Bayesian Meta-analysis #16
Description
This is a tracking note on the issue of Bayesian analysis in MetaWin. One of the biggest fundamental problems with adding Bayesian analysis is that most of the Bayesian packages available in Python are not truly multiplatform.
For example, one of the most widely used packages, PyMC, requires access to a Fortran compiler at runtime; this is generally done through gcc, which functionally makes PyMC use restricted to Linux (and thus also Mac) environments. While it is possible to get it working in Windows, it is generally difficult and not something that can be readily distributed as part of a package such as MetaWin. Developers don't necessarily see this as a problem because they tend to assume that (a) "real" computation is only done in Linux, and (b) users of Windows can go through the effort of installing Linux-emulation (or other similar strategies) if they really want to use these tools in that environment. For the MetaWin user base, it's just not that practical.
Another newer package, with a growing fan base, NumPyro has the same sort of problem as it is reliant on the just-in-time compiler JAX, which is again not really compatible with Windows for all practical purposes (like with gcc, it is possible to get it running in Windows, but not in a way that makes it generally useful for our purposes).
The good news is that there is another potentially workable solution. The package Pyro (which uses another package called PyTorch) appear to not have the platform limitations that the above named tools have. I'm currently in the process of learning more about it to see if I can figure out how to start implementing models that I could use within MetaWin.
There is no timeline on when this may become available, but it is a work in progress.