search

DReichLab / AdmixTools

Tools test whether admixture occurred and more
183 stars 64 forks source link

qpAdm sometimes copies admixture proportions in other lines #24

Closed TCLamnidis closed 7 years ago

TCLamnidis commented 7 years ago

Hello!

Sometimes qpAdm will give me identical admixture proportions between different models, and even forego forcing some to 0. (see part of results below). I am fairly sure the chi square statistics are accurate for the model, I just don't get the admixture proportions.

fixed pat  wt  dof     chisq       tail prob
    00000  0     4     2.925        0.570392     0.006    -0.022     0.557     0.048     0.411  infeasible
    00001  1     5   101.589     2.44391e-20    -4.084     0.558     0.092     4.434     0.000  infeasible
    00010  1     5     3.026        0.696022    -4.084     0.558     0.092     4.434     0.000  infeasible
    00100  1     5    83.532     1.52896e-16    -4.084     0.558     0.092     4.434     0.000  infeasible

Clearly in the latter 2 models, the final component shouldn't be forced to 0, and instead the other components should.

Is this a problem with my use, or an internal problem? Cheers

TCLamnidis commented 7 years ago

I should add, as far as I can see, this happens consistently with certain sample sets, copying over the same models every time (at least in my 3 replicates).

bumblenick commented 7 years ago

This does look like a bug. Does it ever happen when the model is "feasible"

A workaround of course is to rerun deleting sources you don't want, and rerun. Please tell me the version number (grep on version: in logfile)

N

On Thu, Apr 27, 2017 at 11:17 AM, TCLamnidis notifications@github.com wrote:

Hello!

Sometimes qpAdm will give me identical admixture proportions between different models, and even forego forcing some to 0. (see part of results below). I am fairly sure the chi square statistics are accurate for the model, I just don't get the admixture proportions.

fixed pat wt dof chisq tail prob 00000 0 4 2.925 0.570392 0.006 -0.022 0.557 0.048 0.411 infeasible 00001 1 5 101.589 2.44391e-20 -4.084 0.558 0.092 4.434 0.000 infeasible 00010 1 5 3.026 0.696022 -4.084 0.558 0.092 4.434 0.000 infeasible 00100 1 5 83.532 1.52896e-16 -4.084 0.558 0.092 4.434 0.000 infeasible

Clearly in the latter 2 models, the final component shouldn't be forced to 0, and instead the other components should.

Is this a problem with my use, or an internal problem? Cheers

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/DReichLab/AdmixTools/issues/24, or mute the thread https://github.com/notifications/unsubscribe-auth/AQn_h-uM1qn8bZmmsEOE5kFXHtHfXTPqks5r0LF3gaJpZM4NKXcQ .

TCLamnidis commented 7 years ago

## qpAdm version: 610

Yes, it also happens with feasible models:

        00000  0     2     4.803       0.0905804     0.592     0.097     0.183     0.101     0.028 
        00001  1     3    17.570     0.000539394     0.493     0.098     0.210     0.199     0.000 
        00010  1     3     7.677       0.0531877     0.493     0.098     0.210     0.199     0.000

And it seems to overwrite models that are not rejected too, since in the first example one of the models had a high tail probability.

bumblenick commented 7 years ago

I think I found the problem and am in the process of fixing. It likely is caused by having 2 populations in the left set that are very very similar genetically. Is that plausible here?

Nick

On Thu, Apr 27, 2017 at 12:36 PM, TCLamnidis notifications@github.com wrote:

qpAdm version: 610

Yes, it also happens with feasible models:

    00000  0     2     4.803       0.0905804     0.592     0.097     0.183     0.101     0.028
    00001  1     3    17.570     0.000539394     0.493     0.098     0.210     0.199     0.000
    00010  1     3     7.677       0.0531877     0.493     0.098     0.210     0.199     0.000

And it seems to overwrite models that are not rejected too, since in the first example one of the models had a high tail probability.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/DReichLab/AdmixTools/issues/24#issuecomment-297770168, or mute the thread https://github.com/notifications/unsubscribe-auth/AQn_h9etoUH6tIlzT-Q5BmJObYHPAHJbks5r0MQYgaJpZM4NKXcQ .

TCLamnidis commented 7 years ago

Here, the five source left populations are genetically distinct. Moreover, they remain mostly intact across multiple analyses, the changing variable being the population being explained as a 5-way mixture. Only specific target populations seem to trigger this effect, while most run normally without copied lines in the outputs.

Note: I have managed to run the reduced models without problem, as you suggested.

bumblenick commented 7 years ago

I just pushed a new version of Admixtools to github. Probably this fixes your problem. Please let me know!

Nick

On Thu, Apr 27, 2017 at 11:17 AM, TCLamnidis notifications@github.com wrote:

Hello!

Sometimes qpAdm will give me identical admixture proportions between different models, and even forego forcing some to 0. (see part of results below). I am fairly sure the chi square statistics are accurate for the model, I just don't get the admixture proportions.

fixed pat wt dof chisq tail prob 00000 0 4 2.925 0.570392 0.006 -0.022 0.557 0.048 0.411 infeasible 00001 1 5 101.589 2.44391e-20 -4.084 0.558 0.092 4.434 0.000 infeasible 00010 1 5 3.026 0.696022 -4.084 0.558 0.092 4.434 0.000 infeasible 00100 1 5 83.532 1.52896e-16 -4.084 0.558 0.092 4.434 0.000 infeasible

Clearly in the latter 2 models, the final component shouldn't be forced to 0, and instead the other components should.

Is this a problem with my use, or an internal problem? Cheers

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/DReichLab/AdmixTools/issues/24, or mute the thread https://github.com/notifications/unsubscribe-auth/AQn_h-uM1qn8bZmmsEOE5kFXHtHfXTPqks5r0LF3gaJpZM4NKXcQ .

TCLamnidis commented 7 years ago

I tried rerunning some of the models that had problems before, and all mixture proportions are printed normally now.

bumblenick commented 7 years ago

Great; I think you have a boundary case in which some populations have very thin data. In the new release I trap these cases and do something (fairly) reasonable.

Nick

On Tue, May 16, 2017 at 5:20 AM, TCLamnidis notifications@github.com wrote:

I tried rerunning some of the models that had problems before, and all mixture proportions are printed normally now.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/DReichLab/AdmixTools/issues/24#issuecomment-301724727, or mute the thread https://github.com/notifications/unsubscribe-auth/AQn_hxlnrt9DeEXAGRh43h8E8QU-XJCWks5r6WpAgaJpZM4NKXcQ .