- Bayesian stochastic search variable selection
- Spike-and-slab variable selection
- Kuo-Mallick indicator selection
- Posterior inclusion probabilities
- Variance component selection
Categories
- Analytics & Behaviour
- Statistics & Modelling