Table 3.

Average out-of-sample test errors (rounded average model sparsity in parenthesis) for trac ($a\in \{1,1/2\}$) and sparse log-contrast models, respectively. Each row represents the results for base level OTU, genus, and family and the corresponding dimensionality of the base level. Each value is averaged over ten different training/test splits of the Tara data.

Base level	p	trac (a $=$ 1)	trac (a $=$ 1/2)	Sparse log-contrast
OTU	8916	2.1 (7)	1.8 (14)	1.3 (24)
Genus	4220	2.0 (7)	1.5 (14)	1.4 (34)
Family	1869	2.1 (6)	1.7 (10)	1.6 (13)