Transition-specific costs and cCTSTMs

[mclements]

Devin,

I have been working on a few changes to the hesim package.

For this pull request, I have implemented transition-specific costs, which currently only has an application for iCTSTMs.

Note that this folds in the unresolved inline extension from pull request #122.

As a signpost: I am preparing a presentation for the International Society of Clinical Biostatistics for later this month that implements cCTSTMs using ordinary differential equations. I have a gist and will prepare a pull request. I have a number of questions to ask you about that pull request:).

Sincerely, Mark.

[mclements]

I have added some changes to address the recent R check notes and warnings:

• I incremented the version number
• To remove a note about R6 not having any imports, I added @importFrom R6 R6Class in ctstm.R -- although this is not strictly necessary.
• Most of the other notes had to do with documentation external links not having packages.

[mclements]

Devin: pkgdown is complaining about the documentation for CohortCtstm in ctstm.R. I can't see the error - would you be able to help here, please?

[mclements]

Some commentary on the implementation:

• I was not able to find another package on CRAN that allows LinkingTo: for ordinary differential equations in C++. I investigated adapting the rehuel library for CRAN, but found that the package is not well maintained and some of the tests were failing. I have kept the BH dependency. For further discussion?
• As discussed at The individual-based Markov CTSTM vignette uses a fixed time: is this sensible? #126, I have assumed that the point mass distribution is Markov. I am not happy with the implementation for the point masses: I needed dynamic_cast to determine (a) if a transmod was an mstate_list, and then (b) whether an element of mstate_list->survmod_ is a point mass (after adding an hesim::stats::distribution* get_dist(int trans) method to mstate_list). The code would be cleaner and clearer if we did not allow for point masses.
• For the standard hip replacement vignette, I have assumed that the fifth transition is exponential. This is strictly Markov (clock-forward), whereas the individual-based vignette is a mix of clock-forward and clock-reset transitions.
• I have some test code in preparation, but not completed.

[mclements]

I have now fixed the pkgdown issue.

[mclements]

@dincerti, I have now added some tests for this. This is now ready for review -- although admittedly this is a larger pull request than initially anticipated.

[mclements]

Devin:

• I have now added an experimental cCTSTM implementation for semi-Markov models.
• The implementation is similar to the Markov case, except that one needs to discretise for duration and keep track of both time in state and time in study.
• However, the implementation is slow and needs +to be optimised.
• TODO: I need to address some check warnings.