TMRCA calculations

[ironroad41]

I know that I can take a homogeneous set of data such as the Ian Cam, Kerchners family or even a larger set like M222 and get a TMRCA which is in concert with most other ways of knowing about these sets of data, historical et.al.

I also have learned that I have to use independent lines; I can't include the descendant of a person who has inherited the mutation, that overcounts, if the descendant has a new mutation than I include him in the data set but don't count the common mutation of his ancestor.  In the Ian cam, you can almost identify relationships by common mutations, especially if they are medium to slow mutators.

How do you implement this requirement when you just use a data set which has a common SNP, such as R-L21?  One way might to filter out "successful SNP's"  SNP's which have started lines which have had great success in propagation.

The point is that these lines will have less variance, than lines that go back to the R-L21 founder, and that will shorten the time estimate to the founder.  All tree lines are not necessarily equal.  Think of Markos work where he tried to predict the most probable tree back to the common ancestor.  He had individual lines.  In a tree that contains M222 such as R-L21 only one man, the founder of M222, traces back to L21.  We have a large convergence at M222, but we consider each line as ending as L21.  So all individual tree lines are not equal, and do not have had the chance for equal diversity.

So, I would argue that any SNP that has large sub SNP populations will have its age underestimated as the calculation is now made.  We want all the tree lines to be the same length and we want them to be independent.  JMHO

[Mikewww]

I know that I can take a homogeneous set of data such as the Ian Cam, Kerchners family or even a larger set like M222 and get a TMRCA which is in concert with most other ways of knowing about these sets of data, historical et.al.

I also have learned that I have to use independent lines; I can't include the descendant of a person who has inherited the mutation, that overcounts, if the descendant has a new mutation than I include him in the data set but don't count the common mutation of his ancestor.  In the Ian cam, you can almost identify relationships by common mutations, especially if they are medium to slow mutators.

How do you implement this requirement when you just use a data set which has a common SNP, such as R-L21?  One way might to filter out "successful SNP's"  SNP's which have started lines which have had great success in propagation.

This is why I bring up the concept of random sampling. It should be by cross-section though to make sure it is representative.

The point is that these lines will have less variance, than lines that go back to the R-L21 founder, and that will shorten the time estimate to the founder.  All tree lines are not necessarily equal.  Think of Markos work where he tried to predict the most probable tree back to the common ancestor.  He had individual lines.  In a tree that contains M222 such as R-L21 only one man, the founder of M222, traces back to L21.  We have a large convergence at M222, but we consider each line as ending as L21.  So all individual tree lines are not equal, and do not have had the chance for equal diversity.

So, I would argue that any SNP that has large sub SNP populations will have its age underestimated as the calculation is now made.  We want all the tree lines to be the same length and we want them to be independent.  JMHO

Intraclade variance and TMRCA calculations do have this problem of bias caused by the larger groups within the sample.

I think the most useful work around is to use interclade methodologies at various levels of the Y DNA tree above and below the desired level of study.  If the two clades compared are somewhat similar in age, the precision of the estimate is greatly increased.  The biases are significantly reduced. I think of the interclade methodology as a way to filter out biases.

I ran a set of R-L21 subclade interclade TMRCAs and tried to display the results here.
http://tech.groups.yahoo.com/group/RL21Project/files/Haplogroup_R-L21_Subclades_MRCA_Timeline.jpg

I show it more generally in a different (branching) format here.
http://tech.groups.yahoo.com/group/RL21Project/files/Haplogroup_R-L21_Tree-Timeline.jpg

You'll see the L21/U152 and L21/Z196 estimates have very tight (more precise) one and two sigma ranges. You'll also see that their mid-points are slight younger than the midpoint for DF23/Irish II, the oldest of the sub L21 interclade pairs. This is not a problem because the error ranges all fit.  

Effectively you can start to corner in the TMRCAs for these subclades.  I think the use of random sampling can help reduce bias, but the interclade method is actually our best bet.

[Mikewww]

I don't have access to the Yahoo Project, you will have to tell me how to sign up using a hotmail account.....

I don't really like Yahoo but they provide a lot of free storage and have a structure set up.  Here is the registration site.  You might have to create a Yahoo ID but that doesn't mean you need to use it other than to be allowed into get the files. You can also "edit member" to "web only" so you get no emails.
https://edit.yahoo.com/registration?.src=ygrp&.intl=us&.lang=en-US&.done=http://tech.groups.yahoo.com/group/R-P312Project

[ironroad41]

I know that I can take a homogeneous set of data such as the Ian Cam, Kerchners family or even a larger set like M222 and get a TMRCA which is in concert with most other ways of knowing about these sets of data, historical et.al.

I also have learned that I have to use independent lines; I can't include the descendant of a person who has inherited the mutation, that overcounts, if the descendant has a new mutation than I include him in the data set but don't count the common mutation of his ancestor.  In the Ian cam, you can almost identify relationships by common mutations, especially if they are medium to slow mutators.

