downSample: Down sample the observations in a mixture

In CNPBayes: Bayesian mixture models for copy number polymorphisms

Description

For large datasets (several thousand subjects), the computational burden for fitting Bayesian mixture models can be high. Downsampling can reduce the computational burden with little effect on inference. This function draws a random sample with replacement. Batches with few observations are combined with larger batches that have a similar median log R ratio.

Usage

downSample(dat, size = 1000, min.batchsize = 75)

Arguments

dat	data.frame with required columns medians, batches, and plate
size	the number of observations to sample with replacement
min.batchsize	the smallest number of observations allowed in a batch. Batches smaller than this size will be combined with other batches

Value

A tibble of the downsampled data (medians), the original batches, and the updated batches after downsampling

Examples

## TODO: this is more complicated than it needs to be
library(dplyr)
mb <- MultiBatchModelExample
mapping <- tibble(plate=letters[1:10],
                  batch_orig=sample(c("1", "2", "3"), 10, replace=TRUE))
full.data <- tibble(medians=y(mb),
                    batch_orig=as.character(batch(mb))) %>%
  left_join(mapping, by="batch_orig")
partial.data <- downSample(full.data, 200)
## map the original batches to the batches after down-sampling
mapping <- partial.data %>%
  select(c(plate, batch_index)) %>%
  group_by(plate) %>%
  summarize(batch_index=unique(batch_index))
mp <- McmcParams(iter=50, burnin=100)
## Not run: 
    mb2 <- MultiBatchModel2(dat=ds$medians,
                            batches=ds$batch_index, mp=mp)
    mb2 <- posteriorSimulation(mb2)
    if(FALSE) ggMixture(mb2)
    full.dat2 <- full.data %>%
      left_join(mapping, by="plate")
    ## compute probabilities for the full dataset
    mb.up <- upSample2(full.dat2, mb2)
    if(FALSE) ggMixture(mb2)

## End(Not run)

CNPBayes documentation built on May 6, 2019, 4:06 a.m.