A patch for the pnbd.LL() function in the original BTYD package, first proposed by Theo Strinopoulos.
The original version of the pnbd (Pareto/NBD) part of the BTYD package failed for me, as it did for Theo, for reasons he explained here. So, I implemented Theo's proposed log-sum-exp trick almost identically to his solution, rebuilt from source, and then it worked. This is the short version. The long version is that in the process of implementing the log-sum-exp fix I made some changes to the choice of optimization routine (now using optimx as opposed to optim), Gaussian hypergeometric (now you have the option of using the hypergeo package) and I found some opportunities for refactoring functions defined in both the pnbd and the bgnbd (Beta-Geometric Negative Binomial) groups. Along the way I defined some new supporting functions in the dc (data conversion) group. Finally, I rebuilt BTYD2 and checked that it would return the same numbers as BTYD did when running the examples in the original BTYD vignette (see threeway_walkthrough.R).
BTYD2 is for the curious and impatient: you can install it now, side-by-side with the unaltered BTYD from CRAN. There is a GitHub-hosted version of BTYD but it's read-only. If the original BTYD were still maintained, and there were a way to submit a pull request, I would. As it is, BTYD2 replicates the functionality of BTYD exactly in the original set of functions whose names start with dc (data conversion) and bgbb (Beta-geometric/Beta-binomial model) with one exception: there's an internal helper to bgbb.LL, called addition, which now also implements the log-sum-exp trick.
This means that you could use BTYD2 like I do, to replace BTYD altogether, once you're satisfied that the patches it proposes
to the pnbd and bgnbd groups of functions meet your needs. But, caveat emptor.
The devtools way is the easiest and it works: install_github('ghuiber/BTYD2', build_vignettes = TRUE).
First git clone. Then, install it the devtools way: document(); build(); install(build_vignettes = TRUE). This process, like install_github(), will obliterate inst/doc. Recover it from the source repo if you want to build your cloned source package the standard way, which is a two-step process:
First, at the command line, do this:
R CMD build BTYD2
R CMD check BTYD2
There may be errors, warnings, notes. Work them out. Then, at the R console, do this:
install.packages("BTYD2_2.4.tar.gz", repos = NULL, type = "source")
system(paste('gunzip', file.path(.libPaths(), 'BTYD2', 'data', 'discreteSimElog.csv.gz')))
system(paste('gunzip', file.path(.libPaths(), 'BTYD2', 'data', 'cdnowElog.csv.gz')))
# works only in RStudio:
.rs.restartR()
If you change any functions or add yours, keep it modern: use devtools (while it's still there) and follow the advice from R Packages. I went by the first edition and I only read the parts I needed as I went along to boot, but the general idea is to document your functions with Roxygen2 headers and do document(); build(); install() as often as needed. If that workflow is no longer the recommended way, do as R Packages says.
A work-in-progress example of BTYD2 extensions is in R/equi.R. See also ?ace_snackfoods.