Alexei-Kassian
commented
4 years ago Hi Tiago,
-
The good news is that MrBayes is able to elaborate datasets with fossils only. See, e.g., our forthcoming paper https://www.academia.edu/39903804/ (Supplement_technical-files.zip with MrBayes files).
-
Having prepared the paper I encountered a similar problem as you did: with a certain sequence of the input commands, chains don't converge. My specific case was reported by Igor Yanovich here https://github.com/NBISweden/MrBayes/issues/43#issuecomment-400369802, but one should suspect that the real number of such problematic command sequences is bigger.
danielmcasali
Hello ,
I’m reporting here an important issue that I recently had with Mr. Bayes regarding the analysis of data sets including only fossil species. I have discussed this with others and it seems to be an actual issue with the program that no one seems to find an answer to.
When using fossil only data for tip-dating with relaxed clock models Mr. Bayes seems to not be able to handle the lack of taxa with age = 0Myrs and there is no mixing between chains or runs never reaching convergency. I initially thought this was a problem with the tree prior, but now I’m less certain about it. Here is the three major sets of analyses I did. I used the fossilized birth-death tree model (FBD) and tested both the IGR and TK02 clock models as well as fossiltips and random sampling strategies for each, always having the same problematic output:
1) Following the command reference strictly, I set the rho parameter (through the “sampleprob” prior on the probability of sampling extant taxa) value to “0”. But under this option in Mr. Bayes 3.2.6, I get an error message indicating that this parameter has to be “set between 0 and 1”. Using Mr. Bayes v. 3.2.7a, the analyses will start running, but the MCMC chains are somehow stuck and the average standard deviation of split frequencies between chains maintains the exactly same value from the first to the last generation. The final results are nonsensical since the chains actually didn’t sample anything.
2) I subsequently tried repeating the prior settings from the only morphological clock paper that I am aware of in which all taxa were extinct and analyzed with Mr. Bayes (Bapst et al. 2016— reference at the end). In the latter, the authors simply omitted the sampleprob prior from Mr. Bayes block, which I assume was then kept at the default value of 1.0 by the program (an incorrect model assumption, but I give it a shot anyways). In this case, both versions of Mr. Bayes run fine at the beginning, but despite detecting some good mixing between chains in the final log file, the average standard deviation of split frequencies at the end was still indicating a very poor convergence between runs. As a result, most parameters could not be sampled from the stationarity phase, with PSRFs very far from 1.0 (e.g., TH = 1.598!) and log likelihood for each run very distinct from each other. As expected, tree topology is mostly full of polytomies. Bapst et al. also reported that their runs with Mr. Bayes using only extinct taxa had very poor mixing between chains… I reran their analyses and I do get very similar behavior to running my dataset (although their dataset performance is somewhat better, but still problematic).
3) I tried using the uniform tree prior for the morphological clock analysis instead of the FBD model, in order to check if this was an issue with the FBD model implementation in Mr. Bayes. However, I still got very similar results, indicating this may not be an issue related to the tree model itself, although I still cannot figure out the reason for this behavior when the only difference is having no extant taxa.
Interestingly , in another dataset in which only one taxon is extant the analysis ran perfectly fine. I had not faced this issue before since all of my previous interactions with Mr. Bayes included analyses inclusive of extant taxa. My question is, what may be causing this apparent misbehavior and is that an issue known to the software developers? With the increasing adoption of Mr. Bayes and Bayesian inference by paleontologists, this is a major apparent flaw that will highly restrict its use for me and my peers.
All the best, Tiago Simões
Bapst, D. W., Wright, A. M., Matzke, N. J. & Lloyd, G. T. 2016. Topology, divergence dates, and macroevolutionary inferences vary between different tip-dating approaches applied to fossil theropods (Dinosauria). Biology Letters, 12(7).
How may we reproduce this bug?
See attached dataset files with embedded Mr. Bayes blocks: 1-SS_BayesCal_LN-Sym_IGR-ga_Rho=0: input and output log files for the analysis with rho=0 as reported above for analysis 1 (this is a steppingstone run that I was performing to compare the model marginal likelihood for the IGR clock model).
2- BayesCal_LN-Sym_IGR-ga_NoRho.txt: input and output log files without the rho parameter as reported above for analysis 2.
3- BayesCal_LN-Sym_IGR-ga_FT_Uniform: input and output log files using the uniform tree prior as reported above for analysis 3.
Steps to reproduce the bug: Simply run each of the input files above to get the observed outputs reported (1-3).
Would you be able to compile and run MrBayes to test fixes to this bug?
What is the environment that you run MrBayes in?