Combine multiple spatial priors via weighted geometric mean

Given a list of priors (each a data.table with rsg_code, probability), pools them into a single distribution using a weighted geometric mean and renormalises to sum to 1.

Usage

combine_priors(priors, weights = NULL, epsilon = 1e-06)

Arguments

priors

A list of data.tables with columns rsg_code and probability.

weights

Optional non-negative numeric vector of length length(priors). Defaults to equal weights. Will be renormalised to sum to 1.

epsilon

Smoothing floor for classes missing from a prior (default 1e-6). Must be > 0 – otherwise any class missing from a single prior is suppressed entirely.

Value

A data.table with columns rsg_code, probability, sorted by descending probability.

Details

Geometric pooling has two desirable properties for soil-class priors:

1. externally Bayesian (the pooled posterior under any common likelihood matches what one would get by individual updates), and

2. zero-preserving: a class assigned probability 0 by any prior is suppressed in the pooled distribution. To avoid that, classes absent from a given prior are imputed with the smoothing constant epsilon.