Dear Andre,
(I repeat here the same message that I've just sent through the ccp4bb; sorry for answering late)
I would strongly advice you to look at the article by Lunin and Skovoroda (Acta Cryst, A, 1995) that addresses exactly your question:
https://scripts.iucr.org/cgi-bin/paper?vs0124
The authors remind a very important point that after model refinement ML phase errors are strongly underestimated if using all reflections, as that was done in the original works (see references in the article). While the same ML estimates work perfectly for unrefined models, that's not the case for refined ones, as was observed yet in the beginning of 80ths.
These authors show then that using the test-set of reflections (the same as for R-free) is crucial to get the correct phase error estimates and respective FOMs for all cases, as this is implemented now in modern refinement programs, e.g. phenix.refine. See also the article by Pannu & Read (Acta Cryst, A, 1996)
https://onlinelibrary.wiley.com/doi/pdf/10.1107/S0108767396004370
A more recent important article on this topic is that by Praznikar and Turk (Acta Cryst, D, 2009)
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4257616/
who discuss what can be done if a statistically significant test set of reflections is not available.
I hope this helps you. I am ready to answer you further eventual questions (probably, in a private correspondence).
With best wishes,
Sacha Urzhumtsev
________________________________
-------- Forwarded Message --------
Subject: [phenixbb] Figure of merit from refinement
Date: Tue, 1 Oct 2019 08:15:14 -0300
From: Andre LB Ambrosio