Memory for Items and Associations: Distinct Representations and Processes in Associative Recognition

Abstract

In two experiments, participants studied word pairs and later discriminated old (intact) word pairs from foils, including recombined word pairs and pairs including one or two previously unstudied words. Rather than making old/new memory judgments, they chose one of five responses: (1) Old-Old (original), (2) Old-Old (rearranged), (3) Old-New, (4) New-Old, (5) New-New. To tease apart the effects of item familiarity from those of associative strength, we varied both how many times a specific word-pair was repeated (1 or 5) and how many different word pairs were associated with a given word (1 or 5). Participants could discriminate associative information from item information such that they recognized which word of a foil was new, or whether both were new, as well as discriminating recombined studied words from original pairings. The error and latency data support the view that item and associative information are stored as distinct memory representations and make separate contributions at retrieval.

Memory theorists distinguish between two qualitatively different sources of information underlying the process of recognition (Hockley, 1992; Humphreys, Bain, & Pike, 1989; Kelley & Wixted, 2001; Murdock, 1997; Reder, Nhouyvanisvong, Schunn, Ayers, Angstadt, & Hiraki, 2000; Yonelinas, 1997). Memory for associative information, such as the co-occurrence of two words, can be distinguished from memory for the individual items (for reviews, see Clark & Gronlund, 1996; Raaijmakers & Shiffrin, 1992; Yonelinas, 2002). However, memory theorists differ with respect to just how item information and associative information are represented in memory and how each contributes to decision-making in the recognition of word pairs. One controversy is whether associative information is stored separately from item information or is stored in common with item information (Gillund & Shiffrin, 1984; Hintzman, 1988; Hockley, 1991, 1992; Murdock, 1993). A second core issue is how item and associative information are accessed and utilized in subsequent decision-making during recognition. The research reported in this article addresses these two issues.

The Representation of Item and Associative Information in Associative Recognition

Global models of associative recognition, such as REM (Shiffrin & Steyvers, 1997), TODAM (Murdock, 1997), MINERVA II (Hintzman, 1988), and Matrix (Humphreys, Bain, & Pike, 1989), assume that item and associative information are inseparable and are stored as part of a common memory system, whereas local models such as SAC (Reder et al., 2000) assume that associative information is represented separately from item information. In associative recognition, the REM, TODAM and Matrix models use similar mathematical operations to represent associative information as a conjoined representation of two sets of item features – either a concatenation (REM), convolution (TODAM), or tensor product (Matrix) of two item vectors. Local models, in contrast, store item and associative information in separate memory locations. For instance, the SAC model (Reder et al., 2000) representation of a studied word pair is explicitly defined by two semantic nodes, one for each word in the pair, with each word node linked to the same episodic node encoding their association – the co-occurrence of both words in a particular study context (see models 4 and 5, Buchler & Reder, 2007, pp. 111-112; Reder et al., 2007).

In the research reported in this article, we test these assumptions underlying local and global models. If associative information is a product of item information, as assumed by global models, it follows that the strengthening of item information necessitates the strengthening of associative information. However, if item and associative information are stored separately, as assumed by local models, it should be possible through appropriate manipulation of item and pair repetitions to strengthen item information without strengthening associative information.

In our study of word pair recognition, we use two experimental manipulations – repetition (differential strength of the word pairing) and fan or multiple word pair associates (differential associative interference) – to further differentiate item and associative sources of information. In two experiments, we manipulated how many times a specific word-pair was repeated and how many different word pairs were associated with a given word to tease apart the effects of item strength from associative strength. If associative information is stored as a conjoined memory representation of item information as assumed by global models, then the memory representation for pairs with multiple overlapping associates should be equivalent in strength to word pairs in which the same word-pair is repeatedly presented. Examples are given in Table 1 (see Fan 5-5 and Rep×5). In both cases the underlying item strength is identical, and by extension, global matching models predict that the associative strength should be identical as well. In contrast, local models of memory predict that it should be possible to additionally strengthen associative information separately from item information. In this case, repeated word pairs should be better recognized than overlapping word pairs. Although these would seem to be straightforward predictions, we were unable to find any evidence of a test between these competing views. We explicitly confirmed these predictions with a Monte Carlo simulation of both a global matching model of memory, REM, and a local memory model, SAC. A description of the REM and SAC model mechanics and the results of our Monte Carlo simulations are provided in the General Discussion.

Table 1. An illustration of word-pair stimulus composition for the study and test phases in relation to the level of fan during study and the stimulus or foil type at test.

Study: Condition:	Number of Stimuli:	Study Word-Pair	Test Condition:	Level of Fan:	# of Stimuli:	Tested Word-Pair
Rep ×5	50	KEY-BEACH	Intact	1-1	10	KEY-BEACH
		KEY-BEACH
		KEY-BEACH
		KEY-BEACH
		KEY-BEACH
Fan 1-1	30	CHEMICAL-SMILE	Intact	1-1	10	CHEMICAL-SMILE
		ENERGY-HOUSING	Rearranged	1-1	10	ENERGY-BEAUTY
		ARTIST-BEAUTY
		MOON-FRAME	Item	1-new	10	MOON-SHOT
			Item	new-1	10	FACULTY-FRAME
Fan 1-5	20	SHELTER-PHASE	Intact	1-5	10	SHELTER-PHASE
		WOOD-PHASE
		CREDIT-PHASE
		CIRCLE-PHASE
		BOTTLE-PHASE
		JURY-COMMERCE	Rearranged	1-5	10	JURY-MESSAGE
		CUTTING-COMMERCE
		FOREST-COMMERCE
		TRADITION-COMMERCE
		RELATIONSHIP-COMMERCE
		MINE-MESSAGE
		PERMIT-MESSAGE
		REALITY-MESSAGE
		HAT-MESSAGE
		REFERENCE-MESSAGE
Fan 5-1	20	(see Fan 1-5)	Intact	5-1	10	(see Fan 1-5)
			Rearranged	5-1	10
Fan 5-5	30	SOLUTION-SKY	Intact	5-5	10	SOLUTION-SKY
		TEACHER-SKY
		VILLAGE-SKY
		SHOULDER-SKY
		ENEMY-SKY
		SOLUTION-MODEL
		TEACHER-MODEL
		VILLAGE-MODEL
		SHOULDER-MODEL
		ENEMY-MODEL
		RESOLUTION-BACKGROUND	Rearranged	5-5	10	SOLUTION-BACKGROUND
		TITLE-BACKGROUND
		DEVICE-BACKGROUND
		REGION-BACKGROUND
		CELL-BACKGROUND
			Item	5-new	10	SOLUTION-PARK
			Item	new-5	10	GAS-BACKGROUND
			Novel	new-new	10	TRIP-CHAIR
Total pairs studied : 150			Total pairs tested : 160

The Use of Item and Associative Information in Recognition Decision-Making

A second, related issue is whether item and associative sources of information make discrete contributions at retrieval or conversely, whether item and associative information contribute to an aggregate measure of memory strength. This question addresses a critical distinction between two classes of memory decision models – global matching and dual-process. Global matching models (Hintzman, 1988; Humphreys, Bain, & Pike, 1989; Murdock, 1997) assume that the retrieval process aggregates item and associative information into a single composite measure of memory strength. Whether a stimulus is called “old” or “new” depends on whether the composite memory score falls either above or below a decision criterion. For instance, Hockley (1991, 1992) conceptualized word pair recognition as a signal detection process based on a single memory strength statistic, the sum of item information and associative information that must exceed a decision threshold.

In contrast, the dual-process model of Yonelinas (1997, 2002) argues that recognition is not based on memory strength aggregated across sources, but is instead based on the outputs of two separable memory processes, familiarity and recollection (Mandler, 1980; Reder et al., 2000). The process of recollection involves retrieving specific contextual associations (i.e. episodic traces) and the familiarity process is based on the general strength of items in memory. Recollection is treated as a threshold process that can be distinguished by highly confident recognition and vivid mnemonic detail (for reviews, see Diana, Reder, Ardnt & Park, 2006; Yonelinas, 2002). Whether a stimulus is recognized as “old” or “new” depends on whether either memory process is successful.

