Re: [dinosaur] What is the earliest known bird

[David Marjanovic <david.marjanovic@gmx.at>]

> While it was common to think of calibrations in terms of "maximum ages to go 
> with the minimum ages" in the 1990s and early 2000s, the field has moved on, 
> and this is not a common practice anymore. With some notable exceptions 
> (e.g., treePL), modern software packages (e.g., MCMCTree, MrBayes, BEAST 1, 
> BEAST 2, RevBayes) generally give the user the option to specify a parametric 
> calibration density with support on [0, infinity) or [offset, infinity), 
> where the offset represents the minimum age. Exponential, lognormal, and 
> truncated Cauchy densities are often used, and there is now a moderately 
> large body of literature on how these can be made less arbitrary by fitting 
> them to the distribution of fossil occurrences through time.

Yes, I should have mentioned that. I was thinking in the context of my upcoming 
paper, which is based on a paper from 2017 (Irisarri et al.: 
Phylotranscriptomic consolidation of the jawed vertebrate timetree. Nat Ecol 
Evol 1, 1370â1378. DOI 10.1038/s41559-017-0240-5) that used pretty classic 
node-based dating with minima and maxima.

> The user still can (and I've done this before) choose to treat, say, the 95th 
> percentile of an exponential as a "soft" maximum and scale the rate parameter 
> so as to place 95% of the total probability mass between the minimum and 
> maximum ages, but that's just one possible way to think about it.

In PhyloBayes, the software used in that particular paper, calibrations can 
only be treated as hard or soft. "Soft" means that 5% of the probability mass 
of the result if there's only a minimum or a maximum for a particular 
calibration, or 2.5% if there are both, _must_ be younger than the minimum or 
older than the maximum. Having a soft maximum and a hard minimum for the same 
calibration is not an option, and the shape of the probability distribution 
cannot be tuned. That way, even the mean age of a calibrated node can end up 
younger than its minimum.

Soft minima almost never make sense. The one exception I came up with is when 
the phylogenetic position of a calibrating fossil is sufficiently well known, 
but its age is not. The opposite situation is much more common these days.

> Parametric densities also give you the freedom to specify your prior beliefs 
> in different ways, some of which may make more sense depending on the 
> situation: e.g., using a mean and a standard deviation, the distance between 
> the minimum and the mode, the 95% prior credibility interval, etc. Specifying 
> a soft maximum and parameterizing your density accordingly can help avoid 
> implausibly old age estimates when your fossil record sucks, but when it 
> doesn't, the framework allows you to do much better than that.

Yes. None of that, unfortunately, is possible in PhyloBayes.