How do you implement this requirement when you just use a data set which has a common SNP, such as R-L21?  One way might to filter out "successful SNP's"  SNP's which have started lines which have had great success in propagation.

This is why I bring up the concept of random sampling. It should be by cross-section though to make sure it is representative.

The point is that these lines will have less variance, than lines that go back to the R-L21 founder, and that will shorten the time estimate to the founder.  All tree lines are not necessarily equal.  Think of Markos work where he tried to predict the most probable tree back to the common ancestor.  He had individual lines.  In a tree that contains M222 such as R-L21 only one man, the founder of M222, traces back to L21.  We have a large convergence at M222, but we consider each line as ending as L21.  So all individual tree lines are not equal, and do not have had the chance for equal diversity.

So, I would argue that any SNP that has large sub SNP populations will have its age underestimated as the calculation is now made.  We want all the tree lines to be the same length and we want them to be independent.  JMHO

Intraclade variance and TMRCA calculations do have this problem of bias caused by the larger groups within the sample.

I think the most useful work around is to use interclade methodologies at various levels of the Y DNA tree above and below the desired level of study.  If the two clades compared are somewhat similar in age, the precision of the estimate is greatly increased.  The biases are significantly reduced. I think of the interclade methodology as a way to filter out biases.

I ran a set of R-L21 subclade interclade TMRCAs and tried to display the results here.
http://tech.groups.yahoo.com/group/RL21Project/files/Haplogroup_R-L21_Subclades_MRCA_Timeline.jpg

I show it more generally in a different (branching) format here.
http://tech.groups.yahoo.com/group/RL21Project/files/Haplogroup_R-L21_Tree-Timeline.jpg

You'll see the L21/U152 and L21/Z196 estimates have very tight (more precise) one and two sigma ranges. You'll also see that their mid-points are slight younger than the midpoint for DF23/Irish II, the oldest of the sub L21 interclade pairs. This is not a problem because the error ranges all fit.  

Effectively you can start to corner in the TMRCAs for these subclades.  I think the use of random sampling can help reduce bias, but the interclade method is actually our best bet.

 My specific question is not how good or tight your SD is, but how do you know your answer is correct?  I pounded on VV and Ken about this.  Work 1 or 2 problems which have a known answer and see how your model does?  I can tell you the Ian Cam is a well studied data set and it would be interesting to me to see how your approach works.  For an intraclade estimate I would use the R1b of Clan Donald from Doug Mac donalds website.  I don't know if you have a data set in another Hg with similar properties?

[JeanL]

Mike what's the intraclade TMRCA mean and sigmas that you get for R-M153 using Ken's Gen7 Calculator on the 36 most linear STRs?

[Mikewww]

My specific question is not how good or tight your SD is, but how do you know your answer is correct?  I pounded on VV and Ken about this.  Work 1 or 2 problems which have a known answer and see how your model does?  

Ken's interclade spreadsheet is available for free from his web site. There are some instructions albeit they are limited.

He provides sigma (confidence) ranges with his result so you should have an understanding of the risks.

I can tell you the Ian Cam is a well studied data set and it would be interesting to me to see how your approach works.  For an intraclade estimate I would use the R1b of Clan Donald from Doug Mac donalds website.  I don't know if you have a data set in another Hg with similar properties?

Do we know that the Clan Donald data has as a most recent common ancestor?  What is terminal SNP, L21?  Same for Ian Cam?   How long ago did each live?   I think you are talking about subclades that are not marked by an SNP downstream of L21, correct?

[Jean M]

I know that I can take a homogeneous set of data such as the Ian Cam, Kerchners family or even a larger set like M222 and get a TMRCA which is in concert with most other ways of knowing about these sets of data, historical et.al.

Howard and McLaughlin have used a new method to obtain a date for M222 of between 1400 and 2000 BC. See A dated phylogenetic tree of M222 snp haplotypes.

[JeanL]

Effectively you can start to corner in the TMRCAs for these subclades.  I think the use of random sampling can help reduce bias, but the interclade method is actually our best bet.

If you could get me the data for the 36 most linear STRs that you are using for L21 I'd greatly appreciate it. The same thing for R-U152 would be good. I want to run a couple of experiments, just to test the variance.

By this I mean an excel file with 36 columns and x number of people that have tested for those 36 STRs.

[Mikewww]

Effectively you can start to corner in the TMRCAs for these subclades.  I think the use of random sampling can help reduce bias, but the interclade method is actually our best bet.

If you could get me the data for the 36 most linear STRs that you are using for L21 I'd greatly appreciate it. The same thing for R-U152 would be good. I want to run a couple of experiments, just to test the variance.

By this I mean an excel file with 36 columns and x number of people that have tested for those 36 STRs.

