vblogit: Fit logistic regression model using VB approximation

In antiphon/vblogistic: Logistic regression using Variational Bayes approximation

Fit logistic regression model using VB approximation

Description

Bayesian fit of logistic regression model. p coefficients, n observations.

Usage

vblogit(
  y,
  X,
  offset,
  eps = 0.01,
  m0,
  S0,
  S0i,
  xi0,
  verb = FALSE,
  maxiter = 1000,
  ...
)

Arguments

y	binary vector of responses, length n
X	n x p matrix of covariates, including 1-column for intercept
offset	n-vector of offsets (or 1-vector which will be replicated)
eps	convergence criterion, increase in log-likelihood is no more than this
m0	p-vector of prior means
S0	p x p prior covariance matrix
xi0	p-vector of initial
verb	verbose output, logical
maxiter	upper limit for iterations
...	ignored. Computes the posterior distribution of regression coefficients in logistic regression using the method of Jaakkola&Jordan 1996.

Examples

## some data
n <- 100
p <- 10
X <- matrix( rnorm(n*p), ncol=p)
theta <- rnorm(p)
prob <- 1/(1+exp(-X%*%theta))
y <- rbinom(n, 1, prob)

## See that it works:
## vb:
fit_vb <- vblogit(y, X, verb=TRUE)
## glm:
fit_glm <- glm(y ~ -1+X, family=binomial)

coefs <- cbind(vb=fit_vb$coef, glm=fit_glm$coef)

summary(fit_vb)

## compare vb and glm
plot(coefs, main="Estimates")
abline(0,1)

## Compare to true coefficients
plot(coefs[,1]-theta)
points(coefs[,2]-theta, col=3, pch=4)
abline(h=0)
legend("topright", c("glm","vblogit"), col=c(1,3), pch=c(1,4))

antiphon/vblogistic documentation built on Oct. 14, 2023, 1:14 a.m.