Among dual-process models, however, there is no universal agreement on how item and associative sources contribute to an associative recognition judgment. That is, dual-process models all assume an attempt to recollect the association, but offer a number of different characterizations of the familiarity process. Conceptually, a word pair recognition task has three potentially discrete sources of information – each word in the pair (word₁, word₂) and the association. All three sources (word₁, word₂, association) contribute to familiarity in the Kelley and Wixted (2001, Wixted, 2007) model, both item sources contribute to familiarity in the Yonelinas (1997, Yonelinas & Parks, 2007) model, and each item can contribute separately in the SAC model (Buchler & Reder, 2007; Reder et al., 2000).

Theoretical treatments of associative recognition generally lump together the two words in the pair as contributing to ‘item’ information. However, there is evidence that participants are able to accurately make frequency judgments for words and word pairs and thus distinguish the frequency of occurrence of specific item information (each word) and associative information (pairing). Hockley and Cristi (1996) orthogonally varied the study presentation frequency of word pairs and the frequency of the individual items of the pairs across different study pairings. In separate tests, participants were asked to judge the frequency of the word pairs, a word's frequency as an individual item, its frequency as a member of word pairs, and the combined frequency of the word. Each of the four judgments was made with considerable accuracy, indicating that participants can distinguish between item and associative information when the task demands such discrimination.

If participants are able to specifically distinguish the item strength of tested word pairs in the context of an associative recognition judgment, this suggests that multiple sources of information can be queried discretely in memory, including the left and right word. Given the opportunity, participants may be able to distinguish among all sources of information in memory (word₁, word₂, association). This possibility would support dual-process models such as SAC (Buchler & Reder, 2007; Reder et al., 2000) which assume a local representation and a decision process that differentiates not only the associative information, but also both item sources.

Our approach is two-fold as we increase both the specificity of the recognition decision and the number of stimulus-types that the subject has to differentiate. Before describing our experimental paradigm, we introduce our second manipulation, how using multiple foil-types can dissociate the contribution of item and associative information.

Using Multiple Foil-types in Word Pair Recognition

Word pair recognition tasks have been used to distinguish the contributions of item and associative sources of information in a recognition judgment. In a typical associative recognition task, participants are required to distinguish previously studied “old” word-pairs (KEY-BEACH, CHEMICAL-SMILE) from “new” foil pairs composed of recombined study pairs not paired together during study (KEY-SMILE). Correct identification of previously studied word pairs requires the retrieval of associative information – that two words were presented together during study. Only associative information can be used to distinguish between targets and foils because the foil pairs all involve previously studied words that are combined into new pairings. That is, item information is not diagnostic since all words are equally familiar.

In order to also examine the contribution of item information, Castel and Craik (2003, see also Humphreys, 1976, 1978) presented three different types of word pair foils in a paired-associate recognition test. Word pairs studied in an initial learning phase are denoted by A-B and C-D, where A, B, C, and D are studied words. In the test phase, participants respond “old” or “new” to four types of word pairs: (a) intact pairs (A-B, C-D), (b) rearranged pairs (A-D, C-B, with word position in the pair preserved), (c) item pairs (A-X, X-D), and (d) novel pairs (Y-Z), where X, Y, and Z are new words. The pattern of hits and false alarms helped separate the contributions of item and associative information to word pair recognition. The discrimination of intact from rearranged word pairs requires the retrieval of associative information, whereas item information can be used to discriminate between the various types of item, and novel word pair foils. Castel and Craik (2003) found that false alarm rates increased with the number of study words reinstated at test, from novel pairs (zero), to item pairs (one), to rearranged pairs (two). They attributed this pattern to differences in accumulated memory strength of the items in the foil word pairs.¹ As discussed below, our research examines the assumption proposed by a number of composite strength accounts (Donaldson, 1996; Dunn, 2004; Hirshman & Master, 1997) that multiple recognition decisions can be assigned by using more than one decision threshold (i.e. low versus high confidence).

Five-Choice Paired Associate Recognition (5-PAR) Task

We provide a fine grain-size for examining the basis of recognition decision-making. Our experimental procedure emphasizes the decision-making aspects of recognition, which we operationally define as the process of assigning test stimuli to response categories. Participants studied word-pairs and later discriminated original word-pairs from various types of foils including recombined word-pairs and foils composed of one or no previously studied words. Instead of old/new memory judgments, they chose one of five responses: (1) Old-Old (original), (2) Old-Old (rearranged), (3) Old-New, (4) New-Old, (5) New-New. This five-choice paired-associate recognition (5-PAR) experimental procedure allowed us to tease apart the contributions of item and associative information to a word pair recognition decision, and to examine whether participants are able to explicitly dissociate the specific item strengths of the left and right words in the pair, as well as the ability of participants to detect novel stimuli.

Global matching models account for multiple recognition responses, whether the additional specificity of a “remember / know” response (Gardiner & Java, 1991) or confidence level, by specifying response thresholds at gradations of memory strength. For instance, some researchers maintain that the increased specificity of “remember / know” responses indicates low and high levels of confidence in a unitary memory strength statistic and that these responses do not differentiate between the dual-processes of recollection and familiarity (Donaldson, 1996; Dunn, 2004; Hirshman & Master, 1997). It is conceivable that a similar procedure will be used by participants in assigning our 5-PAR responses to the word pair stimuli. If so, a strong prediction of models assuming composite memory strength is that participants will be unable to differentiate between item word pairs with the studied word (A) in the left (A-X) or right position (Y-A) since the overall memory strength is identical. On the other hand, participants may be able to recognize specific item information (word₁, word₂) and discriminate the separate source contributions of the right and left word in the pair by responding Old-New or New-Old in our 5-PAR task. The 5-PAR decisions were also assessed in response to two experimental manipulations – repetition and multiple word pair associates (fan) – used to orthogonally vary the item strength of words and the associative strength of studied word pairs.

The Role of Associative Information in Associative Recognition

A number of recent findings suggest that prior exposure to associative information can have negative consequences for recognition memory. Ability to retrieve the associative component of the memory trace can be manipulated by presenting pairs that share words with other pairs, such that each word in the pair is also presented in the same left or right position with other words from other pairs (e.g. SOLUTION-SKY, TEACHER-MODEL, SOLUTION-MODEL, TEACHER-SKY). In this example both the left and right words of a pair were studied in exactly two pairs, which we denote as Fan 2-2 after Anderson and Bower (1973) who demonstrated that both error rates and response latencies increase with fan (Anderson, 1974; Anderson & Lebiere, 1998; Anderson & Reder, 1999; Reder & Anderson, 1980; Reder & Ross, 1981; Sohn, Anderson, Reder & Goode, 2004), a result ascribed to associative interference. This effect has been modeled by assuming that activation is spread to associated memories and second, that this spreading activation is increasingly diffuse for higher levels of fan. Activation is distributed in proportion to the total number of associations so that with increased fan each association receives a smaller amount of activation, leading to increases in both response latencies and the number of intrusion errors.

Verde (2004) used interference to examine associative recognition and manipulated the fan of the word pairs up to Fan 4-4. Like others (Diana et al, 2006; Reder et al., 2000), Verde found results consistent with associative interference models, such as a diminished ability to discriminate intact from rearranged pairs. Using remember-know judgments to supplement an ‘old’ recognition judgment, Verde (2004) found that the proportion of remember responses decreased whereas know responses increased with fan. Remember responses have historically been attributed to associative recollection and know responses to item familiarity (see Diana et al., 2006; Hockley & Consoli, 1999). Using that logic, these results suggest that the availability of associative information decreased as a result of the associative interference generated by the fan manipulation.

The Time Course of Recognition Responses

Recognition models should be able to explain patterns of response times as well as patterns of accuracy. Timing considerations have generally favored dual-process interpretations because the retrieval of item and associative information has different temporal characteristics. For instance, differential forgetting rates have been found for item (rapid decay) and associative (relatively stable) information (Hockley, 1991). This finding challenges the notion that item and associative information are stored and retrieved as part of the same memory system representation (but see Murdock, 1997). Furthermore, a number of researchers have found that the accuracy of recognizing rearranged word pairs varies with response deadlines; as the time allotted for a recognition response is increased, the probability of incorrectly endorsing rearranged word pair foils initially increases and then decreases (Dosher, 1984; Gronlund & Ratcliff, 1989; Hintzman & Curran, 1994; McElree et al., 1999). That the false acceptance function is initially inflected upward indicates that associative information takes longer to recruit than item information (but see Dewhurst, Holmes, Brandt, & Dean, 2006). In defense of composite strength accounts, it is possible that timing differences reflect differing degrees of completeness of a memory trace, whose gradual accumulation has been curtailed by deadline procedures. That is, the composite strength account may be correct, but for some reason, associative traces take longer to accrue.

