nMonte3: Iron Age for NW/W Europe

**MitchellSince1893** · 06-23-2022, 02:40 AM

Originally Posted by MitchellSince1893

Post 323 results. Again my pie chart is most like English_North and German_NW

I need to correct that. Actually 323 result makes me look more Scottish and to a lesser extent Southern English...areas that have almost and even split of 40 something percent for Briton & West-Germanic

If you took the Scotland result and replaced the Nordic with SlavicEU and replaced Hispani with Gallic, Scotland would be a good fit for my result (my pie chart is in North Sea)

Areas with some Gallic in Yellow, some SlavicEU in blue, and overlap in dark green.

**Huijbregts** · 06-23-2022, 06:50 AM

Originally Posted by Telfermagne

Word

just found this nugget: https://towardsdatascience.com/linea...d-f88be6c1e00b

With the problem of number of samples to the side, in the following instance the definition of 'groups' are based upon their respective study, so they're independent of the LDA:

In PAST when I use these guys Axis 1 seems to account for 83.59% of the variation, while Axis 2 seems to account for 16.41%. The error matrix shows 98.25% accuracy (jacknifed), all Salme Ship remains were correctly assigned, all Imperial Romans were correctly assigned, 32 out of 34 Arras remains were correctly assigned (2 were mistaken for Imperial Roman).

So, within the context of a k=3 model these would be satisfactory?

For several reasons I have my doubts.
IMO your model is only valid within the condition that your three components East Yorkshire MIA_LIA, VK2020_EST_Saaremaa_EVA and ITA_Rome_Imperial really are the three separate components of the West-European submodern variation.
If the first dimension accounts for 83.59% of the variation, my hunch is that your model is more or less one-dimensional.
An accuracy of 98.25% appears incredible for a 2-dimensional model, especially with so few data.
That all Salme samples are correctly assigned makes me really suspicious.
As a sanity check I suggest that you fit a kmeans model with k = 2,3,4 on the aggregate of your 3 three groups.
But even if the model is valid, I think that is missing too many admixtures to be representative for all West-Europeans.

**jadegreg** · 06-23-2022, 11:34 AM

Originally Posted by Huijbregts

For several reasons I have my doubts.
IMO your model is only valid within the condition that your three components East Yorkshire MIA_LIA, VK2020_EST_Saaremaa_EVA and ITA_Rome_Imperial really are the three separate components of the West-European submodern variation.
If the first dimension accounts for 83.59% of the variation, my hunch is that your model is more or less one-dimensional.
An accuracy of 98.25% appears incredible for a 2-dimensional model, especially with so few data.
That all Salme samples are correctly assigned makes me really suspicious.
As a sanity check I suggest that you fit a kmeans model with k = 2,3,4 on the aggregate of your 3 three groups.
But even if the model is valid, I think that is missing too many admixtures to be representative for all West-Europeans.

When you previously referred to 'cherry picking', do you mean removal of samples to get clear separation of K, or groups within the model? This of course is a post hoc selection. Therefore, would it be acceptable to use a priori 'cherry picking'. For example, let's say we would like a 'Migration Period' group, that included Bauvarii, Allemanii, Longobards and Saxons. However, we know some samples are heavily admixed with a different populations. Would it be reasonable to exclude these hypothetically non-representative samples from your initial grouping for the 'Migration Period' group?

**Anglecynn** · 06-23-2022, 02:09 PM

For post 323 it prefers to take away from West Germanic and make it a combination of Hispani, Slavic and Nordic instead, but the Briton is still quite similar. It's an odd mix compared to the previous one so not sure where it fits 'on the map'.

