Admixtools My qpAdm Results Please Share Yours!

[qh777]

Add Italy_Epigravettian.SG and replace bichon with it. Should help with your WHG percentage.

And then experimented with more references:

 

[qh777]

Here are some Iron and Bronze Age models:

And a few Medieval:

This one contains pooled Southern Medieval Italians

I chose to share this one instead of Spain Medieval because those produce much higher standard errors

 

[qh777]

France_Medieval_o.AG_I15027.AG is a medieval Breton

 

[qh777]

In this model and others that I have ran it looks to me like the iron age Germanic groups are absorbing some Celtic (insular&continental) ancestry. My guess is that its because the two groups have similar steppe/ANF/WHG admixtures and it's hard for qpAdm to assign where it's coming from. With other references, the standard errors get pretty high.

Using other references, the proportions shift at the expense of higher standard errors. Higher pvals though

 

[Jovialis]

In light of the of Recchia et al. 2025 these results are not only statistically robust, but archeologically appropriate.

There was indeed a trans-Adriatic migration into Italy in the BA-IA:

 

[qh777]

Subtle admixtures could be missed with this model, I think.

 

[Jovialis]

View attachment 19085
View attachment 19086

Subtle admixtures could be missed with this model, I think.

I ran f4(Left_i, Left_j; Right_k, Right_l) grid with the Left panel vs the right panel of many distal sources, and it seems like this model may be doomed to fail (13% SE is high). I think the reason is that they are too low in coverage, as the F2 extract filters the SNP count from 500K+ to a mere 10K. They may also be too closely related, Idk.

You could probably still do qpAdm the way you are doing, but if you set up Admixtools2 with AADR, you could run this test to find out your most informative (asymmetrical) outgroups according to the left groups you are testing.

 

[Stefano]

In light of the of Recchia et al. 2025 these results are not only statistically robust, but archeologically appropriate.

There was indeed a trans-Adriatic migration into Italy in the BA-IA:

View attachment 19082

View attachment 19080
View attachment 19083
View attachment 19081

May i suggest you to try Croatia_Cetina, Montenegro_mlba, or Albania_ia? It surely is more historically accurate. You can also try the singleton Italy_IA_Republic_o (R1)who has IBD with illyrians and clusters with them. Be sure to check the groups with g25/PCA before using them because the labeling from AADR often contains outliers.

 

[Jovialis]

May i suggest you to try Croatia_Cetina, Montenegro_mlba, or Albania_ia? It surely is more historically accurate. You can also try the singleton Italy_IA_Republic_o (R1)who has IBD with illyrians and clusters with them. Be sure to check the groups with g25/PCA before using them because the labeling from AADR often contains outliers.

I recall trying some of the others, but could double check later when I have some time, at any rate,

I tried Montenegro_MLBA.AG, however there's issues with that sample.
Filtering in f4(Left_i, Left_j; Right_k, Right_l) causes massive collapse of SNP data:

As a result, the there are barely any asymmetrical informative outgroups in the panel for the model, which isn't enough for a qpAdm run:

I'm aware of the occasional mislabeling, and created some corrections myself. However, if test with PCA it will be with smartpca, as g25 is erroneous due to it's dependency on the pre-determined reference set (25 dimensions flattening complex genetic data in values, instead of reading polymorphic SNPs), and follows an unknown methodology. I like prefer smartpca, because g25 is derivative of it. It is kind of like buying a Chef Boyardee sauce vs making your own, knowing exactly what goes into it. It is serviceable, but who know what is going into it, and if it is even good for you.

 

[Jovialis]

Cetina_MBA is a flop for me, but perhaps a third component would suffice.

 

[Jovialis]

Ah ha! I found an issue, because my f4(Left_i, Left_j; Right_k, Right_l) is a later version of the pervious one I was using.

I had a practical fix where I set maxmiss to 0.2 because of the aggressive filtering. But apparently I can keep it set to 0, which is most conservtive, while making qpfstats = TRUE . I'll modify my f4(Left_i, Left_j; Right_k, Right_l) to see how it plays out.

Edit: Unapproved the f4 thread in the mean time, computation seems to be very slow with this setting, I will have to optimize the processing with the n_cores used, I will re-approve once I have it set.

Edit 2: I've updated the script, be advised it is slow, I will provide a version that takes advantage of n_cores, but that is idiosyncratic to your processor, and you to manually enable it exploitation.

 

[qh777]

View attachment 19085
View attachment 19086

Subtle admixtures could be missed with this model, I think.