We examine decision latencies for the various 5-PAR responses under self-paced conditions when accuracy is stressed. Dual-process models assume that when accuracy is stressed, participants first attempt recollection, which can yield a highly-detailed memory trace and a corresponding high confidence response. If recollection fails, participants can respond based on the familiarity of the word pair (see Kelley & Wixted, 2003; Yonelinas, 2002). In our 5-PAR task, dual-process models predict dissociation in the timing of the recognition responses based on the type of information retrieved. On the other hand, we hypothesized that if the accumulation of the composite signal is gradual, then we would expect the timing of the various recognition responses to be continuously arrayed as a function of combined item and associative strength. The recognition response should speed up with the strength of the underlying memory trace, being fastest for intact studied word pairs, faster for rearranged word pairs, and slowest for item word pairs. Our analysis of decision latencies should be informative about how information is queried in memory, either en masse or as part of a serial retrieval process, or a bit of both.

Experiment 1

Method

Participants

Thirty-four participants (14 males), recruited from the Carnegie Mellon University undergraduate community, participated in exchange for $5.

Materials and Design

Table 1 illustrates the different experimental conditions in this within subject design. We varied the fan or number of different pairings associated with a word (either 1 or 5) and orthogonally varied the fan of the left and right words of a pair. During the study phase, this meant that participants were presented with Fan 1-1, Fan 1-5, Fan 5-1, and Fan 5-5 word pairs. Some 1-1 pairs were repeated 5 times (Rep×5) such that the exact same pair was studied five times. All other word pairs were studied only once. No pairs involving a Fan 5 word were repeated 5 times. The study phase had a total of 150 word pair study events and the exact numbers for each type of study pair are given in Table 1.

The recognition test included 160 word pairs that were either reinstated study pairs (intact) or were one of three different types of foils: rearranged, item, or novel. Foil-type was crossed with the repetition and fan manipulation factors with the constraint that rearranged pairs involved reassignment within the same level of fan condition. That is, a Fan 1-5 pair would be rearranged with another Fan 1-5 word pair. Word position within a pair was always preserved in rearranged and item foils. New, previously unstudied, words were used in constructing the item and novel foil stimuli. The recognition test probes were constructed so that all of the previously presented study words were used, either in intact word pairs or in foil stimuli.

Word pairs consisted of two words, each four to 12 characters in length (M = 6.1) 18-point font separated by a dash, presented in the center of the computer screen. The assignment of words to conditions (including unstudied) was randomly determined for each participant from a pool of 320 common nouns with word-frequencies (Kucera & Francis, 1967) between 55 and 95 occurrences per million (M = 71.3, SD = 11.8) generated from the MRC psycholinguistic database (Wilson, 1988). The order of presentation of word pairs and test pairs was also randomly determined for each participant. The dependent measures were proportion of responses and response latencies to each of the 5 response categories: (1) Old-Old (original), (2) Old-Old (rearranged), (3) Old-New, (4) New-Old, and (5) New-New. We focus here on the accuracy data, deferring discussion of the latency data until we have described the results of Experiment 2.

Procedure

Participants were informed prior to study that some word pairs would be repeated and that some words would be repeatedly presented, but each time with a different associate. During the study phase, word pairs were presented one at a time for 4 sec, followed by a 1.5 sec delay before the next pair was presented. After study and just prior to the recognition test, participants were told about the five possible response choices and also informed about the different types of foil stimuli. The test phase consisted of 160 test pairs (see Table 1), presented one at a time. The pair remained on the screen until the participant selected one of the five alternatives that were listed in the following order – Old-Old (original), Old-Old (rearranged), Old-New, New-Old, New-New – at the bottom of the screen by pressing one of the five designated keys. The screen cleared after the response and remained blank for .5 sec followed by a fixation stimulus of two dashed lines for 1 sec before the next test trial began.

Results and Discussion

Table 2 lists the proportion of responses to each of the five possible responses for the four different types of test probes: intact, rearranged, item, and novel word pairs. Each row represents the mean proportion of responses for each of the five possible responses for one of the fan or repetition conditions for a given type of test probe. Of the five choices of response, each type of test probe had its own correct response that is indicated in Table 2 in bold font. Note that for item word pairs, the correct response was either ‘Old-New’ or ‘New-Old’ depending on whether the reinstated word was on the left or the right. Visual inspection of the bolded correct responses (and 95% confidence intervals) in Table 2 strongly suggests that participants were able to correctly identify each type of test probe. A level of p < .05 was used for all significance tests, unless otherwise noted.

Table 2. Mean proportion of responses to each of the four word-pair types in the word-pair recognition test as a function of the number of associates (fan) for word1 and word2 and whether the pair was repeated in Experiment 1.

	Response
Word Pair:	Old-Old (original)	Old-Old (rearranged)	Old-New	New-Old	New-New
Intact Pair
Fan 1-1	.45 (.09)	.09 (.04)	.09 (.04)	.09 (.03)	.28 (.06)
Fan 1-5	.44 (.09)	.18 (.05)	.04 (.02)	.25 (.07)	.09 (.06)
Fan 5-1	.44 (.07)	.17 (.05)	.29 (.07)	.04 (.02)	.06 (.04)
Fan 5-5	.55 (.09)	.34 (.08)	.03 (.02)	.06 (.03)	.01 (.02)
Rep ×5	.84 (.06)	.06 (.02)	.03 (.02)	.03 (.02)	.02 (.02)
Rearranged Pair
Fan 1-1	.06 (.03)	.24 (.06)	.19 (.04)	.16 (.04)	.34 (.06)
Fan 1-5	.13 (.05)	.41 (.08)	.03 (.03)	.34 (.07)	.07 (.05)
Fan 5-1	.14 (.05)	.34 (.07)	.41 (.07)	.04 (.02)	.06 (.03)
Fan 5-5	.32 (.08)	.54 (.08)	.06 (.03)	.04 (.02)	.04 (.03)
Item Pair
Fan1-New	.04 (.02)	.12 (.05)	.29 (.07)	.06 (.03)	.48 (.08)
New-Fan1	.04 (.04)	.09 (.04)	.09 (.04)	.31 (.07)	.45 (.08)
Fan5-New	.04 (.03)	.19 (.06)	.65 (.09)	.04 (.03)	.09 (.04)
New-Fan5	.08 (.04)	.16 (.05)	.04 (.03)	.58 (.10)	.13 (.07)
Novel Pair
New-New	.04 (.02)	.07 (.03)	.10 (.03)	.09 (.03)	.70 (.09)

Correct responses are shown in bold. Variability is listed in parentheses as 95% confidence intervals.

Although the pattern of results in Table 2 strongly suggests that the assignments of correct (in bold) and incorrect response proportions are not due to chance (see 95% confidence intervals in parentheses), we developed a way to assess the reliability of these numbers. If participants were able to distinguish among the various word pair types, then the proportion of correct responses should differ reliably from the proportion of responses to novel word pairs for that particular response category. That is, we took the proportion of incorrect responses to novel word pairs as an estimate of response bias for each response category. This analysis examined the proportion of correct and incorrect responses for stimuli in which words were studied once² – the intact (Fan 1-1), rearranged (Fan 1-1), and item (Fan 1-new, new-Fan 1) word pair types – compared to novel (new-new) word pairs. Multiple pairwise comparisons using the Dunnett (1955) procedure established that the proportion of correct responses for each type of word pair was significantly different from the response bias to novel word pairs. All correct responses were significantly different from the response bias to novel word pairs.³ For example, the proportion of hit ‘Old-Old (original)’ responses to intact (Fan 1-1) word pairs (M = .45) differed from the proportion of incorrect ‘Old-Old (original)’ responses to novel (new-new) word pairs (M = .04). Just as the correct responses were reliably different from the response proportions to novel pairs, the incorrect responses were not. This finding adds more evidence that participants were able to calibrate responses to the various stimuli. For example, the proportion of incorrect ‘Old-Old (rearranged)’ responses to intact (Fan 1-1) word pairs (M = .09) did not differ from those to novel (new-new) word pairs (M = .07).

In summary, the proportion of correct responses was significantly different from baseline for each type of word pair, whereas the proportions of incorrect responses were not, clearly demonstrating the ability of participants to distinguish among all five word pair stimulus-types on the recognition test. We now examine the roles of other factors in the study, specifically the effects of associative interference and repetition strengthening on performance, and turn to our first question: How are item and associative sources of information represented in memory?

