R: Predict posterior expectation, median, variance and quantiles...

predict.regAbcrf {abcrf}

R Documentation

Predict posterior expectation, median, variance and quantiles given a new dataset using a reg-ABC-RF object

Description

Based on a reg-ABC-RF object this function predicts the posterior expectation, median, variance, quantiles for the corresponding parameter given new dataset. Somes posterior errors can be computed at an higher computational price.

Usage

## S3 method for class 'regAbcrf'
predict(object, obs, training, quantiles=c(0.025,0.975),
paral = FALSE, ncores = if(paral) max(detectCores()-1,1) else 1, rf.weights = FALSE,
post.err.med = FALSE, ...)

Arguments

`object`	a `regAbcrf` object.
`obs`	a data frame containing the summary statistics of the observed data sets.
`training`	the data frame containing the reference table used to train the `regAbcrf` object.
`quantiles`	numeric vector of probabilities with values in [0,1]. The default value is equal to `c(0.025, 0.975)`.
`paral`	a boolean that indicates if random forests predictions should be parallelized.
`ncores`	the number of CPU cores to use for the regression random forest predictions. If paral=TRUE, it is used the number of CPU cores minus 1. If ncores is not specified and `detectCores` does not detect the number of CPU cores with success then 1 core is used.
`rf.weights`	a boolean that indicates if the random forest weights used to predict quantities of interest should we returned. The default value is FALSE.
`post.err.med`	a boolean that indicates if posterior errors based on posterior medians should be computed. The default value is FALSE. If computed, this function might take a much more time.
`...`	optional arguments to be passed on to the function `predict.ranger`.

Value

An object of class regAbcrfpredict, which is a list with the following components:

`expectation`	predicted posterior expectation for each oberved data set,
`med`	predicted posterior median for each oberved data set,
`variance`	predicted posterior variance for each observed data set, computed by reusing weights, this quantity is also the posterior mean squared error,
`variance.cdf`	predicted posterior variance for each observed data set, computed by approximation of the cumulative distribution function,
`quantiles`	predicted posterior quantiles for each observed data set,
`weights`	a matrix composed of the weights used to predict quantities of interest. Returned if `rf.weights` is TRUE,
`post.NMAE.mean`	posterior normalized mean absolute error obtained using the out-of-bag posterior expectation (mean) and previously computed random forest weights, for each observed data set,
`post.MSE.med`	posterior mean squared error obtained using the out-of-bag posterior median and previously computed random forest weights, for each observed data set,
`post.NMAE.med`	posterior normalized mean absolute error obtained using the out-of-bag posterior expectation (mean) and previously computed random forest weights, for each observed data set,
`prior.MSE`	prior mean squared error computed with training out-of-bag prediction based on mean of response variable,
`prior.NMAE`	prior normalized mean absolute error computed with training out-of-bag predictions based on mean of response variable,
`prior.MSE.med`	prior mean squared error computed with training out-of-bag predictions based on median of response variable,
`prior.NMAE.med`	prior normalized mean absolute error with training out-of-bag predictions based on median of response variable,
`prior.coverage`	prior credible inteval coverage computed for training instances, if only two quantiles are of interest, NULL otherwise.

References

Raynal L., Marin J.-M. Pudlo P., Ribatet M., Robert C. P. and Estoup, A. (2019) ABC random forests for Bayesian parameter inference Bioinformatics doi:10.1093/bioinformatics/bty867

Examples

data(snp)
modindex <- snp$modindex
sumsta <- snp$sumsta[modindex == "3",]
r <- snp$param$r[modindex == "3"]
r <- r[1:500]
sumsta <- sumsta[1:500,]
data2 <- data.frame(r, sumsta)
model.rf.r <- regAbcrf(r~., data2, ntree=100)
data(snp.obs)
predict(model.rf.r, snp.obs, data2)

[Package abcrf version 1.9 Index]