EMGCR 0.3.0
Breaking changes
- The latency component is now parametrized through the scale
parameter
lambda = exp(x %*% beta) for all distributions,
as in an accelerated failure time model (the same convention as
survival::survreg()). Previously, the exponential,
Rayleigh, Weibull and log-normal distributions used a rate
parametrization. The fitted model and the log-likelihood are unchanged,
but the latency coefficients of these four distributions now have the
opposite sign (and, for Weibull and Rayleigh, a different scale)
relative to version 0.2.0. A positive coefficient now means longer
survival for the uncured. The log-logistic and inverse Gaussian results
are unchanged. rMCM(), residuals(),
qqMCR() and plot() follow the new
parametrization.
- The
liver dataset was replaced by the liver cancer data
used in the accompanying paper (1,736 patients; covariates
sex, age, medh,
race, grade, chemo and
radio).
Bug fixes
- Standard errors of the latency coefficients for the log-normal
distribution were too small; they are now correct.
summary() failed when the data had been modified after
the fit (for example, when running the help-page examples).rMCM() failed when x had a single
covariate besides the intercept.plot() no longer triggers the ggplot2 deprecation
warning about size.- For the inverse Gaussian distribution, the starting values of the
latency coefficients followed the old rate parametrization, which could
make the EM algorithm stop at a degenerate solution for some
models.
New features and
improvements
summary() shows significance stars and a summary of the
quantile residuals, and returns the coefficient tables as numeric
matrices.residuals() returns an object that prints the first
residuals and a summary instead of every value.MCRfit() warns when the EM algorithm does not converge
within maxit iterations and stores a logical
convergence element; summary() reports
it.print() and summary() show the link
function (with tau only for the power links), label the
incidence coefficients consistently as the “Uncured part”, show the
shape parameter only when it is estimated, and return their argument
invisibly.- Documentation of
MCRfit(), rMCM() and
plot() was corrected and expanded.
knitr is no longer a dependency.