The Representation of Item and Associative Information in Memory

A Discrete Strengthening of Associative Information

A comparison of Rep ×5 and Fan 5-5 word pairs indicates that associative information can be discretely strengthened in memory, as separate from a strengthening of the underlying item memory representation. Repetition was found to increase the availability of associative information, aside from the increased availability of item information. For instance, the five-fold repetition of an intact word pair (Rep ×5 condition) resulted in significantly higher proportions of hits [Old-Old (original) responses] than in the intact Fan 5-5 condition, F(1, 67) = 25.16, MSE = 1.41. In both cases, each word was studied five times, so it can be assumed that the underlying strength of the two words was equivalent. Thus, the increase in hit rate observed for intact Rep ×5 word pairs compared to intact Fan 5-5 word pairs was due to increased success in retrieving the association. A key difference was that the same association has been seen five times in the Rep ×5 condition. This result supports a local memory representation (see Figure 4, panel B), such that item and associative information are discretely stored as separate memory representations. We now examine our second question: How are item and associative information used in recognition decision-making?

Figure 4.

Schematic description of a candidate global model, REM (panel A) and a local model SAC (panel B) of associative recognition. Leftmost section headers delineate the memory representations of item and associative information for three classes studied word pairs (Fan 1-1, Rep ×5, and Fan 5-5) in the REM (sections A1, A2, A3) and SAC (sections B1, B2, B3) models. The results of the Monte Carlo simulation are given on the right.

Signal Detection Analysis of the Contribution of Associative Information to Recognition

In this analysis, we compared ‘Old-Old (original)’ responses to intact and rearranged word pairs. Hits were defined as ‘Old-Old (original)’ responses to intact word pairs and false alarms were defined as ‘Old-Old (original)’ responses to rearranged word pairs. We examined whether associative interference impaired the ability of participants to correctly or falsely recognize word pairs. Hits and false alarms were examined as a function of increasing associative interference (i.e. the number of overlapping word pairs) – Fan 1-1, (Fan 1-5, Fan 5-1), and Fan 5-5. Figure 1 (panel A) shows a large incremental increase in the proportion of false alarms whereas the proportion of hits increased only for Fan 5-5 pairs. The proportion of hits did not significantly increase across the three levels of the fan manipulation, F(2, 134) = 2.33, MSE = 0.15, but the proportion of false alarms did significantly increase with associative interference, F(2, 134) = 23.38, MSE = 0.63. Unfortunately, this analysis could not be extended to the Rep ×5 condition as our design did not include rearranged Rep ×5 word pairs (but see Experiment 2). A series of contrasts indicated that false alarms increased significantly between the first two levels of the fan manipulation – Fan 1-1 and (Fan 1-5, Fan 5-1) – F(1, 100) = 4.01, MSE = 0.11, but there was no increase in hits, F(1, 100) = 0.04, MSE = 0.002. However, between the last two levels of the fan manipulation – (Fan 1-5, Fan 5-1) and Fan 5-5 – significant increases were observed for both the hits, F(1, 100) = 4.38, MSE = 0.28, and the false alarms, F(1, 100) = 30.21, MSE = 0.82. This result suggests that interference from the fan manipulation may impair the retrieval of associative information, since we did not observe a linear increase in hit rate with pair strength. By contrast, the false alarm rate did increase as a function of word pair strength.

Figure 1.

Associative recognition results of Experiment 1. Panel A. Proportion of ‘Old-Old (original)’ hit responses to intact word pairs and ‘Old-Old (original)’ false alarm responses to rearranged word pairs as a function of associative interference (Fan 1-1, [Fan 1-5, Fan 5-1], Fan 5-5). Panel B. Mean d′ statistic as a function of associative interference (fan). The error bars are standard errors of the mean.

Panel A. Proportion of ‘Old-Old (original)’ hit responses to intact word pairs and ‘Old-Old (original)’ false alarm responses to rearranged word pairs as a function of associative interference (Fan 1-1, [Fan 1-5, Fan 5-1], Fan 5-5). Panel B. Mean d′ statistic as a function of associative interference. The error bars are standard errors of the mean.

The false alarm rate in Figure 1 (panel A) reflects false ‘original’ attributions based on the memory strength of the items in the pair. However, the hit rate also includes correct ‘original’ attributions based on this memory strength of the items in the pair. That is, the false alarm proportion is subsumed in the hit rate, which has the additional contribution of associative information to the recognition process. Thus, the contribution of associative information can be determined by the difference in hits and false alarm rates.⁴ We calculated d′(Swets, 1961; Tanner & Swets, 1954), a common statistical expression of signal detection (hits) out of noise (false alarms), to determine the sensitivity of the recognition process to associative information across the fan conditions. Shown in Figure 1 (panel B), d′ declined monotonically as a function of increasing associative interference – Fan 1-1, (Fan 1-5, Fan 5-1), and Fan 5-5, F(2, 133) = 4.13, MSE = 2.67.

How is Item Information Used in Associative Recognition?

Previous studies have shown that repetition increases the ‘old’ hit rate in ‘old’/’new’ judgments. The increased specificity of the recognition response in our 5-PAR procedure allows us to examine whether repetition increases the general item strength of a word pair [Old-Old (rearranged) response] and whether participants are able to make more specific item-based recognition decisions (‘Old-New’ and ‘New-Old’ responses). Our analysis used multiple pairwise comparisons with the Tukey-Kramer correction (Kramer, 1956; Tukey, 1953) to examine the response proportions listed in Table 2 across the fan manipulations (i.e. as a function of repetition) for intact, rearranged, item and novel test pairs. For example, for rearranged word pairs, we examined the proportions of correct ‘Old-Old (rearranged)’ responses across the repetition (fan) manipulation (Fan 1-1, Fan 1-5, Fan 5-1, Fan 5-5): 24%, 41%, 34%, 54%.

The results of the multiple pairwise comparisons (p < .05) indicate that the underlying memory strength of words increased as a function of repetition, even though in the fan conditions the words were studied each time with a different associate. First, with repetition, participants made fewer incorrect ‘New-New’ responses and more correct responses. Second, we examined unbalanced word pairs (e.g. Fan 5-new, new-Fan 5, or Fan 1-5, Fan 5-1) where the item strength of the left and right word in the pair were differentially strengthened. Participants were able to correctly identify unbalanced word pairs and assess each word independently by responding ‘Old-New’ or ‘New-Old’ to discern the stronger of the two words in the pair. For item word pairs, in which the studied word had only been studied in a single pair (i.e. Fan 1-new, new-Fan 1), participants were able to discriminate the position of the new versus old item (‘Old-New’ or ‘New-Old’ hit responses, respectively). And, the identification of item word pairs was reliably better when the studied word was presented in five different pairings (i.e. Fan5-new, new-Fan5) (M = .61 versus .30), F (2, 133) = 58.05, MSE = 3.30, than Fan1-new, new-Fan1. These results indicate that associative recognition decisions can also involve an independent assessment of the specific item strength of each word in the pair. This result has implications for current memory models which are discussed further below.

Third, the pattern of errors also shows that participants were sensitive to the differential strength of the two words in an unbalanced pair. For example, when participants missed recognizing an intact pair, they were more likely to assume that the word presented only once was new and to judge the word that was seen five times as old. As can be seen in Table 2, participants were extremely good at discriminating Fan 5-1 from Fan 1-5 in all types of targets and foils. Further, there was no evidence of bias in tendency to select the right or left member of the word pair.

These results indicate that associative recognition memory decisions can involve an independent assessment of the specific item strength of each word in the pair. If item-based recognition decisions were based on an aggregation of item information (word₁ + word₂), without regard to the left or right word source of item familiarity, then we would expect the ‘Old-New’ and ‘New-Old’ response proportions to be indistinguishable. Overall, these results also support the conclusion that, in addition to decreasing the availability of associative information, a second effect of the fan manipulation was to increase both item-based [Old-Old (rearranged)] and item-specific (Old-New, New-Old) responding by increasing item strength through stimulus repetition.

Experiment 2

The results of Experiment 1 raise a number of questions that we addressed in Experiment 2. The results of Experiment 1 suggested that item and associative sources were dissociable as our interference manipulation affected one kind of retrieval (association) but not the other (item). Experiment 2 includes rearranged Rep ×5 word pairs, which were repeatedly presented together five-times during study, but re-paired with another word as a foil pair for the recognition test. This new condition will allow us to obtain a measure of false associative recognition, the proportion of ‘Old-Old (original)’ response to rearranged Rep ×5 word pairs.

