deconv: A function to compute Empirical Bayes estimates using...

Description Usage Arguments Value Details References Examples

View source: R/deconv.R

Description

A function to compute Empirical Bayes estimates using deconvolution

Usage

1
2
3
deconv(tau, X, y, Q, P, n = 40, family = c("Poisson", "Normal", "Binomial"),
  ignoreZero = TRUE, deltaAt = NULL, c0 = 1, scale = TRUE,
  pDegree = 5, aStart = 1, ...)

Arguments

tau

a vector of (implicitly m) discrete support points for θ

X

the vector of sample values: a vector of counts for Poisson, a vector of z-scores for Normal, a 2-d matrix with rows consisting of pairs, (trial size n_i, number of successes X_i) for Binomial. See details below

y

the multinomial counts. See details below

Q

the Q matrix, implies y and P are supplied as well; see details below

P

the P matrix, implies Q and y are supplied as well; see details below

n

the number of support points for X. Applies only to Poisson and Normal. In the former, implies that support of X is 1 to n or 0 to n-1 depending on the ignoreZero parameter below. In the latter, the range of X is divided into n bins to construct the multinomial sufficient statistic y (y_k = number of X in bin K) described in the references below

family

the exponential family, one of c("Poisson", "Normal", "Binomial") with "Poisson", the default

ignoreZero

if the zero values should be ignored (default = TRUE). Applies to Poisson only and has the effect of adjusting P for the truncation at zero

deltaAt

the theta value where a delta function is desired (default NULL). Applies to Normal only if non-null

c0

the regularization parameter (default 1)

scale

if the Q matrix should be scaled so that the spline basis has mean 0 and columns sum of squares to be one, (default TRUE)

pDegree

the degree of the splines to use (default 5). In notation used in the references below, p = pDegree + 1

aStart

the starting values for the non-linear optimization, default is a vector of 1s

...

further args to function nlm

Value

a list of 9 items consisting of

mle

the maximum likelihood estimate \hat{α}

Q

the m by p matrix Q

P

the n by m matrix P

S

the ratio of artificial to genuine information per the reference below, where it was referred to as R(α)

cov

the covariance matrix for the mle

cov.g

the covariance matrix for the g

mat

an m by 6 matrix containing columns for theta, g, std. error of g, G (the cdf of g), std. error of G, and the bias of g

loglik

the negative log-likelihood function for the data taking a p-vector argument

statsFunction

a function to compute the statistics returned above

Details

The data X is always required with two exceptions. In the Poisson case, y alone may be specified and X omitted, in which case the sample space of the observations $X$ is assumed to be 1, 2, .., length(y). The second exception is for experimentation with other exponential families besides the three implemented here: y, P and Q can be specified together.

References

Bradley Efron. Empirical Bayes Deconvolution Estimates. Biometrika 103(1), 1-20, ISSN 0006-3444. doi:10.1093/biomet/asv068. http://biomet.oxfordjournals.org/content/103/1/1.full.pdf+html

Bradley Efron and Trevor Hastie. Computer Age Statistical Inference. Cambridge University Press. ISBN 978-1-1-7-14989-2. Chapter 21.

Examples

1
2
3
4
5
6
7
set.seed(238923) ## for reproducibility
N <- 1000
theta <- rchisq(N,  df = 10)
X <- rpois(n = N, lambda = theta)
tau <- seq(1, 32)
result <- deconv(tau = tau, X = X, ignoreZero = FALSE)
print(result$stats)

deconvolveR documentation built on May 29, 2017, 3:06 p.m.