Dadi performing and time estimation

[Camille Voisin]

Dear Ryan,

I have a question regarding the time parameter obtained with dadi.

Just to summarise my analysis :

1/ I had compared 9 models to test the demographic evolution of a plant. 

2/ After obtaining my best model, I ran a goodness of fit analysis with 100 replicats to check the validity of my best model. 

3/ Thanks to this last, I also get the distribution of the parameters and use it to convert time into “real” biological meaning.

But I’m in trouble with one point. Because in previous analysis I obtained a really badly supported phylogenetic tree with more or less no structure. But if I trust the distribution of time parameter, I obtain a really old and large estimation.

I’m not surprised by the large confidence interval (between 80 to 5000 kya), but I'm wondering how dadi could find a really old split (mean = 1700 kya), while the phylogenetic tree showed nothing.

Do you have an explanation why dadi performed so well compared to phylogenetic analysis?

Thanks in advance,

Camille

[Ryan Gutenkunst]

Hello Camille,

The large confidence interval does suggests there’s not a lot of information about a potential split in the data. What are the times you infer in genetics units (before converting to “real” units). That’s usually more informative about the population genetics context.

You could carry out a formal composite likelihood ratio test using dadi, comparing models with and without and split. That would be a more formal way to test. Even though your confidence interval doesn’t include zero, you might be on the verge of not significant in a likelihood ratio test.

Best,
Ryan

> --
> You received this message because you are subscribed to the Google Groups "dadi-user" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to dadi-user+...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/dadi-user/a1680e57-d0cd-439d-8ad5-b301763a9297n%40googlegroups.com.

[Camille Voisin]

Dear Ryan,

Thanks for your answer.

First, the times use to infer the genetics unit is : 

T1 : 3.021 [CI95: 0.06-9.22]

T2 : 0.684 [CI95: 0.113-1.726]

I could also interpret it as an old split because of the selection of a model with two epochs.

Secondly, you said : “comparing models with and without a split”.

In my case, I didn’t add any model of no split in my model selection because my species occurs currently in two areas deeply separated and forms two genetic groups. I try this dadi method to infer if the split is recent (after Last Glacial Maximum) or old (before).

Do you have any suggestions regarding this information?

Thanks in advance,

Best,

Camille

[Ryan Gutenkunst]

Hello Camille,

My apologies for the slow reply. Given your phylogenetic results, it may be worth trying a no-split model in dadi. It may be that you have high gene flow, which makes the populations close to panmictic.

That T1 is remarkably old in genetic units. But the CI is obviously huge. I assume you see big uncertainties in the migration parameter as well, with high covariance between migration and divergence time.

Best,
Ryan

> To view this discussion on the web visit https://groups.google.com/d/msgid/dadi-user/97359c72-82bb-46af-9af1-b2febc8edb00n%40googlegroups.com.

[Camille Voisin]

Dear Ryan, 

Thanks for your answer,  and thank you for taking the time to reply to all questions!

I will try by adding a non split model because, as you said, gene flow seems to be really important between my groups.

Anyway, in all my model tests, T1 is really old in genetic units. Even with the uncertainty, it’s a great indication for my study.

All the best,