This condition can critically distinguish between dual-process and composite-strength models, which make opposite predictions. Composite strength models assume that ‘Old-Old (original)’ responses are reserved for word pairs with the highest overall strength. Based on this assumption, rearranged Rep ×5 word pairs are expected to have a false alarm rate that is similar to rearranged Fan 5-5 word pairs, which has equivalent overall item strength. Dual-process models assume that for rearranged word pairs, associative information is either not retrieved or, if successfully retrieved, is used to oppose the item strength of the foil pair. Even if this process is only intermittently successful, we would expect the associative false alarm rate to be substantially lower for Rep ×5 word pairs compared to Fan 5-5 word pairs, since opponent processing for Rep ×5 word pairs benefits from a five-fold strengthening of associative information. It is also possible that these two accounts can be distinguished based on response time. Decision latencies should be prolonged if participants are engaged in opponent processing (i.e. recall-to-reject strategy). Thus, the second goal of Experiment 2 was to examine decision latencies for making the various five recognition decisions, which may help inform our earlier findings. For instance, participants were able to dissociate the specific item strengths of the right and left words in the pair by responding ‘Old-New’ and ‘New-Old’ in an associative recognition judgment. One possibility is that in response to the task demands, each word is queried in memory after first interrogating the word pair. In this case, we would expect average decision latencies for response categories pertaining to associative information (i.e. ‘Old-Old (original)’ and ‘Old-Old (rearranged)’) to occur before those indicating item information (i.e. ‘Old-New’ and ‘New-Old’). A study of latencies in recognition decision-making allows us to examine this possibility, among others.

Method

Participants

Thirty participants (8 males) recruited from the Pitzer College community, participated in exchange for $10.

Materials and Design

The materials and procedures of Experiment 2 were identical to those of Experiment 1 with the exception of the additional foil condition described above: rearranged Rep×5. This necessitated that during study, an additional ten word pairs were repeatedly presented with the same associate five times each, lengthening the study list to 200 study events. At test, half of the Rep ×5 pairs (i.e. 10 of the 20 word pairs) were re-assigned so that a word was presented with another Rep ×5 word, preserving their respective left-right positions in the pair. With these additional 10 foils, there were 170 test trials. As in Experiment 1, assignment of words to condition and foil condition and the order of study and test were randomly determined for each participant.

Results and Discussion

The results of Experiment 2 are presented in Table 3 in an analogous fashion to Table 2 for Experiment 1. Table 3 contains one additional row to present the results from the new condition of rearranged Rep ×5 word pairs.

Table 3. Mean proportion of responses to each of the four word-pair types in the word-pair recognition test as a function of the number of associates (fan) for word1 and word2 and whether the pair was repeated in Experiment 2.

	Response
Word Pair:	Old-Old (original)	Old-Old (rearranged)	Old-New	New-Old	New-New
Intact Pair
Fan 1-1	.45 (.07)	.11 (.05)	.06 (.03)	.10 (.03)	.28 (.07)
Fan 1-5	.40 (.07)	.19 (.05)	.04 (.02)	.29 (.06)	.07 (.05)
Fan 5-1	.40 (.07)	.19 (.07)	.31 (.06)	.02 (.01)	.08 (.05)
Fan 5-5	.53 (.08)	.37 (.07)	.04 (.03)	.05 (.02)	.01 (.01)
Rep ×5	.84 (.07)	.06 (.03)	.04 (.02)	.03 (.02)	.03 (.02)
Rearranged Pair
Fan 1-1	.06 (.03)	.20 (.05)	.19 (.04)	.20 (.04)	.35 (.07)
Fan 1-5	.13 (.04)	.35 (.08)	.04 (.02)	.42 (.07)	.07 (.05)
Fan 5-1	.13 (.05)	.32 (.07)	.45 (.06)	.02 (.01)	.07 (.03)
Fan 5-5	.26 (.08)	.61 (.08)	.05 (.03)	.06 (.03)	.01 (.01)
Rep ×5	.12 (.04)	.53 (.08)	.14 (.04)	.14 (.04)	.08 (.04)
Item Pair
Fan1-New	.02 (.03)	.08 (.03)	.29 (.07)	.08 (.02)	.53 (.08)
New-Fan1	.04 (.03)	.07 (.03)	.06 (.03)	.31 (.07)	.52 (.08)
Fan5-New	.04 (.03)	.17 (.06)	.69 (.08)	.02 (.02)	.08 (.05)
New-Fan5	.07 (.04)	.10 (.04)	.02 (.01)	.72 (.08)	.09 (.05)
Novel Pair
New-New	.01 (.01)	.05 (.02)	.10 (.03)	.10 (.02)	.74 (.07)

Correct responses are shown in bold. Variability is listed in parentheses as 95% confidence intervals.

The Response Proportions of Experiment 2 Replicate Those of Experiment 1

Our initial analysis examined whether the results of Experiment 2 replicated those of Experiment 1. Multiple pairwise comparisons with the Tukey-Kramer correction (Tukey, 1953; Kramer, 1956) compared the corresponding proportions of five choice responses to each word pair type from Table 3 (Experiment 2) to those in Table 2 (Experiment 1). All of the response proportions in Experiment 2 closely matched those of Experiment 1. No value listed in Table 3 was significantly different from the corresponding value listed in Table 2. We also replicated prior results by subjecting the data of Experiment 2 to the same set of analyses as in Experiment 1. The pattern of findings were identical at a significance level of p <.05. To save space, these results are not reported, but can be obtained upon request to the first author. Below, we examine selected accuracy results and then move on to the latency results of Experiments 1 and 2. First, with the inclusion of rearranged Rep×5 word pairs, the results of Experiment 2 bear on the availability of associative information and the nature of false associative recognition. Second, we examine the time-course of associative recognition decision-making.

Associative Attribution Errors Increase with Interference, not Memory Strength

Our study measured false associative recognition, whereby a participant mistakenly reports that a word pair was previously associated during study when it was not. For example, this occurs when a participant makes an ‘Old-Old (original)’ response to a rearranged word pair foil. Experiment 2 included rearranged Rep ×5 word pairs. Participants correctly identified intact Rep ×5 word pairs 84% of the time. When these words were reconstituted as part of a rearranged word pair foil, participants false alarmed only 12% of the time. Rep ×5 word pairs can be compared to Fan 5-5 word pairs to demonstrate the gains in recognition performance due to associative strengthening. For instance, the hit rate to intact Fan 5-5 word pairs was much lower, at 53%, and the false alarm rate to rearranged Fan 5-5 word pairs was much higher, at 26%.

The Experiment 2 results are presented in Figure 2 (panel B).⁵ The mean d′ statistic for the Rep ×5 word pairs was 2.29 with a standard error of 0.15. Contrasts of the d′ statistic demonstrated that participants were better able to discriminate the Rep ×5 word pairs than either Fan 1-1 word pairs, F(1, 59) = 32.64, MSE = 15.72, or Fan 5-5 word pairs, F(1, 59) = 66.47, MSE = 32.02. It is evident in Figure 2 (panel A) that the hits to Rep ×5 word pairs were much higher than the hits to both the Fan 1-1 word pairs, F(1, 59) = 49.65, MSE = 2.32, and the Fan 5-5 word pairs, F(1, 59) = 30.84, MSE = 1.44. Participants made fewer false alarms to Rep ×5 word pairs than Fan 5-5 word pairs, F(1, 59) = 14.08, MSE = 0.32, but made about the same number of false alarms in the Fan 1-1 and Rep ×5 conditions, F(1, 59) = 2.10, MSE = 0.05. Constancy in false alarms across repetition conditions has been previously reported (e.g. Kelley & Wixted, 2001; Light et al., 2004; but see Malmberg & Xu, 2006).

Figure 2.

Associative recognition results of Experiment 2. Panel A. Proportion of ‘Old-Old (original)’ hit responses to intact word pairs and ‘Old-Old (original)’ false alarm responses to rearranged word pairs as a function of associative interference (Rep×5, Fan 1-1, [Fan 1-5, Fan 5-1], Fan 5-5). Panel B. Mean d′ statistic as a function of associative interference (fan). The error bars are standard errors of the mean.

