NEWS

cirls 0.4.0-6

checkCmat has been replaced by reduceCons a more comprehensive function to reduce constraints, removing redundant and underlying equality constraints. See ?reduceCons.
reduceCons can now be controlled from cirls.control.
Changed default of dim when constraining crossbases. Now defaults to var whatever the specification.
Some changes to examples for better clarity.

boundConstr.crossbasis was sometimes removing constraints.
Bounds for constraints on crossbasis, were sometimes forced to 0 even when the initial bounds were different.

Shape-constrained DLNM: methods dedicated to crossbasis objects. Add constraints to the exposure-response association or on the lag-response association (for DLMs).
New Constr method: boundConstr to contraint the bounds of an association
New parameter range fhor shapeConstr: constrain subrange of x

On attach message for the package.
New package overview help page.
Addition of three datasets
New examples
checkCmat now contains switches to return the reduced Cmat and disable warning when redundant/equality constraints are found.

Some changes in internal functions streamlining the building of the constraint matrix and bounds. Should not impact the use apart from the odd error message.
checkCmat now returns reduced Cmat/lb/ub, and better identifies equality constraints.

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).

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.

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.

Added check_cmat and coef_simu to the list of exported functions as they can be useful for specific use cases.
Added full documentation.

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.

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

cirls.fit has the same behaviour as glm.fit when model is singular: fill with NAs