RIVPACS-type models infer impairment from the level of disagreement between observed and expected assemblages. The popular O/E index measures only the disagreement between the observed and expected number of taxa. Thus, O/E ignores additional information offered by a taxon-by-taxon contrast between assemblages. I compare O/E to a new index, the mean absolute difference (MAD) and also to the log-likelihood (LL) index. MAD and LL both summarize, across taxa, the disagreement between each taxon's observed occurrence and its model-predicted occurrence probability. However, neither MAD nor LL appear superior to O/E in practical terms. First, the site-specific rescalings required for MAD and LL obscure interpretations of their numeric scores. In contrast, O/E can be directly interpreted in terms of taxa loss. Second, in a macroinvertebrate predictive model applied to 499 Mid-Atlantic Highlands stream sites, O/E  differentiated between 5 independently-assigned site disturbance classes, and also between reference and nonreference condition, as strongly as did LL or MAD. Correlations between O/E and either MAD or LL, across Mid-Atlantic sites, were strong (>0.85) when indices included only locally-common taxa (Pc > 0.5), but were weak (<0.25) when all taxa were included, probably because MAD and LL have greater sensitivity to occasional rare-taxon occurrences.