The information, formulas and whole bit are all there between the R-L21 and R-P312 Yahoo group file sections.  You can easily copy/paste from it.  I keep pretty busy just trying to keep it updated.  I post it as a spreadsheet so anyone can do whatever they want with it.

[ironroad41]

My specific question is not how good or tight your SD is, but how do you know your answer is correct?  I pounded on VV and Ken about this.  Work 1 or 2 problems which have a known answer and see how your model does?  

Ken's interclade spreadsheet is available for free from his web site. There are some instructions albeit they are limited.

He provides sigma (confidence) ranges with his result so you should have an understanding of the risks.

I can tell you the Ian Cam is a well studied data set and it would be interesting to me to see how your approach works.  For an intraclade estimate I would use the R1b of Clan Donald from Doug Mac donalds website.  I don't know if you have a data set in another Hg with similar properties?

Do we know that the Clan Donald data has as a most recent common ancestor?  What is terminal SNP, L21?  Same for Ian Cam?   How long ago did each live?   I think you are talking about subclades that are not marked by an SNP downstream of L21, correct?

[ironroad41]

As I recall, I had to split Clan Donald R1b into two groups, one akin to the Clan Gregor chief, and an older set.  Its apparent in the range of STR mutations each entry has.  I'd have to look at the data to give you my opinion.  As I recall, the younger group had a TMRCA of about 1100 AD.   So, off the top, this might not be such a good group to use.  I could spend some time with it if you have real interest in pursuing my question.

[ironroad41]

I know that I can take a homogeneous set of data such as the Ian Cam, Kerchners family or even a larger set like M222 and get a TMRCA which is in concert with most other ways of knowing about these sets of data, historical et.al.

Howard and McLaughlin have used a new method to obtain a date for M222 of between 1400 and 2000 BC. See A dated phylogenetic tree of M222 snp haplotypes.

I'll have to look that over next Winter;  its a complex paper which creates a new set of rules.  It is interesting that they refer to John McEwan, and his estimated of c. 3700 BP.I respect his work and wish he had stayed active in the field.

[Mark Jost]

KenN posted his latest Generations calculation spreadsheet for 111 markers. I managed to mod it to show TMRCA in years and broke out the various L21 Subclades sorted by coalescence Generations. I used 30 yrs per generation but fell 25 is would be more representative. Also, KenN reports that the mutation rates were from M Heinila.

SNP   YrsPerGen*   Count   AGE   Generations   YBP   Founder   Generations   YBP
L21All   30.0   N=648   GB coal=   119.5   3,585.8   GB=   133.3   3,997.9
DF21ALL   30.0   N=66   GA coal=   118.0   3,539.8   GA=   139.6   4,186.5
L253   30.0   N=15   GA coal=   97.6   2,927.3   GA=   126.4   3,790.6
DF21/Z246/DF25/DF5   30.0   N=11   GA coal=   94.2   2,825.6   GA=   127.1   3,812.3
L513All   30.0   N=62   GA coal=   91.2   2,734.9   GA=   106.8   3,204.8
L144   30.0   N=5   GA coal=   83.5   2,506.3   GA=   107.7   3,231.5
L253/L226   30.0   N=33   GA coal=   78.8   2,363.7   GA=   93.4   2,803.2
DF23   30.0   N=3   GA coal=   78.6   2,357.2   GA=   #N/A   #N/A
DF23All   30.0   N=124   GA coal=   59.2   1,775.9   GA=   62.3   1,869.0
DF23/M222   30.0   N=121   GA coal=   56.7   1,700.9   GA=   59.6   1,787.9
DF41All   30.0   N=4   GA coal=   47.0   1,410.7   GA=   64.8   1,942.6
R-L21/Z255/L159   30.0   N=28   GA coal=   41.5   1,243.9   GA=   46.5   1,396.4
L226   30.0   N=18   GA coal=   39.6   1,188.1   GA=   44.5   1,336.2
DF21/P314   30.0   N=3   GA coal=   39.3   1,178.6   GA=   #N/A   #N/A
DF21/L720   30.0   N=3   GA coal=   35.6   1,068.9   GA=   #N/A   #N/A
DF21/Z246/DF25/DF5/L627   30.0   N=4   GA coal=   33.9   1,017.6   GA=   53.4   1,603.4
L513/L193   30.0   N=28   GA coal=   33.5   1,003.5   GA=   37.1   1,114.5
L371   30.0   N=7   GA coal=   25.8   775.3   GA=   30.5   916.2
L513/L706.2/L705.2   30.0   N=3   GA coal=   22.8   685.2   GA=   #N/A   #N/A
DF41 x176148-Duffy   30.0   N=3   GA coal=   21.0   630.4   GA=   31.5   945.6
L513/L69.5   30.0   N=9   GA coal=   20.0   599.9   GA=   23.3   698.9
R-L21/L744/L745   30.0   N=10   GA coal=   17.3   518.0   GA=   19.7   592.0
L21-1130-A-1-A    30.0    N=4    GA coal=   12.3   370.0    GA=   16.4   493.4
L555   30.0   N=3   GA coal=   8.2   246.7   GA=   12.3   370.0
L513/L69/5/P66   30.0   N=3   GA coal=   5.5   164.5   GA=   8.2   246.7

