logbin (version 2.0.4)

conv.test: Convergence Test Based on L2 Norm

Description

Performs a test of convergence based on the L2 norm of the change in the parameter estimates.

Usage

conv.test(theta1, theta2, epsilon)

Arguments

theta1

vector of parameter estimates at previous step.

theta2

vector of parameter estimates at current step.

epsilon

positive convergence tolerance.

Value

A logical; TRUE if sqrt(sum((theta1-theta2)**2))/sqrt(sum(theta1**2)) < epsilon; FALSE otherwise.

Details

This is used as the convergence test in the logbin fitting functions: it is passed as convfn.user to turboem. It is used instead of a test based on deviance (such as that applied in glm.fit), which may report convergence at a point away from the actual optimum when the EM algorithm converges slowly.

Examples

# NOT RUN {
theta.old <- c(-4,-5,-6)
theta.new <- c(-4.05,-5,-6)

conv.test(theta.old, theta.new, 0.01)
conv.test(theta.old, theta.new, 0.005)
# }