Panel A. Proportion of ‘Old-Old (original)’ hit responses to intact word pairs and ‘Old-Old (original)’ false alarm responses to rearranged word pairs as a function of increasing associative interference (Rep ×5, Fan 1-1, [Fan 1-5, Fan 5-1], Fan 5-5). Panel B. Mean d′ statistic as a function of associative interference. The error bars are standard errors of the mean.

False associative recognition was found to increase with associative interference (see ‘false alarms’ in Figure 2, panel A) or the number of overlapping associates paired with a given word. Words studied with many different associates were more likely to be falsely judged as part of an intact word pair; an effect that increased with the number of associates (fan). By comparison, participants demonstrated little false associative recognition when the same association was strengthened five-times by repetition (i.e. Rep ×5 word pairs). This result discounts the possibility that false associative recognition is based only on item strength. If this were the case then the proportions of false associative recognition would be equivalent in the Fan 5-5 and Rep ×5 conditions. A key difference between the conditions was that the recognition of Fan 5-5 word pairs was subject to associative interference but not Rep ×5 word pairs.

As noted above, the Fan 1-1 and Rep ×5 word pair conditions had equivalent false alarms, whether one association was presented or the same association was presented five times. This result supports the conclusion that false associative recognition is determined in part by the number of available associations (see also Malmberg & Xu, 2006). This suggests that for rearranged word pairs, participants retrieved an associate, but perhaps not the correct one. Retrieving the correct associate would allow participants to use a recall-to-reject strategy to correctly identify and not falsely associate a rearranged word pair foil. It is likely that false associative recognition involves retrieving the wrong associate, a point we return to later.

An Analysis of Decision Latency in Associative Recognition

The times taken to make the five types of recognition judgments were examined in response to the recognition test probes (intact, rearranged, item, novel) across the fan manipulation test conditions. The mean median response latencies are displayed in Figure 3 for correct responses across both Experiments 1 and 2. The patterns of latency data across Experiments 1 and 2 were identical for the conditions shared by the two studies, which justifies an aggregation of the data to reduce standard errors. Because the Rep ×5 rearranged lures were present only in Experiment 2, no aggregation was possible for these items and, not surprisingly, the error bars are larger for this Experiment 2 condition.

Figure 3.

Mean median response latency for correct responses to intact, novel, rearranged, and item tested word pairs. Variability is shown as 95% confidence intervals.

There was a clear ordering in the time taken to make various types of recognition decisions across the test conditions: Old-Old (original) < New-New < Old-Old (rearranged) < Old-New ≈ New-Old. Participants were fastest at deciding whether the words were previously associated. Participants were also fast at assessing novelty when neither word of the pair was familiar. This novelty detection was more rapidly than deciding that the words were both old but not studied together. They were slowest when one word was old and one word was new. One striking finding was the speed of the novelty response. This finding is consistent with two standard accounts of choice RT (see Ratcliff, 1978) – random walk models and diffusion models – that predict fast responses for response categories that are at extreme values (i.e. ‘New-New’ and ‘Old-Old (original)’ responses) rather than at an intermediate level of evidence.⁶ Furthermore, it is possible that participants could more quickly decide ‘new-new’ based on low item strength than ‘Old-Old (rearranged)’ based on high item strength: Only if the words seem familiar would a search be initiated to see whether they were associated. The overall ordering of the time taken to make the various types of recognition decisions was also maintained for all incorrect responses (not shown) across all of the test conditions.⁷ See the appendix for the full tabulation of correct and incorrect RT data.

As shown in Figure 3, RTs generally increased with the repetition manipulation.⁸. For intact test pairs, there was marginal speedup in correctly identifying previously associated study pairs across the three levels of fan repetition (Fan 1-1, [Fan 1-5, Fan 5-1], Fan 5-5), F(2, 246) = 2.16, MSE = 1880405. Associative strengthening significantly sped up recognition as participants were faster at correctly identifying intact Rep ×5 word pairs than Fan 5-5 word pairs, of equivalent item strength, F(1, 125) = 5.16, MSE = 39134371.

For rearranged test pairs, the pattern of results was different. There was a marked decrease in the time taken to correctly identify rearranged study pairs across the repetition (Fan 1-1, [Fan 1-5, Fan 5-1], Fan 5-5) conditions, F(3, 232) = 6.44, MSE = 11547132. This was not the case, however, for the rearranged Rep ×5 test pairs. Compared to the Fan 5-5 conditions, which has equivalent item strength, the identification of Rep ×5 word pairs was considerably slower, F(3, 232) = 6.44, MSE = 11547132, and more error prone. We take this dramatic reversal as compelling evidence that participants are using opponent processing – a recall-to-reject strategy – to correctly classify these responses. When we consider the prolonged decision latencies to rearranged Rep ×5 foil pairs, it is clear that participants are able to use associative information to effectively reject some foils, such as rearranged Rep ×5 word pairs, but not others, such as Fan 5-5 pairs. This phenomenon solidifies the case for a dual-process interpretation of our data.

For the item pairs, either the left or the right word was strengthened five times and paired with a novel word. The recognition latencies of item test pairs improved with the five-fold strengthening of the item, for the left word, F(1, 122) = 7.41, MSE = 15280304, but only marginally so for the right word, F(1, 117) = 2.71, MSE = 4957994. It is possible that this last latency result reflects the sequential reading of the words from left to right in a test pair.

General Discussion

In two experiments, our results demonstrate that associative-recognition makes use of three discrete sources of information – word₁, word₂, association – each of which could be distinguished at test, especially if they were selectively strengthened during the study period. Participants were able to accurately distinguish source contributions (word₁, word₂, association) by selecting among five recognition responses: (1) Old-Old (original), (2) Old-Old (rearranged), (3) Old-New, (4) New-Old, and (5) New-New. Participants recognized studied word pairs (intact pairs) and three types of foils composed of two (rearranged pairs), one (item pairs), or no (novel pairs) previously studied words by choosing the correct five-choice recognition response to each type of word pair. The retrieval of associative information was necessary to distinguish intact pairs from rearranged pair foils. And the discrimination of various word pair foils required that participants separately consider the item strength of each word in the pair. Furthermore, our latency results demonstrate that there is a specific time-course for making various recognition decisions.

We were able to further dissociate item strength from associative strength by manipulating how many times a specific word or word-pair was repeated (1 or 5) and how many different word pairs were associated with a given word (1 or 5). Generally, as a result of our five-fold repetition of words and associations, participants were increasingly able to distinguish among the item and associative sources. In particular, we found that the association could be strengthened separately from the items, and this produced a corresponding increase in the retrieval of the association. Item information could also be distinguished in unbalanced word pairs, in which the right or left word was differentially strengthened. Participants were able to discern the stronger word in the pair. These results suggest that item and associative sources of information are stored as part of separate memory representations, as each can be discretely strengthened and retrieved

In the associative interference conditions, one or both words in a word pair were studied five times, each time with a different associate. The results of the associative interference (fan) manipulation suggest that item and associative sources of information also make separate contributions at retrieval. This was demonstrated by an associative interference manipulation that adversely affected the retrieval of one kind of information (i.e. association) but not the other (i.e. item). Associative interference decreased the retrieval of the association, which was necessary for identifying intact word pairs, and also increased the tendency to make errors. In some cases, participants incorrectly responded that rearranged word pair foils were previously studied when they were not. This increase in false associative recognition was not due to the underlying item strength of the two words in the pair (but see Kelley & Wixted, 2001). For instance, although item strength was identical in rearranged Rep ×5 and Fan 5-5 word pairs, participants were much more likely to incorrectly associate word pairs in the interference Fan 5-5 condition. A second effect of associative interference was to strengthen items as a function of how often a given word was repeated. These two results replicate the work of Verde (2004) who demonstrated that associative interference generates opposing effects on remember and know (R/K) judgments: a decrease in the number of ‘remember’ responses but an increase in the number of ‘know’ responses. As R/K responses have been found to correspond to the dual-processes of recollection and familiarity (e.g. Hockley & Consoli, 1999), Verde's results suggest that associative interference decreases recollection and increases familiarity-based responding.

