Word frequency and receiver operating characteristic curves in recognition memory: Evidence for a dual-process interpretation.

Abstract

Dual-process models of the word-frequency mirror effect posit that low-frequency words are recollected more often than high-frequency words, producing the hit rate differences in the wordfrequency effect, whereas high-frequency words are more familiar, producing the false-alarm-rate differences. In this pair of experiments, the authors demonstrate that the analysis of receiver operating characteristic (ROC) curves provides critical information in support of this interpretation. Specifically, when participants were required to discriminate between studied nouns and their plurality reversed complements, the ROC curve was accurately described by a threshold model that is consistent with recollection-based recognition. Further, the plurality discrimination ROC curves showed characteristics consistent with the interpretation that participants recollected low-frequency items more than high-frequency items.One of the most replicable empirical results in the recognition memory literature is the wordfrequency effect. The word-frequency effect is the finding that low-frequency words show superior recognition relative to high-frequency words, both in terms of a higher hit rate and a lower false-alarm rate (Glanzer & Bowles, 1976;Gorman, 1961). Such a pattern of results (i.e., that higher hit-rates are often accompanied by lower false-alarm rates) has been dubbed the mirror effect by Glanzer and colleagues (Glanzer & Adams, 1985;Glanzer, Adams, Iverson, & Kim, 1993). Although the mirror effect is more general than manipulations of word frequency (e.g., encoding manipulations such as increased study time also produce mirror effects; Hirshman, 1995;Ratcliff, Clark, & Shiffrin, 1990;Stretch & Wixted, 1998), accounting for the word-frequency mirror effect has proven especially difficult for many theories of recognition memory. In particular, the word-frequency mirror effect has been difficult for global matching models (Gillund & Shiffrin, 1984;Hintzman, 1988;Murdock, 1982) to explain. The reason for this is simple: Models that assume that a single strength dimension underlies recognition memory must explain why low-frequency words show lower levels of memory strength than high-frequency words when they are unstudied but higher levels of memory strength when they are studied.In light of the difficulty that strength-based single-process models have faced in explaining the mirror effect in general, and the word frequency effect in particular, new theories have been proposed that account for mirror effects, including the word-frequency effect. Some of these theories maintain the assumption that a single process underlies recognition memory

The mirror effect and mixture signal detection theory.

“…One problem, however, is that the deviations do not appear to be random but tend to show a systematic pattern, in that the plots suggest nonlinearity. This is most evident for Arndt and Reder's (2002) data, shown in the bottom panels. For example, the black and gray circles in the left bottom panel show the LO and HO data, and it is apparent that the leftmost and rightmost circles are above the fitted line, whereas those in between are below the fitted line.…”

“…This is an important result that is consistent with the interpretation that the LN* distribution represents lowfrequency new words that are treated like LO words; a failure to find this result would cast doubt on the "false recognition" interpretation offered here. The mixing proportions in Table 4 show that the proportion of LN* words was about 30% in Experiment 1a, 34% in Experiment 1b, and 4% in Arndt and Reder's (2002) Experiment 2, and so there was less mixing for low-frequency new words in Arndt and Reder's study; this might be due to differences in the lowfrequency words used across the studies. A reviewer asked whether the LN* distribution can simply be restricted to have the same location as the LO distribution, as the LN* words are viewed as being low-frequency new words that are treated like LO words.…”

“…Second, a reversal in the order of the LO and HO distributions was not found in any of the studies. More specifically, the results for Experiments 1a and 1b suggest that the LO and HO distributions have the same location, and those for Arndt and Reder's (2002) Experiment 2 suggest that LO and HO have either the same location or HO is to the right of LO. In any case, a reversal in strength did not appear in any of the experiments once effects of mixing were controlled for.…”

“…A reviewer asked whether the LN* distribution can simply be restricted to have the same location as the LO distribution, as the LN* words are viewed as being low-frequency new words that are treated like LO words. The parameter estimates and standard errors in Table 4 show that this does not hold for Experiments 1a and 1b (but possibly for Arndt and Reder's 2002 Experiment 2), whereas Table 7 (later) shows that this also does not hold for Arndt and Reder's (2002) Experiment 1. Thus, the results suggest that there are differences between low-frequency new words that are considered as old (LN* words) and low-frequency words that are actually old (LO words), as reflected by the different levels of familiarity-that is, the LN* words do not appear to be as familiar as actual LO words.…”

“…Low-frequency words had a frequency count of less than 4 per million; highfrequency words had a count of greater than 24 per million. Further details can be found in Arndt and Reder (2002). Table 1 presents the frequencies for Experiments 1a and 1b pooled over subjects.…”