Mega QAPDM Models thread

[Grace O'Malley]

I haven't had much luck using two ancient Northwest Europeans as sources. Even if I use a third "Southern" population. It does the same thing you are seeing. I guess qpAdm struggles to figure out where the Steppe is coming from.

This one is working for me. Scotland_Viking is just a Gael so this appears good. More realistic as well. Irish are mostly Gael with a little bit of Germanic.

[Gannicus]

This one is working for me. Scotland_Viking is just a Gael so this appears good. More realistic as well. Irish are mostly Gael with a little bit of Germanic.

I'm not satisfied with the standard errors that I get. I might be able to experiment with different references to fix it. I noticed Drb was able to model with two or more NW Euros and get decent standard errors. He said it involved using related Iron Age and Bronze Age populations in the references. I'm not sure if he kept the older West Eurasian populations in his references though such as AHG (Pinarbasi), EHGs etc.



Edit: This one isn't too terrible for SEs but still higher than I'd like. Especially since the Gallo-Roman admixture isn't small. I hear that a higher standard error is more excusable for smaller admixture percentages like 5-10% due to more uncertainty, because it's small.

[Grace O'Malley]

I'm just not satisfied with the standard errors that I get. I might be able to experiment with different references to fix it. I noticed Drb was able to model with two or more NW Euros and get decent standard errors. He said it involved using related Iron Age and Bronze Age populations in the references. I'm not sure if he kept the older West Eurasian populations in his references though such as AHG (Pinarbasi), EHGs etc.

So that bit at the side with percentage is standard errors 21.38%? Yes most likely the references are the key.

[Gannicus]

So that bit at the side with percentage is standard errors 21.38%? Yes most likely the references are the key.

Yes, that plus or minus X% to the left of the admix bar is the standard error. The interface doesn’t show it but there is another metric with these models and that is Z-Scores. The Z-score is your admixture divided by the standard error. Some hobbyists think that a Z-score of less than 3 means you don’t have that source. Through email correspondence with Andrei I don’t think that’s correct. And the reason I think that is because there have been published qpAdm admixture models in academic papers with Z scores less than 3. Btw I learned this stuff with the help of AndreiDNA, plus looking around myself and discussions on another forum.


This model would satisfy that hobbyist that held the belief that Z-Scores of 3 or better should only be considered:



My 20.55% admixture of EGY TIP divided by the standard error of 6.32% would equal a Z score of ~3.25.


It seems the general approach in peer-reviewed work is p-value + archaeological plausibility* > strict Z cutoff.

*I think that this can be interpreted with nuance as well

[Gannicus]

What reference populations did you use? I’m still trying to figure out which ones work best for modern samples, so far my results have been a bit inconsistent.

For this model, I used these 30 references:

[Xavier727]

For this model, I used these 30 references:

Thanks, mate, any idea why the Human Origins dataset is so limited, there's only like two Germanic populations to choose from, it seems a bit restrictive, which could affect reliability. I’ve been getting some weird FST results, when using Scottish as the target, French, Spanish and Hungry are showing closer distances than Orcadian.

[Gannicus]

Thanks, mate, any idea why the Human Origins dataset is so limited, there's only like two Germanic populations to choose from, it seems a bit restrictive, which could affect reliability. I’ve been getting some weird FST results, when using Scottish as the target, French, Spanish and Hungry are showing closer distances than Orcadian.

That is one of my criticisms of that dataset is the lack of Germans. No Germans, Austrians etc. Have you tried doing FST with just individuals from a population?

[Fistora]

[Xavier727]

That is one of my criticisms of that dataset is the lack of Germans. No Germans, Austrians etc. Have you tried doing FST with just individuals from a population?

There aren’t very many Celtic populations either, which isn’t ideal since it could probably skew the results. I haven’t gone through the individual samples yet, I’ll check them now

[Fistora]

I wonder where the russian.ho samples are from. Is it just the western part of Russia?