Post 293:

Distance: 1.5141% / 0.01514093
54.4 Briton
24.8 West-Germanic
13.4 Gallic
7.4 Nordic

My mother scores a bit of Balto-Finnic in this one:

Distance: 1.8765% / 0.01876547
54.4 Briton
25.0 West-Germanic
19.0 Gallic
1.6 Balto-Finnic

Post 323:

Distance: 1.3568% / 0.01356831
60.4 Briton
13.6 West-Germanic
9.2 Hispani
8.8 Slavic
8.0 Nordic

My mother scores quite a lot of Slavic:

Distance: 1.6163% / 0.01616282
65.0 Briton
17.0 Slavic
7.6 Hispani
7.2 West-Germanic
3.2 Gallic

**Huijbregts** · 06-23-2022, 03:34 PM

Originally Posted by jadegreg

When you previously referred to 'cherry picking', do you mean removal of samples to get clear separation of K, or groups within the model? This of course is a post hoc selection. Therefore, would it be acceptable to use a priori 'cherry picking'. For example, let's say we would like a 'Migration Period' group, that included Bauvarii, Allemanii, Longobards and Saxons. However, we know some samples are heavily admixed with a different populations. Would it be reasonable to exclude these hypothetically non-representative samples from your initial grouping for the 'Migration Period' group?

There is an ex-ante selection of groups to get a better separation between the groups. This results in an unnatural high accuracy of the results, which does not necessarily prove an extremely good fit of the model, but just a good separation of the (possibly erroneously estimated) groups.
An ex-post selection of the groups would happen if you think: I am not quite pleased with the result, let's change the groups so as to get a better fit. This is of course a no go.
In #354 Telfermagne used the term 'cherry picking' for the 100% accuracy he got when he used the option 'jacknifing' in his LDA. This is a technical question which a will not deal with here. But a 100% percent accurate estimate is of course very suspect.

I think it is a bad idea to introduce a 'Migration Period' group. Indeed many samples are heavily admixed with a different population. Maybe you can think of them as cultural groups, but not as genetic groups. Many of these samples are more Asian than European. But the Saxons are not Asian.
However, the Sarmatians plus Cimmerians are relatively homogenous.
IMO excluding hypothetically non-fitting samples from an initial grouping for the 'Migration Period' group is clearly an ex-post change of the model, which is not admitted.
I actually prepared a selected-LDA model with a Sarmatians_like group. Maybe you are interested.
combi3.png

**Telfermagne** · 06-23-2022, 03:38 PM

Originally Posted by Huijbregts

For several reasons I have my doubts.
IMO your model is only valid within the condition that your three components East Yorkshire MIA_LIA, VK2020_EST_Saaremaa_EVA and ITA_Rome_Imperial really are the three separate components of the West-European submodern variation.
If the first dimension accounts for 83.59% of the variation, my hunch is that your model is more or less one-dimensional.
An accuracy of 98.25% appears incredible for a 2-dimensional model, especially with so few data.
That all Salme samples are correctly assigned makes me really suspicious.
As a sanity check I suggest that you fit a kmeans model with k = 2,3,4 on the aggregate of your 3 three groups.
But even if the model is valid, I think that is missing too many admixtures to be representative for all West-Europeans.

Schwifty, wish I had asked these questions a long long time ago

When I do add more realistic groups the Salme accuracy does decrease, as does that for East Yorkshire MIA_LIA and the number of axis increase on the LDA readout (4 Axis instead of 2).

**Telfermagne** · 06-24-2022, 03:39 AM

Did a Kmeans for North-Central_Langobards n = 26, Imperial Romans n = 48, Britain Late Iron Age n = 34, and Gaul (Fench IA2 samples) n = 15. I used 'North-Central Langobards' because Understanding 6th-century barbarian social organization and migration through paleogenomics made a recurrent point to discern between them as a group of 'Northern ancestry' distinct from the other burials. As far as sample numbers, hands are kinda tied with what's in the dataset.

The Kmeans CSV is in the spoiler tags. 4 clusters were proposed.

Cluster 1 has n = 25: 12 were 'Gaul', 7 were 'North-Central Langobard', 4 were 'Britain Late Iron Age' and 2 were 'Imperial Roman'.
Cluster 2 has n = 46: all 46 were 'Imperial Roman'.
Cluster 3 has n = 23: 10 were 'North-Central Langobard', 9 were 'Britain Late Iron Age', and 3 were 'Gaul'.
Cluster 4 has n = 29: 21 were 'Britain Late Iron Age' and 8 were 'North-Central Langobard'.