Simulations of Local and Global Matching Models of Associative Recognition

A major theoretical advance in our understanding of human memory has been the development and refinement of formalized models. Given both the complexity of the decision-making in our 5-PAR dataset and the fact that models can sometimes make surprising and counter-intuitive claims, the implications of these findings will not be fully realized until formal models are fit to the full 5-PAR dataset, something that is beyond the scope of our project. As a first step, however, we derived model predictions from Monte Carlo simulations of both a global matching model of memory, REM (Steyvers & Shiffrin, 1997), and a local memory model, SAC (Reder et al., 2000). Our simulation focused on the key comparison between two conditions, the Rep ×5 and Fan 5-5 pairs.

The mechanics of the REM model are described in Figure 4 (panel A). Item information is represented as a vector of feature weights arbitrarily pulled from a geometric distribution. During encoding, a subset of these features (i.e. non-zero weights) is successfully transcribed to a memory representation of the item, but the model also allows for transcription errors (see each item feature in last position in A1). With repetition, more and more features are correctly transcribed to the item memory representation (see A2 and A3). Associative information is represented as two conjoined item vectors, and it can be seen in Figure 4 (panel A) that the REM model produces identical item and associative information for Rep ×5 and Fan 5-5 word pairs. However, memory decisions in global matching models such as REM are based on an overall familiarity value (i.e. Bayesian likelihood) obtained by comparing the test item to the entire contents of a memory array. Thus, although the item and associative memory representations are identical for Rep ×5 and Fan 5-5 word pairs, our modeling efforts suggest that the REM model predicts a higher hit rate for Fan 5-5 word pairs since the overlap with other word pairs on the list results in higher familiarity (i.e. likelihood) values. Specifically, the Monte Carlo simulation⁹ compared the Fan 5-5 and Rep ×5 test probes when the list contained 24 word pairs that had been studied once. The predicted hit rate, the proportion of target pairs above threshold (φ > 1), was equivalent for Fan 5-5 word pairs (49%) and Rep ×5 word pairs (48%). This simulation result suggests that the model representations of associative information is similar in both conditions. Next, we simulated Rep ×5 word pairs by examining performance when this target pair was also studied with 24 other unique Rep ×5 word pairs. For comparison, we also simulated the recognition of Fan 5-5 word pair as these word pairs were studied with the full set of 24 other overlapping Fan 5-5 word pairs. The predicted hit rate, was much higher for Fan 5-5 word pairs (86%) than that for Rep ×5 word pairs (69%). Therefore, our efforts suggest that the REM model incorrectly predicts a higher hit rate in the Fan 5-5 condition relative to the Rep ×5 condition.

In contrast, a Monte Carlo simulation of the local SAC model generates the desired prediction of a lower hit rate for Fan 5-5 word pairs compared to Rep ×5 word pairs. As described below, a key feature of the local SAC model is that it offers a mechanism for generating interference in the retrieval of distinct associative information. Shown in Figure 4 (panel B) the SAC model has separate representations of item (semantic nodes) and associative (episodic nodes) information. Model equations (see Reder et al., 2000) are used to calculate the representation of item and associative information, such as the residual and current node activations, number of links, and link strengths. That is, as a starting point, the semantic node activations (and number of links) for item information are based on the reported normative word frequency of occurrence of a particular word in the lexicon (e.g. Kucera & Francis, 1967). When a word pair is presented at test, activation spreads from the relevant semantic nodes to any linked episodic nodes, and the amount of spreading activation is determined in proportion to the number and strength of the links to other episodic associates. Thus, the amount of spreading activation contributing to an episodic node is much less in the case of Fan 5-5 word pairs (16% spreading activation) than Rep ×5 word pairs (78% spreading activation). Memory decisions are determined by two separate processes, whether the distribution of activation of particular semantic nodes (item familiarity) or episode nodes (associative recollection) surpasses a decision threshold. Both the variability of the distributions and the decision threshold are fitted parameters. The SAC model Monte Carlo simulation was run by calculating the memory representations of words and word pairs randomly generated from our stimulus list. Although the item familiarities for Fan 5-5 and Rep ×5 word pairs are identical, the simulation demonstrates that the resulting average activation strength of the association (episode node) is much less for Fan 5-5 word pairs. Therefore, the SAC model correctly predicts a lower hit rate for Fan 5-5 word pairs than Rep ×5 word pairs. In sum, our Monte Carlo simulations suggest that two candidate global and local models of memory generate opposite predictions for the hit rate for Fan 5-5 and Rep ×5 word pairs. Our goal is not to prove one model correct and another wrong, but rather, to delineate the set of mechanisms necessary to account for this dataset. Toward this end, a more formal and complete analysis of various models on the full 5-PAR dataset is necessary.

To highlight the incremental nature of memory modeling, a dual-process version of the REM model has recently been proposed by Malmberg, Holden, and Shiffrin (2004). Although it is our understanding that this version of the REM model does not change the predictions described above, an additional mechanism is proposed that can account for some aspects of the 5-PAR latency data. When targets and foils are known to be very similar, a secondary recall mechanism is invoked. For instance, Xu and Malmberg (2006) proposed that if the conjoined association vector exceeds threshold (Φ > 1) then specific traces are sampled using a recall-to-reject strategy (see discussion below). This account is compatible with the prolonged decision times for our rearranged word pairs. However, it is unclear whether this sampling process would also prolong decision times for intact word pairs, which our data show to be the fastest. Furthermore, an additional layer of decision-making may be necessary to account for the item-specificity of the recognition response (‘old-new’, ‘new-old’) to item foil pairs.¹⁰

Two Separate Assessments of Item Strength

An important feature of the 5-PAR paradigm is that the task demands require both associative and single-word recognition. In the context of an associative recognition task, participants were able to discern item strength and distinguish the stronger word in the pair. For instance, participants correctly matched unbalanced item word pairs (Fan 1-new, new-Fan 1, Fan 5-new, new-Fan 5) with ‘Old-New’ or ‘New-Old’ responses. Even incorrect responses to unbalanced intact and rearranged word pairs (Fan 1-5, Fan 5-1), correctly matched the ‘Old-New’ and ‘New-Old’ responses to the stronger item in the pair. Clearly, participants separately assessed each word in the pair. Our data set challenges memory models to account for both types of recognition decisions within a single system of mechanisms. This is noteworthy since the associative and single-word recognition literatures have often been treated separately.

Our findings were not anticipated by memory models that aggregate sources of item information, including a number of dual-process models of associative-recognition (Kelley & Wixted, 2001; Yonelinas, 1997, 2002). All of these models describe the familiarity process as a continuous index of memory strength in which both item sources are pooled in forming a familiarity decision. The familiarity process is successful if this pooled item information surpasses a decision threshold. The item-specificity of our data would seem to be a challenge for models that aggregate item information. As an alternative to item strengthening, multiple-trace models such as MINERVA 2 (Hintzman, 1988) can, in principle, account for the item specificity observed in our data. In response to a test probe, recognition and frequency judgments are based on the intensity of an ‘echo’ or a composite response to all the separate item traces stored in memory. Hintzman (1988) maintains that recognition judgments are a special case of memory for frequency. Thus, a forced-choice recognition memory test is simply a frequency-discrimination task. In our task, multiple-trace models could be modified so that if a zero count is retrieved for associative information, then participants could in a second step, assess item frequency of the left and right word in the pair and respond ‘Old-Old (rearranged)’, ‘Old-New’, ‘New-Old’, or ‘New-New’ (see Hockley & Cristi, 1996).

Although multiple-trace models could account for our item strengthening results by assuming that the likelihood of retrieving one of several traces is higher, we suspect that multiple-trace models would have difficulty reproducing the results of our associative interference manipulation. Multiple-trace models represent associative information as separate traces (Hintzman, 1988) and the availability of associative information is based on a global matching process with the test probe. It is unclear whether increasing the number of associates would interfere with this global match process. Furthermore, associative interference was not evident for Rep ×5 word pairs, which would constitute five separate traces, but was evident in the Fan 1-5, Fan 5-1, and Fan 5-5 pairs. To differentiate Rep ×5 pairs, our data suggest that the same associative information has to be strengthened.

The global matching model TODAM2 (Murdock, 1993, 1997) also offers an alternative explanation for item specificity. That is, the composite memory trace can be queried using different retrieval cues, for instance, by either an associative cue (as in associative-recognition) or an item cue (as in single word recognition). It is possible that in response to the demand characteristics of having ‘old-new’ and ‘new-old’ as choices in the 5-PAR task, participants query memory using an item cue after querying the associative cue. This explanation is similar to what is proposed by dual-process models, and may serve as a way to bridge these views.

