New results vs. Past results

[Nick L.]

Dear Tracy,

I have been using the program and am pretty happy with it. However, I am having a bit of a troublesome issue interpreting results. I ran four categories of analysis based on calibration strategy: 1.) all four calibrated nodes used a generous exponential prior; 2.) all four nodes used a uniform prior; 3.) the root node had a uniform prior and the other three had generous exponential priors; and 4.) just the root node had a uniform prior and no other nodes were calibrated. These calibrations were all based on pretty good fossil data. First, I mostly tried to distinguish among these runs just by eyeballing them, but I wonder if there is a more quantitative way (Bayes factor?). Anyway, strategies 1 and 3 gave me my most reasonable seeming results. The problem is that there were a couple of nodes  that I have previously dated (very recent 0.07 - 5 million years) using coalescent approaches and under the Dpp-div analysis they came out as much older than what I previously found. This may be a problem with my previous results, but I thought I would ask. See the runs performed only using uniform prior distributions had dates much more consistent with my previous analyses, but the deep node ages were very unrealistic. On the other hand, the exponential analyses showed much more realistic dates for the deep nodes, but everything near the tips of the tree were pulled quite far back. Could I address this in some way, without trying to tailor analyses to fit a preconceived result? Thanks Tracy.

Nick

[Tracy Heath]

Hi Nick,

I think the important thing to keep in mind is the fact that there really is no information in the sequence data for the absolute node ages. Furthermore, when doing divergence-time estimation, we are trying to estimate two conflated parameters -- rate and time. Thus, much of the information for the node times comes from the priors. And with regard to absolute ages, pretty much all of the information comes from the calibration priors. So it is unsurprising that you get different results with different calibration priors or with different tree priors (coalescent vs. birth-death). 

The models implemented in DPPDiv for the node-age priors really assume you have divergent taxa, where the tips represent species, not populations. Ultimately, since there isn't a lot of machinery for performing model selection, then you should select a prior that best matches your assumptions about your data. Thus, if you have species-level data, then a birth-death prior is probably reasonable. Perhaps, though, it's worth using BEAST and path-sampling to calculate marginal likelihoods and Bayes factors to determine if your data strongly support a birth-death model versus a coalescent tree prior. With regard to calibration, you should select and parameterize calibration densities that reflect YOUR statistical uncertainty in the age of the calibrated node with respect to the calibration fossil. If your uncertainty in older node ages is very high, then this means a diffuse prior on the node times. However, this uncertainty will then be propagated to younger node possibly resulting in wider credible intervals and higher mean ages. DPPDiv allows you to place a hyper prior on the rate parameters of exponential calibration densities. This means you can account for uncertainty in those parameters by marginalizing over all possible values. I suggest you go with this approach.

What you have already done is shown that estimates of node ages are highly sensitive to the choice and parameterization of the priors. This is par for the course in Bayesian divergence-time estimation, unfortunately. So in the end, you simply have to put your money down on the combination of models/calibrations that matches your prior belief in those parameters. 

Unfortunately, I cannot make firm recommendations about how to parameterize your calibration densities. It really isn't a satisfactory way to estimate node ages with fossils, anyway. We have developed a better method which integrates the calibration ages into the birth-death model (http://www.slideshare.net/trayc7/heath-evolution-2013). However, it's not available yet. Hopefully, we'll have a version out this fall.

Cheers,

Tracy

Nick

--
You received this message because you are subscribed to the Google Groups "dppdiv-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to dppdiv-users...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.
 
 

[Nick L.]

Thanks so much, Tracy.