deconv: A function to compute Empirical Bayes estimates using...
In deconvolveR: Empirical Bayes Estimation Strategies

Description Usage Arguments Value Details References Examples

View source: R/deconv.R

A function to compute Empirical Bayes estimates using deconvolution

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,
  ...
)

`tau`	a vector of (implicitly m) discrete support points for θ. For the Poisson and normal families, θ is the mean parameter and for the binomial, it is the probability of success.
`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`). This applies to the Normal case only and even then only if it is 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`

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
`stats`	an m by 6 or 7 matrix containing columns for theta, g, \tilde{g} which is g with thinning correction applied and named `tg`, 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

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.

Note also that in the Poisson case where there is zero truncation, the stats matrix has an additional column "tg" which accounts for the thinning correction induced by the truncation. See vignette for details.

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.

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$stats)
##
## Twin Towers Example
## See Brad Efron: Bayes, Oracle Bayes and Empirical Bayes
## disjointTheta is provided by deconvolveR package
theta <- disjointTheta; N <- length(disjointTheta)
z <- rnorm(n = N, mean = disjointTheta)
tau <- seq(from = -4, to = 5, by = 0.2)
result <- deconv(tau = tau, X = z, family = "Normal", pDegree = 6)
g <- result$stats[, "g"]
if (require("ggplot2")) {
  ggplot() +
     geom_histogram(mapping = aes(x = disjointTheta, y  = ..count.. / sum(..count..) ),
                    color = "blue", fill = "red", bins = 40, alpha = 0.5) +
     geom_histogram(mapping = aes(x = z, y  = ..count.. / sum(..count..) ),
                    color = "brown", bins = 40, alpha = 0.5) +
     geom_line(mapping = aes(x = tau, y = g), color = "black") +
     labs(x = paste(expression(theta), "and x"), y = paste(expression(g(theta)), " and f(x)"))
}