I am have thinking about error estimation for Admixture results for some time since I have heard a lot of arguments about how even 0.1% result is significant. I was skeptical of that and have rounded off my admixture run results to the nearest percent.
There was a memory leak issue in the bootstrapping code for admixture which crashed it every time I tried running it. I emailed David Alexander and he fixed it in version 1.12.
So I ran the default 200 bootstrap replicates to measure standard error in our old Reference I K=12 admixture. Spreadsheet with population level results is here and participant results are here.
Here are some statistics for the standard error estimates:
| Min. | 1st Qu. | Median | Mean | 3rd Qu. | Max. | |
|---|---|---|---|---|---|---|
| C1 S Asian | 0.00% | 0.02% | 0.33% | 0.52% | 0.96% | 1.93% |
| C2 Blch/Cauc | 0.00% | 0.00% | 1.02% | 0.79% | 1.45% | 2.63% |
| C3 Kalash | 0.00% | 0.01% | 0.40% | 0.50% | 0.99% | 3.76% |
| C4 SE Asian | 0.00% | 0.09% | 0.37% | 0.60% | 1.27% | 1.92% |
| C5 SW Asian | 0.00% | 0.00% | 0.60% | 0.66% | 1.28% | 2.90% |
| C6 Euro | 0.00% | 0.00% | 0.35% | 0.56% | 1.12% | 1.82% |
| C7 Papuan | 0.00% | 0.07% | 0.22% | 0.23% | 0.36% | 1.08% |
| C8 NE Asian | 0.00% | 0.07% | 0.36% | 0.67% | 1.36% | 2.45% |
| C9 Siberian | 0.00% | 0.08% | 0.37% | 0.51% | 0.82% | 2.29% |
| C10 E Bantu | 0.00% | 0.00% | 0.00% | 0.35% | 0.72% | 1.93% |
| C11W Afr | 0.00% | 0.00% | 0.00% | 0.28% | 0.50% | 1.51% |
| C12 E Afr | 0.00% | 0.00% | 0.05% | 0.31% | 0.60% | 1.79% |
You can see the mean value of the standard errors per population and realize how many are over 1% (marked in red).
And statistics for bias estimates:
| Min. | 1st Qu. | Median | Mean | 3rd Qu. | Max. | |
|---|---|---|---|---|---|---|
| C1 S Asian | -1.104% | -0.031% | 0.000% | -0.024% | 0.075% | 1.026% |
| C2 Blch/Cauc | -0.835% | -0.280% | -0.009% | -0.133% | 0.000% | 1.049% |
| C3 Kalash | -1.575% | 0.000% | 0.020% | 0.076% | 0.147% | 0.615% |
| C4 SE Asian | -0.629% | -0.021% | 0.011% | 0.018% | 0.087% | 0.478% |
| C5 SW Asian | -0.691% | -0.094% | 0.000% | -0.020% | 0.035% | 0.613% |
| C6 Euro | -0.572% | -0.086% | 0.000% | -0.039% | 0.004% | 0.468% |
| C7 Papuan | -0.171% | 0.008% | 0.059% | 0.070% | 0.120% | 0.312% |
| C8 NE Asian | -0.739% | 0.000% | 0.016% | 0.034% | 0.107% | 0.679% |
| C9 Siberian | -1.044% | 0.000% | 0.015% | 0.035% | 0.103% | 0.692% |
| C10 E Bantu | -0.412% | 0.000% | 0.000% | -0.007% | 0.001% | 0.370% |
| C11 W Afr | -0.261% | 0.000% | 0.000% | 0.009% | 0.005% | 0.304% |
| C12 E Afr | -0.635% | 0.000% | 0.000% | -0.017% | 0.010% | 0.405% |
You can also see the average value of the bias in each ancestral component for each population.
Since the bias is lower than the standard error and distributed around zero, if a large number of samples of a population group have some small percentage of an ancestral component, the likelihood of that not being noise is higher.
Comments are closed.