PC prior for precision

[maxbiostat]

Hi,

I was wondering if there'd be interest in implementing the penalised complexity (PC) prior of Simpson et al. (2017) for the precision in BNPR.

I have implemented it in BEAST and in phylodyn (in a fork). A brief simulation study (notebooks and write-up) is here.

Cheers,

Luiz

[JuliaPalacios]

Hi Luiz, Can you please point me to your code's BNPR function? I am interested in using your code. Did you implement it in the MCMC methods or in the INLA approximation? Thanks Julia Palacios

[maxbiostat]

Dear Julia, the "new" BNPR function is here. I only implemented the prior for the INLA methods, but could implement it for the MCMC if there was interest. Please note that there may be small bugs, as I haven't tested the functions extensively. I should also probably add that while the PC prior does add some regularisation, it's nothing outstanding, as expected -- $\tau$ is quite high up in the hierarchy, so it's hard to control it in when the data are "weak".

[maxbiostat]

Dear @JuliaPalacios ,

I am considering submitting a pull request with these minor changes. Did you have a chance of trying them?

[maxbiostat]

This provides a very minor improvement by using a principled prior. My branch has an implementation for BNPRand BNPR_PS. I'd be glad to have a look at implementing this for other methods, but first we need to make sure it's worth it. @vnminin @JuliaPalacios and @mdkarcher if you want to weigh in, that'd be cool. I understand dealing with COVID-19 might take precedence over principled prior modelling, though ;-)

[vnminin]

Thanks for your comment, @maxbiostat. I am definitely interested in bringing phylodyn smoothing priors up-to-date with the literature. Jim Faulkner, Andy Magee, and I played with similar ideas in another package: https://github.com/jrfaulkner/spmrf. It would be good to consolidate all of this into phylodyn.

But you are right -- COVID-19 projects take priority right now, so we'll have to put this on a back burner for now.