hoa() now computes the modified likelihood root correctly for contrasts
with more than one nonzero coefficient, such as averages, trends or
Tukey-type comparisons in models with correlated coefficients. Previously
the nuisance information was taken from a sub-block of the covariance
matrix, which is only correct for contrasts that select a single
coefficient. Adjusted statistics, intervals and p-values change for such
contrasts, in some cases by as much as the whole higher-order correction.
Pairwise contrasts in cell-means models (for example the Dunnett example in
the vignette) were nearly unaffected.
hoa() also uses the working weights at the final constrained fit, which
changes its results slightly (up to about 1e-4) for all contrasts.
hoa() no longer discards every profile point for larger data sets.
plot() failed for intervals from confint(..., alternative = "less").
expit() is now an exported generic, so expit(confint(...)) transforms
intervals to the probability scale as documented.
mcprofile() failed when all grid points lay on one side of the estimate.
mcprofile() for lm objects failed when the data, weights or formula
were defined inside a function, or when the data had the same name as an
argument of mcprofile() (for example CM).
confint() now checks adjust and alternative. A misspelled
alternative previously gave a confusing error, or could silently return
an unrelated object called ci from the workspace as the intervals.
Unambiguous abbreviations such as alternative = "g" are now accepted.
mcprofile() now gives a clear error for models with aliased (NA)
coefficients, and for a contrast matrix with the wrong number of columns.
Profiling is faster: each constrained fit solves the quadratic program once per iteration, and the starting values on each side of the estimate are chosen better.
Requires ggplot2 3.4.0 or later. plot() for intervals no longer uses the
deprecated geom_errorbarh().
The citation uses bibentry().
Added a test suite.