Current development version
New features
- New method
shapeConstr to build constraint matrices for
shape-constrained splines. Currently works with classes ns,
bs (from package splines), ps and
onebasis (from dlnm). Also includes a default
method for more general bases.
- New function
zerosumConstr for constraint matrices for
a zero sum such as used in compositional regression for instance.
- New function
edf to compute observed and expected
degrees of freedom for a fitted cirls object.
- Method
logLik.cirls for AIC and BIC computation.
- New function
uncons to return the unconstrained
model.
- Constraints can now be passed as a formula through a new argument
called
constr.
- New function
buildCmat to build a constraint matrix
using a model frame and a list of matrices or constraint formula (or
both).
Changes
- Changed the default QP solver to
quadprog after some
expriments.
- Added the argument
complete in inferential functions,
to allow keeping or discarding aliased coefficients. Same interpretation
as in vcov.lm.
vcov now allows returning the usual variance-covariance
returned by vcov.glm when trunc = FALSE.simu_coef renamed as simulCoef and now
includes an argument to set the seed.check_cmat has been renamed checkCmat. It
also now returns logical vectors instead of vectors of indices- Now the
Cmat, lb and ub used
are not returned in the control object from the result of
glm with cirls.fit.
- Now
lb and ub can be passed by term.
- The element
aic of a fitted cirls object
is penalised by the number of active constraints.
Bug fixes
- Fixed issue with R matrix when there was less observations than
variables.
- Fixed error from
solve.QP when there are large values
in the response.
- Now
checkCmat also checks if there are “zero”
constraints.
- Fixed a bug in
simulCoef. Now includes a switch to
simulate under the constrained or unconstrained model.
simulCoef returns a NA matrix with a warning in the
case of a saturated model.
0.3.0
New features
- Added
check_cmat and coef_simu to the list
of exported functions as they can be useful for specific use cases.
- Added full documentation.
Bug fixes
- In
check_cmat, removed the call to
limSolve::nnls() to be replaced by
coneproj::coneB (also a NNLS solver) to reduce the number
of dependencies.
0.2.1
New features
- Initialization of a short documentation for several functions.
Bug fixes
cirls.control now checks for constraint matrix
irreducibility.cirls.control is now exported.- changed the function to determine redundant constraints in
check_cmat
- No warning message on row rank of
Cmat anymore
Cmat is now checked only when there are more than one
row
0.2.0
New features
- Method vcov for cirls to compute corrected covariance matrices
- Method confint for cirls to compute feasible confidence
intervals
- Added several QP solvers: quadprog (the original one), osqp and
coneproj.
Changes
- A warning is now displayed when Cmat is not of full row rank
- vcov and confint return NA matrices if Cmat is not of full row
rank
- Changed residual df computation to account for active
constraints
- Replaced bvec by lb (lower bound) and ub (upper bound). Allows
equality constraints.
- Added cirls class to glm output
Bug fixes
- cirls.fit has the same behaviour as glm.fit when model is singular:
fill with NAs