wish_list

View page source

CppAD Mixed Wish List

Better Asymptotic Statistics

  1. Detect which fixed effects bound constraints are active and remove them from the asymptotic statistics (constant at bound value).

  2. Use the rejection algorithm to get independent samples from the posterior for the fixed effects. The probability of acceptance \(p_a\) is a measure of how good the Laplace approximation is; see section 3.7.7 of Wakefield’s book Bayesian and Frequentist Regression Methods.

Multi-Threading and Statistics

Change fix_likelihood and ran_likelihood so that every data value is a separate element of the return vector. This would require a separate argument to the cppad_mixed constructor that tells which elements of fix_likelihood are absolute value terms. It would also enable asymptotic statistics using Jacobians of the likelihood vector, instead of Hessians of the total likelihood.

Fixed Likelihood Hessian

Move the initialization of fix_like_hes_ to a separate routine so it does not get calculated when quasi_fixed is true. If quasi_fixed is true, fix_like_hes_ would need to be initialized the first time information_mat is called.

Examples

Where appropriate, change virtual function definitions to override definitions. In addition, change template functions to regular functions when only one type is used as the template argument.

d_sparse_rcv

Convert all the sparse matrix structures to use CppAD::d_sparse_rcv instead of Use sparse_mat_info and row , col , val triples.

Aborting Optimization

It would be helpful to return the fixed and random effects that correspond to the last attempted function evaluations that fail; i.e., that result in nan .

Windows Install

Make and test a set of Windows install instructions for cppad_mixed .