retrieval cue does not isolate a specific memory trace, but rather activates many traces simultaneously, each to a degree depending on its similarity to the cue. The retrieved information is a scalar value (variously referred to as strength, familiarity, similarity, or echo intensity) that represents the sum of memory activation produced by the cue. Old-new recognition judgments are made relative to a criterion on the summed activation scale, and forced-choice recognition judgments are made by choosing the test alternative whose summed activation is the highest. Clark and Gronlund (1996) discuss some of the successes and failures of these models.One success of global-matching models has been in explaining the test-pair similarity effect-a phenomenon of forced-choice recognition. Consider two unrelated items, A and B, both of which are old with respect to the study list, and two new items, A¢ ( similar to A) and B¢ (similar to B). Two-alternative forced-choice performance is better on similar pairs such as A versus A¢ than it is on dissimilar pairs such as A versus B¢. In the first case, the lure is related to the old test item with which it is paired, and in the second case, it is related to an old item that is not part of the same test pair. The test-pair similarity effect has been found in experiments manipulating the orthographic similarity of words (Hall, 1979), the semantic similarity of words (Hintzman, 1988), and the visual similarity of photographs (Tulving, 1981). An exception is that the opposite effect-a similar-pair disadvantage-occurs in associative recognition (Clark, Hori, & Callan, 1993).The test-pair similarity prediction derives from a fundamental property of global matching models, which is that correlated inputs or cues produce correlated outputs. Because cues A and A¢ are similar, they tend to activate the same subset of memory traces, prominently including the trace of A itself. Thus, if A produces summed activation that is relatively high for an old test item, A¢ will tend to cause activation that is relatively high for a new item. If A produces activation that is low for an old item, A¢ is likely to produce activation that is low for a new item. No such correlation is predicted for the unrelated test pair, A versus B¢.The choice between an old test item and a new test item is made by selecting the item with the higher summed activation; thus the probability of an error is P[(X 2 Y ) < 0], where X is activation caused by the old item, and Y is activation caused by the new item. In the simplest model, the old-and new-item means ( μ X and μ Y ) and variances (s 2 X and s 2 Y ) are assumed to be unaffected by the type of test pairing. Thus, P[(X 2 Y ) < 0] increases as a direct function of the variance of the differences, X 2 Y:where r XY is the Pearson correlation relating X and Y. When the old item is A and the new item is B¢, r is zero and the error rate is at its maximum. When the old item is A and the new item is A¢, r is positive and the error rate is reduced. To see intuitively why thi...