# When comparing classificators, is McNemar's test applied to the output or to the hits?

Suppose I have the following output, alongside the ground-truth ($$GT$$), for two classifiers $$C_1$$ and $$C_2$$:

GT    C1    (C1 = GT)    C2     (C2 = GT)
a     a         1         b         0
b     a         0         b         1
b     b         1         b         1
a     a         1         a         1
a     b         0         a         1
b     b         1         a         0


Is the confusion table used for the McNemar's test the one built from columns $$C_1$$ and $$C_2$$, or from comparing columns $$C_1 = GT$$ and $$C_2 = GT$$?

My intuition says it is the later, but the material I've found about the subject was a little confusing.

• McNemar's test is a test of the odds ratio in paired binary data. You shouldn't use that to compare classifiers. High OR != Good classification accuracy and vice versa. – AdamO Jun 21 '19 at 16:20