The results support both the representation and decision-making structure of local models such as SAC (Reder et al., 2000) that propose not only that item and associative information are repeatedly strengthened and retrieved as discrete units, but also offer an associative interference mechanism (diffusion of activation) that can account for fan effects. One limitation, however, is that previous SAC models of word recognition do not postulate spurious recollection (but see Reder & Schunn, 1996; Schunn et al., 1997 for accounts of spurious feeling-of-knowing judgments, and the analogous implementation of that idea in models described in Reder, Paynter, Diana, Ngiam & Dickison, 2007). Our data clearly indicate that false associative recollection occurs and increases with associative interference generated from the fan manipulation. Currently, the SAC model is being developed to provide an account of false associative recollection whereby the spread of activation in the model is enough to tangentially activate another episode node. We postulate that in the associative interference conditions, it is possible for any of the five episode nodes associated with a given word to get over threshold when lures with that word are presented (see Figure 4, panel B3).

Recall-to-Reject and False Associative Recognition

A number of recent findings from word pair recognition studies using repetition strengthening and associative interference manipulations are not easily reconciled with the assumption that recognition decisions are based on retrieved composite strength. Repetition can be used to further increase the hit rate and speed of recognizing studied word pairs. The strengthening of word pairs by three, four, or six-fold repetition increased the hit rates in studies of associative recognition (Gallo, Sullivan, Daffner, Schacter, & Budson, 2004; Kelley & Wixted, 2001; Light, Patterson, Chung, & Healy, 2004). A puzzling finding is that while the hit rate increases with repetition, the false alarm rate to rearranged word pairs remains constant for young adults. This is inconsistent with the composite strength assumption since the false alarm rate should also increase based on the strength of accumulated evidence. There is general consensus among the study authors that this phenomenon reflects opponent processing, such that associative information is used to disconfirm item information by a recall-to-reject strategy (Jones & Jacoby, 2001; Kelley & Wixted, 2001; Rotello & Heit, 2000; Rotello, Macmillan, & Van Tassel, 2000; Yonelinas, 1997, 2002). The recall-to-reject account proposes that rearranged word pairs (KEY-SMILE) are rejected because the correct association (KEY-LAMP or CHEMICAL-SMILE) is retrieved for one of the words in the rearranged word pair. The false alarm rate remains constant because the increase in item strength is offset by increased success in rejecting these word pairs based on associative information. This dual-process account suggests that associative information can be used to reduce memory errors. Our results provided compelling accuracy and latency data to suggest that participants do engage in a recall-to-reject strategy to reject rearranged Rep×5 word pairs. However, our results also suggest that participants were not able to effectively use this strategy under our experimental conditions of increasing associative interference (fan).

Why does acceptance of rearranged word pairs increase with associative interference? It may be that the sheer number of possible associations that can be retrieved complicates the rejection of word pair lures. A full disconfirmation involves an exhaustive retrieval of all associates. Objectively, increasing the fan manipulation would increase the chances of retrieving any one of five previously studied associates, which can then be used to disconfirm the lure presented as part of the rearranged or item word pair (c.f. Gallo, 2004) Although a number of researchers (e.g. Castel & Craik, 2003; Light et al., 2004) have discussed recall-to-reject as retrieval of the correct instance, the problem with a recall-to-reject strategy in our associative interference paradigm is that the participant has to retrieve all of the associates. This difficulty was borne out by our results which demonstrated increased false recollection across the levels of fan manipulation. With five different associates during the study phase, retrieving just one of the associations does not disconfirm the lure. It is possible that participants endorsed high-fan word pair foils based on the retrieval of one or more paired-associates matched with either or both words in the pair.

Conclusion

Our experimental results and procedures emphasized the decision-making aspects of memory, which we defined as the assignment of test stimuli to one of five recognition response categories. How memory is queried is often constrained by the experimental procedure used in the laboratory (see Orne, 1962). For this reason, methodological advances play an important role in scientific discovery. Our new 5-PAR paradigm offered a rich pattern of results that examined how item and associative sources of information are queried in memory to support recognition decision-making that will help refine and advance current models of associative recognition. The major theoretical contribution of this paper was a clear demonstration that these results cannot be accommodated by a model of association recognition whereby item and associative information are retrieved en masse. Our results advance a differentiated account of associative recognition – both in terms of the underlying memory representation and the specificity and timing of the decision-processes that operate upon them. Item (word1, word2) and associative sources of information made distinct contributions at retrieval and are stored as discrete units in memory.

Appendix

Table A. Mean median correct response latency to each of the four word-pair types in the word-pair recognition test as a function of the number of associates (fan) for word1 and word2 in Experiment 1.

	Response
Word Pair:	Old-Old (original)	Old-Old (rearranged)	Old-New	New-Old	New-New
Intact Pair
Fan 1-1	2421 (312)	3350 (447)	4927 (1000)	4409 (1035)	3219 (486)
Fan 1-5	2090 (238)	3607 (929)	5087 (1087)	4060 (662)	2122 (275)
Fan 5-1	2437 (286)	3293 (550)	3185 (310)	6288 (1654)	2995 (371)
Fan 5-5	1935 (174)	2743 (459)	8187 (2988)	3944 (643)	1208 (128)
Rep ×5	1701 (137)	2126 (260)	3152 (399)	7036 (1447)	2516 (326)
Rearranged Pair
Fan 1-1	2724 (398)	3544 (464)	4688 (824)	4580 (867)	2989 (412)
Fan 1-5	3092 (715)	3101 (323)	3479 (487)	3449 (494)	4043 (937)
Fan 5-1	3020 (768)	3155 (411)	3450 (370)	3587 (469)	3587 (551)
Fan 5-5	2612 (545)	2494 (225)	3699 (599)	4734 (520)	3002 (809)
Item Pair
Fan 1-new	3592 (1503)	3961 (920)	4010 (397)	4556 (819)	2649 (338)
new-Fan 1	2875 (592)	3668 (1321)	3073 (458)	4487 (493)	2651 (263)
Fan 5-new	2579 (323)	3055 (330)	3525 (381)	6754 (2053)	2876 (592)
new-Fan 5	2514 (428)	3619 (373)	3203 (508)	3519 (373)	4214 (1571)
Novel Pair
new-new	2140 (535)	3776 (687)	3742 (500)	4344 (578)	2444 (179)

Correct responses in bold. Variability is listed in parentheses as 95% confidence intervals.

Table B. Mean median correct response latency to each of the four word-pair types in the word-pair recognition test as a function of the number of associates (fan) for word1 and word2 in Experiment 2.

	Response
Word Pair:	Old-Old (original)	Old-Old (rearranged)	Old-New	New-Old	New-New
Intact Pair
Fan 1-1	2649 (288)	5586 (997)	6010 (1052)	7450 (1469)	4181 (962)
Fan 1-5	2502 (325)	4981 (1137)	5794 (929)	6217 (947)	3726 (787)
Fan 5-1	2715 (293)	6015 (2753)	5465 (907)	7272 (3522)	5065 (2575)
Fan 5-5	2464 (254)	4014 (924)	4704 (1163)	6023 (1267)	4494 (837)
Rep ×5	1977 (163)	3811 (676)	6825 (1915)	8514 (6006)	2553 (1164)
Rearranged Pair
Fan 1-1	4765 (1714)	4617 (553)	5650 (655)	6294 (840)	4200 (1025)
Fan 1-5	4236 (1544)	3852 (514)	6531 (2193)	6467 (1485)	6389 (3434)
Fan 5-1	3467 (1485)	3596 (560)	5034 (602)	4856 (1150)	3622 (887)
Fan 5-5	3479 (749)	3377 (348)	6349 (1989)	6369 (1951)	9974 (4805)
Rep ×5	2487 (287)	4216 (612)	6340 (1566)	6912 (1331)	5570 (1843)
Item Pair
Fan 1-new	3891 (1335)	6619 (1965)	6309 (900)	7697 (2053)	3644 (617)
new-Fan 1	3593 (2211)	6523 (2891)	6917 (1834)	5805 (933)	3949 (782)
Fan 5-new	3009 (637)	4231 (866)	4556 (483)	5007 (355)	5463 (1967)
new-Fan 5	3453 (980)	6463 (3030)	5865 (1893)	5411 (824)	4011 (1050)
Novel Pair
new-new	3689 (1624)	6531 (2692)	7433 (1684)	6880 (1116)	3357 (514)

Correct responses in bold. Variability is listed in parentheses as 95% confidence interval.