If I'm interpreting correctly, then that means that there's no grounds to discern 'clusters' within this selection, at least in terms of Britain Late Iron Age, North-Central Langobard, and Gaul.

**Telfermagne** · 06-24-2022, 05:27 AM

Let's give this a try.

Started with selecting all samples that were England_MIA, all IA_La_Tene samples (La Tene_o and low coverage were removed), all Baltic Bronze Age samples (LVA, EST, LTU), all samples prefixed with 'Sarmatian_', all VK2020 samples that were suffixed with _EVA, and all Rome_Late_Antiquity samples.

Proposed k = 6

LDA error matrix, jacknifed, yields 89.17% accuracy:
Attachment 50209

There were 4 Axis: Axis 1 accounted for 48.3%, Axis 2 accounted for 40.61%, Axis 3 accounted for 7.692%, Axis 4 accounted for 2.843%

I did a Kmeans test, and proposed 6 clusters. Accuracy was determined by dividing the total in a given cluster by the group majority; e.g.) Cluster 5 was 53 samples, roughly 75% of which were Early Viking, therefore I labelled 'cluster 5' Early Viking but while keeping in mind that it's only 75% accurate. In the case of Rome Late Antiquity, even though all the samples in cluster 2 are 'Roman Late Antiquity' it only accounts for 16 out of the 24 Roman Late Antiquity samples included, so I'm inclined to say that it's only 66.66% accurate.

Outcome:
Sarmatian 100%
Baltic Bronze Age 100%
England Middle Iron Age 89%
Rome Late Antiquity 66.66%
Early Viking 75%
La Tene 87.75%

The CSV is in spoiler tags:

PCA with group averages:
Attachment 50208

Model:

,PC1,PC2,PC3,PC4,PC5,PC6,PC7,PC8,PC9,PC10,PC11,PC1 2,PC13,PC14,PC15,PC16,PC17,PC18,PC19,PC20,PC21,PC2 2,PC23,PC24,PC25
Britain_Middle_Iron_Age,0.129948033,0.135595689,0. 060544644,0.047096989,0.040103078,0.016293422,0.00 2172444,0.004925422,0.007587844,0.005888244,-0.004970867,0.006729078,-0.014942056,-0.0153847,0.021486033,0.006408522,-0.005032833,0.002073478,0.001625678,0.003621189,0. 006135,0.004178156,-0.001787111,0.005566922,-0.0016192
Rome_Late_Antiquity_Average,0.113016875,0.14877504 2,0.005735458,-0.029837125,0.023876167,-0.012399,-0.000450417,-0.002663292,0.006791958,0.023903292,0.0014005,0.00 502675,-0.00834975,-0.003107958,-0.004552292,-0.000491708,0.004286458,0.0005385,0.003273458,-0.00149025,-0.000993042,0.001607542,-0.000919208,0.000261083,-0.001247375
La_Tène_Average,0.128497523,0.140768015,0.05396306 2,0.027380462,0.042346323,0.007658769,0.002433169, 0.003617569,0.009537108,0.0144892,-0.003610031,0.005150785,-0.012112415,-0.006425846,0.004865046,0.003663523,0.000684031,0. 001489046,0.002423062,0.002074077,0.002188477,0.00 3431785,-0.001647692,-0.000598769,-0.001803585
Early_Viking_Average,0.127400571,0.127400643,0.072 775286,0.062846571,0.044169333,0.022849071,0.00501 3429,0.009917238,0.004236595,-0.008946929,-0.003282524,0.00131319,-0.003543095,-0.002680405,0.017889214,0.00752919,-0.006630952,0.001918405,0.004339595,0.004913119,0. 007974071,0.002316976,0.001103286,0.011079976,0.00 1787714
Sarmatian_Average,0.109595343,0.044886429,0.036440 743,0.065522857,-0.025411143,0.024773514,0.001027286,-3.95429E-05,-0.027113914,-0.035645429,-0.003646771,-0.000509486,-0.000696514,-0.0215438,0.0208272,0.012425629,-0.0043362,0.000347543,-0.001081,0.000432286,-0.009579514,0.001953686,0.000450743,0.0056324,-0.001645657
Baltic_Bronze_Age_Average,0.131167476,0.124378095, 0.098823571,0.109728476,0.049195857,0.039974381,0. 014470143,0.018867286,-0.001928333,-0.049646333,0.000085,-0.018490714,0.032677048,0.040205476,-0.016713095,0.001092333,0.005010333,-0.000699667,-0.001268905,0.005603905,-0.00258481,-0.005093333,0.009073429,-0.019245333,0.003084857

