Unlinked branch lengths behavior

[tiago-simoes]

When running the current developers version ("3.2.8", but also with previous versions) using unlinked branch lengths across partitions but keeping linked topologies, I see some potential issues with the output files. For example, if running 4 independent runs with 2 partitions, a typical analysis with linked branch lengths outputs 4 tree output (.t) files and 4 parameter file outputs (.p), one for each independent run. When running with UNLINKED BLs, one should expect 4 separate .t, .p, and consensus output files for each partition; in this 2 partitions example, one would have 4 x 2 = 8 .t and .p files and 4.con files in total.

Accordingly, Mr. Bayes does output these 8.t files and 4.con files, but .p files continue to be 4 .p files in total, with labels indicating each came from a different run. This gives me the impression that it is somehow conflating the values of the two partitions under the same file for equivalent runs (e.g., infile.nex.run1.p has parameters averaged from both partitions).

Shouldn't these .p files be unlinked and matched the number of .t files? This lack of equivalency among them makes it hard to run diagnostic parameters with tools like RWTY. Also, it's not clear how to interpret summary statistics that are independent across partitions when conflated under the same variable names under a single partition file.

The issues above seem to reflect on the convergence diagnostics reported on the log output of Mr. Bayes itself. A typical analysis with linked BLs converges well for a particular dataset (96 taxa and 11 loci), but the exact same dataset and parameters but with UNLINKED BLs indicates no convergence according to ASDSF values (they fluctuate around the same value throughout the entire run), high PSRFs, and effective sample sizes are much much lower than using linked BLs. It reads as if diagnostic values were all based on the assumption of linked BLs only. Is that right?

Thanks in advance for any feedback.