NEWS.md
In pedmod: Pedigree Models

pedmod 0.2.4

eval_pedigree_hess is faster.
fixed a bug from release 0.2.0 which in extreme settings could cause the C++ code to run forever.
deal with a breaking change in RcppArmadillo which could possibly cause issues in this package. See https://stackoverflow.com/a/72533955/5861244
better starting values are used by pedmod_profile_prop. It is also possible to pass a bound on the confidence interval using the bound argument.
mvndst_grad is added which computes the gradient with respect to the mean and covariance matrix.

a minor bug fix on Mac when using Apple LLVM version 10.0.0 with R version 4.2.0 and x86_64.

A hessian approximation of objects from pedigree_ll_terms is added in the eval_pedigree_hess function.
pedmod_profile works with object from pedigree_ll_terms_loadings.
pedmod_profile_nleq has been added to construct profile likelihood based confidence intervals for general non-linear transformations of the model parameters.
An undefined undefined behavior bug has been fixed in the C++ code which possibly effects cases where use_aprx = TRUE but only in very extreme settings.
A bug has been fixed in pedmod_profile_prop. minvls_start and maxvls_start were used instead of minvls and maxvls.
The code to compute the limits in pedmod_profile and pedmod_profile_prop has been changed. The previous code could give very wrong points for the conf element if a point was computed very far from one of the confidence limits. The issue was caused by using approx in combination with spline and with points with great distance.

pedigree_ll_terms_loadings is implemented to support models with individual specific covariance scale parameters (e.g. individual specific heritabilities).
The minimax tilting method suggested by Botev (2017) (see https://doi.org/10.1111/rssb.12162) is implemented. The method is less numerically stable and thus required more care when implementing. This yield a higher per randomized quasi-Monte Carlo sample cost. Though, the increased cost may be worthwhile for low probability events because of a reduced variance at a fixed number of samples.
The vls_scales argument is added which allows the user to use more randomized quasi-Monte Carlo samples for some log likelihood terms. This is useful e.g. when one uses weighted terms.