knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "README-"
)

addreg

CRAN_Status_Badge

addreg provides methods for fitting identity-link GLMs and GAMs to discrete data, using EM-type algorithms with more stable convergence properties than standard methods.

An example of periodic non-convergence using glm (run with trace = TRUE to see deviance at each iteration):

require(glm2, quietly = TRUE)
data(crabs)

crabs.boot <- crabs[crabs$Rep1, -c(5:6)]

t.glm <- system.time(
  fit.glm <- glm(Satellites ~ Width + Dark + GoodSpine, data = crabs.boot, family = poisson(identity),
                    start = rep(1, 4), maxit = 500)
)

The combinatorial EM method (Marschner, 2010) provides stable convergence:

require(addreg, quietly = TRUE)
t.cem <- system.time(
  fit.cem <- addreg(Satellites ~ Width + Dark + GoodSpine, data = crabs.boot, family = poisson,
                    start = rep(1, 4))
)

...but it can take a while. Using an overparameterised EM approach removes the need to run $2^3 = 8$ separate EM algorithms:

t.em <- system.time(fit.em <- update(fit.cem, method = "em"))

while generic EM acceleration algorithms from the turboEM package --- implemented in version $\geq$ 3.0 --- can speed this up further still:

t.cem.acc <- system.time(fit.cem.acc <- update(fit.cem, accelerate = "squarem"))
t.em.acc <- system.time(fit.em.acc <- update(fit.em, accelerate = "squarem"))

Comparison of results:

fit.list <- list(fit.glm, fit.cem, fit.em, fit.cem.acc, fit.em.acc)
time.list <- list(t.glm, t.cem, t.em, t.cem.acc, t.em.acc)
res <- data.frame(converged = sapply(fit.list, function(x) x$converged),
                  logLik = sapply(fit.list, logLik),
                  iterations = sapply(fit.list, function(x) x$iter[1]),
                  time = sapply(time.list, function(x) x[3]))
rownames(res) <- c("glm", "cem", "em", "cem.acc", "em.acc")
res

The combinatorial EM algorithms for identity-link binomial (Donoghoe and Marschner, 2014) and negative binomial (Donoghoe and Marschner, 2016) models are also available, using family = binomial and family = negbin1, respectively.

Semi-parametric regression using B-splines (Donoghoe and Marschner, 2015) can be incorporated by using the addreg.smooth function. See example(addreg.smooth) for a simple example.

Installation

Get the released version from CRAN:

install.packages("addreg")

Or the development version from github:

# install.packages("devtools")
devtools::install_github("mdonoghoe/addreg")

References



mdonoghoe/addreg documentation built on May 22, 2019, 4:19 p.m.