I’ve found R package:pomp useful in a recent publication, where I used its capabilities only minimally. But I hold out high hopes for its iterated filtering and particle MCMC (PMCMC) inference methods as I extend this work. So I wonder if any of you have experimented with this package for PK/PD applications, and would share impressions or conclusions about its advantages and limitations vis-à-vis other tools – especially Stan.
Based on your response to a different post, I checked the publication and found the pomp package. I had never heard about it before, but your seem to have used it for the first time in our field. What is the reason for this choice?
If you are looking for open source tools, specifically estimation engines, you will find a lot coming your way from the community in 2017 and 2018, including full Bayesian in Stan via torsten and other maximum likelihood based estimators such as nlmixr. Not to forget Pmx-stan developed by the folks who developed nlmixr. These are just a couple of them, but a more comprehensive toolset is in the making!
Thanks for your reply, Vijay, with its useful links that I’ll certainly keep an eye on! As you guessed, I seek exclusively open source tools.
As for my own rationale, it draws on several factors. Firstly, I’m approaching my PK/PD problems from a Kalman filtering mindset, to which the language of POMP speaks so directly and appealingly. Secondly, I ‘got burned’ once by Stan when I converted a (non-PK/PD) JAGS model to Stan with no speedup – due I’m told to a numerical problem in Stan’s implementation of the beta distribution. (Also, it seems Stan offers little in the way of filtering.) Thirdly, package pomp was developed by ecologists, whose compartmental-model mindset carries over immediately into PK/PD. I had already implemented some of my models in deSolve, and (perhaps because pomp uses deSolve under the hood) carrying these over into pomp proved utterly straightforward. Fourthly, I find the plug-and-play concept of pomp immensely appealing in the context of my DTAT research, because freedom to revise model specification is so valuable.
Of course, I do not feel entirely assured that I’ve made a ‘bulletproof’ tool choice here. (Can one ever have such assurance ?) This motivated my reaching out to hear others’ considered views here, perhaps based on deeper experience. I must say, however, that even if I eventually switch to a more mainstream PK/PD tool, I certainly will not regret having adopted pomp initially – simply on account of the fascinating literature it has pointed me toward, and (for me) its remarkable ease-of-use.
I’ve just checked with the authors of pomp, and can confirm that they also know of no previous pharmacometrics application of their package. So I look forward to returning (2-4 weeks) with more reflections when I feel I’ve got a properly considered opinion to contribute.