Results:
[1] "distance%=2.0827"

Dad

Britain_Middle_Iron_Age,72.2
Early_Viking_Average,21.8
La_Tène_Average,4.8
Rome_Late_Antiquity_Average,1.2

[1] "distance%=2.2292"

Mom

La_Tène_Average,40.2
Britain_Middle_Iron_Age,37
Rome_Late_Antiquity_Average,13
Early_Viking_Average,9.8

[1] "distance%=2.0289"

Me

Britain_Middle_Iron_Age,75.4
La_Tène_Average,18.4
Rome_Late_Antiquity_Average,6.2

**Garimund** · 06-24-2022, 11:02 AM

Originally Posted by Telfermagne

Let's give this a try.

Started with selecting all samples that were England_MIA, all IA_La_Tene samples (La Tene_o and low coverage were removed), all Baltic Bronze Age samples (LVA, EST, LTU), all samples prefixed with 'Sarmatian_', all VK2020 samples that were suffixed with _EVA, and all Rome_Late_Antiquity samples.

Proposed k = 6

LDA error matrix, jacknifed, yields 89.17% accuracy:
Attachment 50209

There were 4 Axis: Axis 1 accounted for 48.3%, Axis 2 accounted for 40.61%, Axis 3 accounted for 7.692%, Axis 4 accounted for 2.843%

I did a Kmeans test, and proposed 6 clusters. Accuracy was determined by dividing the total in a given cluster by the group majority; e.g.) Cluster 5 was 53 samples, roughly 75% of which were Early Viking, therefore I labelled 'cluster 5' Early Viking but while keeping in mind that it's only 75% accurate. In the case of Rome Late Antiquity, even though all the samples in cluster 2 are 'Roman Late Antiquity' it only accounts for 16 out of the 24 Roman Late Antiquity samples included, so I'm inclined to say that it's only 66.66% accurate.

Outcome:
Sarmatian 100%
Baltic Bronze Age 100%
England Middle Iron Age 89%
Rome Late Antiquity 66.66%
Early Viking 75%
La Tene 87.75%

The CSV is in spoiler tags:

PCA with group averages:
Attachment 50208

Model:

,PC1,PC2,PC3,PC4,PC5,PC6,PC7,PC8,PC9,PC10,PC11,PC1 2,PC13,PC14,PC15,PC16,PC17,PC18,PC19,PC20,PC21,PC2 2,PC23,PC24,PC25
Britain_Middle_Iron_Age,0.129948033,0.135595689,0. 060544644,0.047096989,0.040103078,0.016293422,0.00 2172444,0.004925422,0.007587844,0.005888244,-0.004970867,0.006729078,-0.014942056,-0.0153847,0.021486033,0.006408522,-0.005032833,0.002073478,0.001625678,0.003621189,0. 006135,0.004178156,-0.001787111,0.005566922,-0.0016192
Rome_Late_Antiquity_Average,0.113016875,0.14877504 2,0.005735458,-0.029837125,0.023876167,-0.012399,-0.000450417,-0.002663292,0.006791958,0.023903292,0.0014005,0.00 502675,-0.00834975,-0.003107958,-0.004552292,-0.000491708,0.004286458,0.0005385,0.003273458,-0.00149025,-0.000993042,0.001607542,-0.000919208,0.000261083,-0.001247375
La_Tène_Average,0.128497523,0.140768015,0.05396306 2,0.027380462,0.042346323,0.007658769,0.002433169, 0.003617569,0.009537108,0.0144892,-0.003610031,0.005150785,-0.012112415,-0.006425846,0.004865046,0.003663523,0.000684031,0. 001489046,0.002423062,0.002074077,0.002188477,0.00 3431785,-0.001647692,-0.000598769,-0.001803585
Early_Viking_Average,0.127400571,0.127400643,0.072 775286,0.062846571,0.044169333,0.022849071,0.00501 3429,0.009917238,0.004236595,-0.008946929,-0.003282524,0.00131319,-0.003543095,-0.002680405,0.017889214,0.00752919,-0.006630952,0.001918405,0.004339595,0.004913119,0. 007974071,0.002316976,0.001103286,0.011079976,0.00 1787714
Sarmatian_Average,0.109595343,0.044886429,0.036440 743,0.065522857,-0.025411143,0.024773514,0.001027286,-3.95429E-05,-0.027113914,-0.035645429,-0.003646771,-0.000509486,-0.000696514,-0.0215438,0.0208272,0.012425629,-0.0043362,0.000347543,-0.001081,0.000432286,-0.009579514,0.001953686,0.000450743,0.0056324,-0.001645657
Baltic_Bronze_Age_Average,0.131167476,0.124378095, 0.098823571,0.109728476,0.049195857,0.039974381,0. 014470143,0.018867286,-0.001928333,-0.049646333,0.000085,-0.018490714,0.032677048,0.040205476,-0.016713095,0.001092333,0.005010333,-0.000699667,-0.001268905,0.005603905,-0.00258481,-0.005093333,0.009073429,-0.019245333,0.003084857

