lorad_estimate {lorad}

Calculates the LoRaD estimate of the marginal likelihood

Description

Provided with a data frame containing sampled paraneter vectors and a dictionary relating column names to parameter types, returns a named character vector containing the following quantities:

• logML (the estimated log marginal likelihood)

• nsamples (number of samples)

• nparams (length of each parameter vector)

• training_frac (fraction of samples used for training)

• tsamples (number of samples used for training)

• esamples (number of sampled used for etimation)

• coverage (nominal fraction of the estimation sampled used)

• esamplesused (number of estimation samples actually used for estimation)

• realized_coverage (actual fraction of estimation sample used)

• rmax (lowest radial distance: defines boundary of working parameter space)

• log_delta (volume under the unnormalized posterior inside working parameter space)

Usage

lorad_estimate(params, colspec, training_frac, training_mode, coverage)

Arguments

params

Data frame in which rows are sample points and columns are parameters, except that last column holds the log posterior kernel

colspec

Named character vector associating column names in params with column specifications

training_frac

Number between 0 and 1 specifying the training fraction

training_mode

One of random, left, or right, specifying how training fraction is chosen

coverage

Number between 0 and 1 specifying fraction of training sample used to compute working parameter space

Value

Named character vector of length 11.

Examples

normals <- rnorm(1000000,0,10)
prob_normals <- dnorm(normals,0,10,log=TRUE) 
proportions <- rbeta(1000000,1,2)
prob_proportions <- dbeta(proportions,1,2,log=TRUE)
lengths <- rgamma(1000000, 10, 1)
prob_lengths <- dgamma(lengths,10,1,log=TRUE)
paramsdf <- data.frame(
    normals,prob_normals,
    proportions, prob_proportions,
    lengths, prob_lengths)
columnkey <- c(
    "normals"="unconstrained", 
    "prob_normals"="posterior", 
    "proportions"="proportion", 
    "prob_proportions"="posterior", 
    "lengths"="positive", 
    "prob_lengths"="posterior")
results <- lorad_estimate(paramsdf, columnkey, 0.5, 'random', 0.1)
lorad_summary(results)