[ironroad41]

Good work Mark.  Is L253 = Z253?  Is it at all possible to include the Clan Gregor Ian Cam as an entry to see what you get.  I would use 30 y/gen.  There are some 70 entries at Clan Gregor, FtDNA.  thanks bob

[Mark Jost]

Yes thats a type s/b Z253

Can you send me the Csv or excel file with them I will run it.

[ironroad41]

You can go to: www.familytreedna.com/public/MacGregor.  Select the Y STR data and look for the entries under the red bar, Ian Cam, Its in excel format.  good luck!

[Mark Jost]

Ok, here it is

YrsPerGen*   Count   AGE   Generations   YBP   Founder   Generations   YBP
30   N=8   GA coal=   16.1   481.8   GA=   19.5   585.9


2124   R1b1a2   Gregor founder of the clan
2909   R1b1a2   Peter McGregor c 1860
108707   R1b1a2a1a1b4   Robert McGregor m1815, Irvine, Scotland
120820   R1b1a2a1a1b4   daniel mcfarland, 1655-1738
131269   James Henry McGregor b. @ 1769 d. 1826   
133637   R1b1a2   John Dubh of Drumnacharrie and Stronfearnan
190435   R1b1a2a1a1b4   David Stewart, Greenbrier & Nicholas Cos VA 1780-1
191228   R1b1a2   Alexander McGregor bo 1700s 'Lanark'

MJost

[ironroad41]

Well, you did it Mark. Right on.  If I use 1960 as the average birthdate, I get 1375 which is as good as we can do.   I see you picked carefully and only used eight, but thats what it takes to do this well.  Thanks so much.  I'll forward your msg to the clan gregor moderator.

[ironroad41]

Well its time for me to "eat a little crow".  My congrats to MJost and GEn7 and Mikewww..  I still hold reservations about generalized sets of data.  But this sample convinces me it has a lot of merit.

[Jdean]

Why not have a play with it ?

It's easy enough to use and you can delete entire loci from the calculation to see how it effects the outcome.


BTW does anybody know where the mutation rates Ken used in Generations111T are published ?

[Stephen Parrish]

Why not have a play with it ?

It's easy enough to use and you can delete entire loci from the calculation to see how it effects the outcome.


BTW does anybody know where the mutation rates Ken used in Generations111T are published ?

The mutation rates that Ken used in Generations111T are from Marko Heinila; I have so far not been able to find his mutation rates online.

Stephen

[Mikewww]

Well its time for me to "eat a little crow".  My congrats to MJost and GEn7 and Mikewww..  I still hold reservations about generalized sets of data.  But this sample convinces me it has a lot of merit.

I'm just glad something aligned the way it should.... sometimes this stuff does make sense. No one should think this is mature science. I really do think Ken N knows his stuff, though.

[Jdean]

Why not have a play with it ?

It's easy enough to use and you can delete entire loci from the calculation to see how it effects the outcome.


BTW does anybody know where the mutation rates Ken used in Generations111T are published ?

The mutation rates that Ken used in Generations111T are from Marko Heinila; I have so far not been able to find his mutation rates online.

Stephen

Yes I read that elsewhere as well.

There are two sets of mutation rates in the spreadsheet, only one of which he uses (obviously). I'm intrigued what the other ones were and would also like to know the rates for 389I & II.

Probably I should just drop him a line.

[ironroad41]

Why not have a play with it ?

It's easy enough to use and you can delete entire loci from the calculation to see how it effects the outcome.


BTW does anybody know where the mutation rates Ken used in Generations111T are published ?

The mutation rates that Ken used in Generations111T are from Marko Heinila; I have so far not been able to find his mutation rates online.

Stephen

 I got a note from Marko.  He is still recovering, but he said that his new rates are about 15% faster than Chandlers.  That explains most of  the difference between the 800 years BP for the Ian Cam he had calculated previously and MJosts data? I think he will be on line soon(hope).  

[Mark Jost]

After considering the comment about dual set of mutations rates in KenN Gen111T spreadsheet,  Row 10 is labled as 'Backup' and it used in the calculations and the Sum of rates was shown in row10 was then shown in DM14. Row14 was the 2nd set of rates. The result at DM14 is: 0.243231% which is about the 15% faster, as just mentioned, than the 0.2% standard germ rate. FYI, the sum of the 2nd set is: 0.237478.