Predicting U5 haplogroup

I have made some experiments with predicting major haplogroup branch (U5a1, U5a2, U5b1, U5b2, U5b3) using machine learning techniques. I have tested multinominal logistic regression (“multilogit”), random forest, discriminant analysis and K Nearest Neighbor. So far I have had best luck in classification accuracy with multinominal logistic regression so I will present those results. Only the HVR1 region is used as information and the models are estimated using those kits that have a definite haplogroup. Remember that this is only estimats so it is not guarranteed that it is correct!

The classification results can be seen below. All kits in U5a1, U5a2 and U5b3 are correctly classified. U5b1 and U5b2 seems to be harder to separate. In total 97.2% of the kits are correctly classified. Number of predicted in haplogroup (in colums) against observed in haplogroups (in rows). Predicted haplogroup with multinominal logistic regression only using HVR1.

Below are the results for every kit. The results from the multilogitis in column three. In addition haplogroups for branch defining mutations that are found in HVR1 or HVR2 are shown in the following two columns. Don’t worry if the information here is contradictory, mutations can have happened several times.