empb_beta_negbinomial_c: empb_beta_negbinomial_c

In alexandercoulter/DHBayes: Bayesian Hierarchical Sampler for Discrete (Binomial, Negative-Binomial, Poisson) Data

Description

empb_beta_negbinomial_c

Usage

empb_beta_negbinomial_c(
  df,
  eta = 0.1,
  tol = 1e-08,
  maxIter = 10000,
  starting_rab = NULL,
  method = c("newton", "gdescent")
)

Arguments

df	data.frame object, containing at least columns 'x' containing non-negative integer values, and 'g' containing group labels.
eta	positive numeric dampening parameter for Newton's method, gradient descent algorithm.
tol	non-negative numeric tolerance parameter for exiting optimization algorithm.
maxIter	positive integer setting maximum number of iterations for optimization algorithm.
starting_rab	optional 3-long numeric vector, giving initial algorithm starting point for fitting empirical Bayes estimates for r, a, and b, respectively; default NULL.
method	string controlling optimization method; default 'newton'.

Value

list object containing empirical Bayes (EMPB) estimates of r, a, and b hyperparameters, assuming df$x ~ nbinom(r, p_g), and p_g ~ beta(a, b), where 'p_g' denotes a group-level parameter.

Examples

# Generate example data:
set.seed(31)
r = 4
a = 3
b = 9

# Number of groups:
NG = 10

# Creating group IDs:
g = replicate(NG, paste(sample(LETTERS, 10), sep="", collapse=""))

# Generating 'true' p parameters:
p = rbeta(length(g), a, b)

# Number of experiments, i.e. rows in df:
numexps = 100

# Filling df with pseudo data; note the requisite columns 'x' and 'g':
df = data.frame('x' = numeric(0), 'g' = character(0))
for(k in 1:numexps){
  gk = sample(g, 1)
  xk = rnbinom(1, r, p[g == gk])
  df = rbind(df, data.frame('x' = xk, 'g' = gk))
}

# Generating empirical Bayes (EMPB) solutions for r, a, and b:
rab_fit = empb_beta_negbinomial_c(df = df, method = 'gdescent')

# Compare fitted values to known values:
cbind(c(r, a, b), c(rab_fit$r, rab_fit$a, rab_fit$b))

alexandercoulter/DHBayes documentation built on Dec. 19, 2021, 12:29 a.m.