Results:
[1] "distance%=2.0827"

Dad

Britain_Middle_Iron_Age,72.2
Early_Viking_Average,21.8
La_Tène_Average,4.8
Rome_Late_Antiquity_Average,1.2

[1] "distance%=2.2292"

Mom

La_Tène_Average,40.2
Britain_Middle_Iron_Age,37
Rome_Late_Antiquity_Average,13
Early_Viking_Average,9.8

[1] "distance%=2.0289"

Me

Britain_Middle_Iron_Age,75.4
La_Tène_Average,18.4
Rome_Late_Antiquity_Average,6.2

Target: garimund_anc_scaled
Distance: 2.7366% / 0.02736611
72.4 Britain_Middle_Iron_Age
24.0 La_Tène_Average
3.6 Rome_Late_Antiquity_Average

Can you post this model using individual samples?

**lg16** · 06-24-2022, 11:31 AM

Target: lg_scaled
Distance: 1.3974% / 0.01397425
73.0 Britain_Middle_Iron_Age
27.0 La_Tène_Average

Distance to: lg_scaled
0.01657462 Britain_Middle_Iron_Age
0.02781736 La_Tène_Average
0.04260723 Early_Viking_Average
0.10364409 Rome_Late_Antiquity_Average
0.14041279 Sarmatian_Average
0.14057639 Baltic_Bronze_Age_Average

Thread: nMonte3: Iron Age for NW/W Europe

Thread Tools

Display

The Following 5 Users Say Thank You to MitchellSince1893 For This Useful Post:

The Following 5 Users Say Thank You to Huijbregts For This Useful Post:

The Following 4 Users Say Thank You to jadegreg For This Useful Post:

The Following 7 Users Say Thank You to Anglecynn For This Useful Post:

The Following 5 Users Say Thank You to Huijbregts For This Useful Post:

The Following 5 Users Say Thank You to Telfermagne For This Useful Post:

The Following 4 Users Say Thank You to Telfermagne For This Useful Post:

The Following 14 Users Say Thank You to Telfermagne For This Useful Post:

The Following 10 Users Say Thank You to Garimund For This Useful Post:

The Following 10 Users Say Thank You to lg16 For This Useful Post:

Similar Threads

Stable population structure in Europe since the Iron Age, despite high mobility

Eastern Europe G25: Iron Age - Medieval

Iranian Neolithic-like component in Bronze Age and Iron Age Southern Europe

Balkan Iron Gate Hunter-Gatherers (Hint to Possible E-L618 "E-V13" Entry to Europe?)

URBAN EXPERIENCES IN EARLY IRON AGE EUROPE: CENTRAL PLACES AND SOCIAL COMPLEXITY

Posting Permissions