I’ve recently used PyMC3 to tackle an analysis in a study I’m now preparing for publication (field biomedicine and microbiology). Though I find myself in doubt how to write the methods clearly and concisely for that specific part of the study. A search for publications that use PyMC for an aspect of their analyses so far didn’t yield any paper I could use as a reference.
Is anyone aware of studies that used PyMC as part of their analyses which have been published already?
This would be very helpful as a guide how they included their models and described the workflow in a manuscript.
This and this would be a couple of recent examples.
The problem, as you mention, is that the methodological descriptions of Bayesian analyses (conducted in PyMC or not) vary enormously. Worse yet, there seems (to me) to be weird paradoxical pattern where those who have more experience with Bayesian modeling tend to write less (i.e., because they are confident in what they did) and those who are less familiar tend to write more (e.g., because they are insecure). My suggestion would be to take a look at this paper and select the subset appropriate for your paper. I think all this recommendations are good, but I don’t think many journals would accept anything nearly so thorough. Similar guides can be found elsewhere if you want a sense of the diversity and points of (dis)agreement: here, here, here, here, and here.
Yes, there is no agreed upon way of going about things at this point, and it is probably situation specific. For example, a Bayesian methods paper trying to establish a new model would go into more detail, but in an empirical paper then less detail could work.
For empirical papers, I think a good minimal approach would be to include a mathematical description of the model (both priors and likelihood). I’ve got a paper here for example which which shows 2 bayesian models in section 2.6 and 2.7. I tend to include Jypyter notebooks which are (hopefully) reproducible in the supplementary material - and I leave all the prior predictive checks / trace plots etc in there.
Thanks for sharing. I use a very similar approach, but I link additional info (traceplots, autocorrelation plots, etc) to the Open Science Framework (OSF), where I also save traces and raw data to facilitate reproducibility.
In case an example from cognitive neuroscience is helpful to someone, I have an example here.