What are your thoughts on unscaled coordinates?

[Aeolus]

I think they're kinda silly. 

I'm a Turk and with unscaled coordinates Phrygians and other Anatolians are very close to me distance-wise, which makes no sense because of my east Asian ancestry, which would strongly pull me away from their cluster.

On top of that it exaggerates my European admixture. Like I get modelled as half European and Half Anatolian, which again, makes no sense.

What are your takes on unscaled coordinates?

[AimSmall]

From the Global25 author himself, Davidski...

https://eurogenes.blogspot.com/2019/07/g...6102031007

The scaled coordinates produce more stable and realistic results most of the time. But use both and see what works better in each case in comparison to other methods and scientific literature.

July 13, 2019 at 2:32 AM

[Jalisciense]

They are better than scaled definitely, at least for me, my family, Iberians and LATAM friends.

I'd say that just in R+nMonte3 with its penalty works the best, because in Vahaduo the results are horrible and don't look like the papers.

[Aeolus]

They are better than scaled definitely, at least for me, my family, Iberians and LATAM friends.

I'd say that just in R+nMonte3 with its penalty works the best, because in Vahaduo the results are horrible and don't look like the papers.

Where can I access Nmonte? I've tried to in the past, but it seems to have shut down. At least that's what I could gather a year ago when I researched about it.

[Jalisciense]

They are better than scaled definitely, at least for me, my family, Iberians and LATAM friends.

I'd say that just in R+nMonte3 with its penalty works the best, because in Vahaduo the results are horrible and don't look like the papers.

Where can I access Nmonte? I've tried to in the past, but it seems to have shut down. At least that's what I could gather a year ago when I researched about it.

Davidski made a guide here, I just followed the steps:

https://eurogenes.blogspot.com/2019/07/g...2.html?m=1

[AimSmall]

They are better than scaled definitely, at least for me, my family, Iberians and LATAM friends.

I'd say that just in R+nMonte3 with its penalty works the best, because in Vahaduo the results are horrible and don't look like the papers.

Where can I access Nmonte? I've tried to in the past, but it seems to have shut down. At least that's what I could gather a year ago when I researched about it.

nmonte/nmonte3.R at master · tipirneni/nmonte · GitHub

[Aeolus]

From the Global25 author himself, Davidski...

https://eurogenes.blogspot.com/2019/07/g...6102031007

The scaled coordinates produce more stable and realistic results most of the time. But use both and see what works better in each case in comparison to other methods and scientific literature.

July 13, 2019 at 2:32 AM

BTW did Davidski ever explain what the difference between unscaled and scaled is and why the fits are so good with unscaled vs with scaled?

[Jalisciense]

From the Global25 author himself, Davidski...

https://eurogenes.blogspot.com/2019/07/g...6102031007

The scaled coordinates produce more stable and realistic results most of the time. But use both and see what works better in each case in comparison to other methods and scientific literature.

July 13, 2019 at 2:32 AM

BTW did Davidski ever explain what the difference between unscaled and scaled is and why the fits are so good with unscaled vs with scaled?

Unscaled are your real G25 values and scaled is a filtered version of them.

Scaling works by multiplying each coordinate by its squarerooted eigenvalue, so in high dimensions doesn't work fine (Unlike unscaled) because main differences between populations (Especially similar ones) are in the higher dimensions.

So I guess multiplying eigenvalues would increase the distance %.

[AimSmall]

I have so many notes in my collection in OneNote, but the links reference back to AG.  I like to source all my notes.

Anyways, I'll refer you to this thread with particular attention to Dr. Huijbregt's reply.  Note, he is the author of the previously mentioned nMonte3.R code.

https://genoplot.com/discussions/topic/2...-or-off/10

What's missing from genoplot's version is his graphic which I still retain.

   

[AimSmall]

These questions come up routinely over the years on scaled vs unscaled. Another thread I keep in my collection is this one.

https://www.theapricity.com/forum/showth...scaled-why&p=6833064&viewfull=1#post6833064

[Aeolus]

These questions come up routinely over the years on scaled vs unscaled.  Another thread I keep in my collection is this one.

https://www.theapricity.com/forum/showth...ost6833064

Thanks!

[Aeolus]

I have so many notes in my collection in OneNote, but the links reference back to AG.  I like to source all my notes.

Anyways, I'll refer you to this thread with particular attention to Dr. Huijbregt's reply.  Note, he is the author of the previously mentioned nMonte3.R code.

https://genoplot.com/discussions/topic/2...-or-off/10

What's missing from genoplot's version is his graphic which I still retain.

Interesting read. So basically unscaled does give more accurate results with better fits, while scaled often requires overfitting for it to give a proper fit?

That matches with my own experience as well tbh. I need a massive overfit to get a fit of 1.1 for instance on scaled.

The reason I'm so suspicious of unscaled is just the distances are very close. Like the distance to Greece_Logkas is like 3 for me, Turkic Central Asian is 4, which is both way too close. 10 seems to be more realistic.