Incorporating individual samples when modelling oneself, or exclusively averages?

[bovefex]

Hello all,

I recently got into G25 modelling and was wondering whether it is ok to use individual samples in correspondence with averages, for example when trying to find out minute amounts of ancestry from a far away population.

Thanks in advance!

[randwulf]

Hello all,

I recently got into G25 modelling and was wondering whether it is ok to use individual samples in correspondence with averages, for example when trying to find out minute amounts of ancestry from a far away population.

Thanks in advance!

People develop different work patterns. I use all individuals, usually. Sometimes, I do model with averages. I don't usually mix them. I guess that would work, but I am not sure if there are pitfalls to it.

[ph2ter]

My experience is that with individual samples you can get better, more accurate results, but you must take into account obvious outliers which exist in the spreadsheet.
Someone from Germany (or any other nationality or ethnicity) sometimes does not look so.

[JerryS.]

My experience is that with individual samples you can get better, more accurate results, but you must take into account obvious outliers which exist in the spreadsheet.
Someone from Germany (or any other nationality or ethnicity) sometimes does not look so.

I found this to be the case for me, particularly with my north shifted German.

[bovefex]

Do you guys then also use each and every individual sample from a population, or only specific ones? Because some samples gives no results, others give weird results and some give results that make sense, at least in my experience.

[BalkanKiwi]

Do you guys then also use each and every individual sample from a population, or only specific ones? Because some samples gives no results, others give weird results and some give results that make sense, at least in my experience.

Depending on what I'm looking for I'll use the individual sample that I fit best with if I'm using individual samples, however I also run these through a calculator first to make sure they represent the majority of my ancestral component (and aren't going to overfit/underfit). An example for me is I don't use the Maori population when modelling, as its 50/50 Polynesian and Western Euro, and sometimes I score more of it than I should because of this. I always use a Papuan population instead.

[bovefex]

An example for me is I don't use the Maori population when modelling, as its 50/50 Polynesian and Western Euro, and sometimes I score more of it than I should because of this. I always use a Papuan population instead.

Kind of unrelated, but Maori are ~50% Western European?

[BalkanKiwi]

Kind of unrelated, but Maori are ~50% Western European?

The particular Maori sample in the G25 is. Its normal for modern Maori to be mixed with something (for example Croatian due to Yugoslav migrations), however you can still get Maori who are >60% Polynesian. The North Island of New Zealand for example, has had a lot of migration from Samoa, Tonga and Fiji, and have mixed with Maori.

[randwulf]

Do you guys then also use each and every individual sample from a population, or only specific ones? Because some samples gives no results, others give weird results and some give results that make sense, at least in my experience.

I tend to use a regional group of references in full for convenience. This means my reference groups tend to be proxies at times, sort of the way BalkanKiwi is describing. My example: my dad models well as German and Ashkenazi and I can use all of the individuals in his model and it works very well. My mother is South German and Isles, but if I include German in the model, some of the northerly Germans in the set of individuals overlap with some of the English and the percentages become odd. But, if I use other regional Germanic types, like Swiss German or Swiss French, it works very well for her. So, you have to consider mixed or outliers in the set as well as the considerable overlap between some regions.

[Nino90]

I tend to use a regional group of references in full for convenience. This means my reference groups tend to be proxies at times, sort of the way BalkanKiwi is describing. My example: my dad models well as German and Ashkenazi and I can use all of the individuals in his model and it works very well. My mother is South German and Isles, but if I include German in the model, some of the northerly Germans in the set of individuals overlap with some of the English and the percentages become odd. But, if I use other regional Germanic types, like Swiss German or Swiss French, it works very well for her. So, you have to consider mixed or outliers in the set as well as the considerable overlap between some regions.

Well explained.

If I use the individual Swedish samples, It works best when choosing the ones who fit closest to my mother.

Some of the Swedish samples are much more southern/continental shifted, and others have part Finnic and Siberian admixture.

Target: Nino_scaled
Distance: 1.8542% / 0.01854211
66.6 Swedish
25.0 Italian_Tuscany
8.4 Saami

 

Target: Nino_scaled
Distance: 1.8542% / 0.01854211
53.8 Swedish:Sweden10
21.4 Italian_Tuscany:NA20504
12.8 Swedish:Sweden12
8.4 Saami:saami6
3.6 Italian_Tuscany:VO59

VS

All Swedish samples:

Target: Nino_scaled
Distance: 1.7974% / 0.01797445
69.2 Swedish
22.8 Italian_Tuscany
8.0 Saami

 

Target: Nino_scaled
Distance: 1.7975% / 0.01797522
27.6 Swedish:Sweden10
20.6 Swedish:Sweden17
18.2 Italian_Tuscany:NA20504
7.6 Swedish:Sweden10
6.0 Swedish:Sweden18
4.8 Saami:saami6
4.6 Italian_Tuscany:Tuscany27
4.0 Swedish:Sweden12
3.6 Swedish:Sweden16
2.4 Saami:saami9
0.6 Saami:saami3

Not an extreme change in percents. But notice that the other option add way more samples than first model.

Vahaduo_ Global 25 Views (11).png

According to this PCA you see some Swedish samples are very Uralic shifted and others close to Germans.

It is also very easy to manipulate your results:

Here I just choosed the most Uralic shifted sample:

Target: Nino_scaled
Distance: 3.0525% / 0.03052475
46.0 Swedish
42.2 Italian_Tuscany
11.8 Saami

Target: Nino_scaled
Distance: 3.0525% / 0.03052475
46.0 Swedish:Sweden15
42.2 Italian_Tuscany:VO59
11.8 Saami:saami6

Of course increase the South European because of lower levels of EEF I would guess.
Also notice the bad fit.

Also I noticed that I most often get modeled with two of the most "north" shifted Tuscan samples, witch make sense because my Italian ancestors were all from North Tuscany.