I ran f4(Left_i, Left_j; Right_k, Right_l) grid with the Left panel vs the right panel of many distal sources, and it seems like this model may be doomed to fail (13% SE is high). I think the reason is that they are too low in coverage, as the F2 extract filters the SNP count from 500K+ to a mere 10K. They may also be too closely related, Idk.

You could probably still do qpAdm the way you are doing, but if you set up Admixtools2 with AADR, you could run this test to find out your most informative (asymmetrical) outgroups according to the left groups you are testing.

View attachment 19087

You have to keep in mind those 2 populations are not that far apart genetically. qpAdm will have issues distinguishing them. France Yonne IA and Denmark IA are both IA NW Europeans. Z-Score is over 3 also. Which even then Z scores don't have to be 3 or better to validate a model. The crazy high p-value, the fact that both are not too far genetically distant make this model not a problem to me. It all depends. If the p-value just barely passed (p=0.05) with those standard errors, then I would agree. Not an ideal model.

I have modeled myself with a larger reference list. And with most cases, the SEs go down as a result.

 

[Jovialis]

View attachment 19085
View attachment 19086

Subtle admixtures could be missed with this model, I think.

You have to keep in mind those 2 populations are not that far apart genetically. qpAdm will have issues distinguishing them. France Yonne IA and Denmark IA are both IA NW Europeans. Z-Score is over 3 also. Which even then Z scores don't have to be 3 or better to validate a model. The crazy high p-value, the fact that both are not too far genetically distant make this model not a problem to me. It all depends. If the p-value just barely passed (p=0.05) with those standard errors, then I would agree. Not an ideal model.

I have modeled myself with a larger reference list. And with most cases, the SEs go down as a result.

View attachment 19094
View attachment 19095

Yeah, I was going to mention that as well.

qpadm isn't really meant for very close left groups.

But regarding the right groups, you should endeavor to find the most asymmetrical and least redundant to verify the panel isn't just working because it is conditioning the results to pass.

It is also important to know if the polymorphic SNP count is sufficient if you are using F2_extract because a low SNP count could create erroneous results. I recall about 100k after filtering is sufficent with aDNA.

 

[qh777]

Yeah, I was going to mention that as well.

qpadm isn't really meant for very close left groups.

But regarding the right groups, you should endeavor to find the most asymmetrical and least redundant to verify the panel isn't just working because it is conditioning the results to pass.

OpenAI thought it was an alright list of references. Although it can make mistakes. Those other 2 models were done with different right groups:

If you'd like you can suggest a reference list for me to test those sources with. I always like to experiment.

 

[Jovialis]

Sure thing, just tell me what left groups you would like to test.

I know, I have used AI as well. But the advise it was giving can be wrong, especially with highly technical stuff like that. When I uploaded Harney et al. 2021 supplement, it would say stuff contrary to what it state. Then I would admonish it, about it, and it would say stuff like "Oh, sorry you were right!"

 

[qh777]

Sure thing, just tell me what left groups you would like to test.

I know, I have used AI as well. But the advise it was giving can be wrong, especially with highly technical stuff like that. When I uploaded Harney et al. 2021 supplement, it would say stuff contrary to what it state. Then I would admonish it, about it, and it would say stuff like "Oh, sorry you were right!"

Denmark IA + France Yonne IA like I did. Also, IA NW Euro + East Med. I really enjoy testing out people's reference lists!

 

[Jovialis]

Denmark IA + France Yonne IA like I did. Also, IA NW Euro + East Med. I really enjoy testing out people's reference lists!

Here's the Denmark IA and France IA run

I should probably remove mbuti from the testing list, because I wasn't sure if it should apply here or not. At any rate. Always keep it as 1st in the outgroups.

Try the following:

Mbuti.DG,
Russia_Sunghir_UP.SG
Russia_Yuzhniyoleniyostrov_Mesolithic.AG
Turkey_Marmara_Barcin_N.AG
Italy_Epigravettian.AG.BY.AA
Morocco_Iberomaurusian.AG

Then experiment with swapping around the informative_rights.

This is improved f4(Left_i, Left_j; Right_k, Right_l) with qpfstats-based extraction to handle high missingness pretty computationally, and time demanding (took about 20 minutes or more), so I unfortunately can only offer you this for now.

 

[Jovialis]

Here's a preview I tried with English.HO

 

[Jovialis]

Here is a better sample that works with it GBR.DG with is the British cohort.

It worked better for it, with decent SE. (under 0.1)

 

[Jovialis]

above 2 Z is permissable, but 3 is optimal.