Explaining Recollection without Remembering

Abstract

Recollection without remembering is a counterintuitive phenomenon that violates a traditional assumption of source memory models—namely, that accurate item memory is a necessary precondition for remembering source details that accompanied an item’s presentation. The dual-recollection model explains recollection without remembering as a by-product of the contrasting effects of target and context recollection on item tests versus source tests. We pitted that explanation against two others that preserve the traditional assumption, one based on hypothesized testing artifacts and the other derived from multivariate signal detection theory. Our experiments focused on a manipulation that according to dual-recollection theory, should drive source memory and item memory in opposite directions. In two experiments, studied items were tagged with three source details (voice gender, taxonomic category, and list), such that either (a) the three details were consistent with each other or (b) one detail was inconsistent with the other two. As predicted, source memory was better but item memory was worse when source details were consistent with each other. The recollection without remembering effect was observed in both experiments, and as predicted by dual-recollection theory, it was more robust when item memory was worse than when it was better. A further instance of recollection without remembering was detected that involved distractors rather than presented items.

Historically, models of source memory have posited that it requires accurate item memory; that the conjunction of remembering an item’s source details without remembering the item itself is a null state (cf. Batchelder & Riefer, 1990; Bayen, Murnane, & Erdfelder, 1996). Dual-recollection theory does not make that assumption (Brainerd, Gomes, & Moran, 2014). It assumes instead that item memory is supported by a bivariate recollection process (target recollection and context recollection) and a familiarity process, whereas source memory is supported primarily by context recollection. Context recollection refers to the conscious reinstatement of realistic details that accompanied the prior occurrence of an item, either objective details (e.g., position, font, and color) or subjective ones (e.g., associations, thoughts, and emotions). Target recollection, on the other hand, refers to the conscious reinstatement of presented items per se.

The two types of recollection both produce realistic phenomenology that supports true memory for presented items. However, because they are distinct processes there should be manipulations that have opposite effects on them (Brainerd et al., 2014), which would have the counterintuitive effect of driving the accuracy of item and source memory in opposite directions. We discuss such a targeted manipulation below. It has also been shown that target and context recollection have opposite effects on the two most common forms of false memory, false memory for distractors that preserve semantic features of presented items and false memory for contextual details that accompanied presented items (Brainerd, Nakamura, & Gomes, 2015). With semantic false memory, context recollection supports such errors (e.g., Arndt, 2012), but target recollection suppresses them (“No, Coke and orange were not on the list because I clearly remember reading Pepsi and tangerine instead”; see Lampinen & Odegard, 2006). With a particularly difficult form of semantic false memory in which distractors overlap almost completely with the meaning and surface structure of list items (e.g., piano vs. pianos), affirmative rejection seems to demand target recollection. With the second type of false memory, source-monitoring errors, target recollection supports such errors (Brainerd et al., 2015), but context recollection suppresses them (“No, Pepsi and tangerine were not read in a male voice because I clearly remember hearing them in a female voice”; see Gallo, 2013). Brainerd et al. (2015) reported a series of experiments in which the two forms of false memory were dissociated by manipulations that selectively affected target versus context recollection.

Recollection without Remembering

Suppose that subjects study lists of items (e.g., words such as apple, feet, goat, and house) in two or more distinct contexts (e.g., apple and feet are read in a male voice, whereas goat and house are read in a female voice) and that they later respond to tests of item memory (e.g., old/new recognition) and source memory (e.g., recognizing items’ presentation voices). Curiously, subjects are able to remember the presentation contexts of studied items that are identified as new on item tests (e.g., goat and house are incorrectly judged to be new but their presentation voice is correctly judged to be female). This is known as recollection without remembering. The phenomenon is puzzling because it violates the aforementioned assumption that contextual details cannot be recollected unless the item itself can be remembered, and hence, it has been the focus of some recent studies (Brainerd et al., 2014; Ceci, Fitneva, & Williams, 2010; Cook, Marsh, & Hicks, 2006; Kurilla and Westerman, 2010; Starns, Hicks, Brown, & Martin, 2008).

In the research that we report, we continued this line of work by evaluating three explanations of recollection without remembering, two of which preserve the assumption that item memory is a necessary precondition for source memory. The first, which has been mentioned, falls out of dual-recollection theory. This theory predicts recollection without remembering on the ground that different retrieval processes have different effects on source and item memory—in particular, target recollection supports hits and suppresses false alarms on item tests, but it increases false alarms more than hits on source tests by fomenting acceptance of proffered sources when subjects cannot recollect contextual details. The other two accounts are an explanation derived from multivariate signal detection theory by Starns et al. (2008) and an artifactual explanation that was discussed by Kellen, Singmann, and Klauer (2014) and by Malejka and Bröder (2016). We briefly sketch these latter explanations and related findings before reporting our experiments.

Multivariate Signal Detection Explanation

According to the multivariate signal detection model of item and source memory (e.g., Slotnick & Dodson, 2005), the same retrieval processes support both, which preserves item-source dependency. Those processes are continuous variables, so that a decision criterion must be applied to the information that accumulates from memory on source and item tests. For experiments in which items are presented in two contexts, A and B, this model posits three bivariate normal memory strength distributions and two memory discrimination dimensions. The three strength distributions are for Source A items, Source B items, and new items (distractors), whereas the two discrimination dimensions are for item tests (discriminating old from new) and source tests (discriminating Source B from Source A). The mean strength of one of the two source distributions, Source B by convention, is assumed to be greater than the other. On an item test for a Source B item, subjects sample a value from the Source B distribution, and the item is judged to be old if the value exceeds the decision criterion between the Source B distribution and the new distribution. On a source test for the same item, subjects again sample a value from the Source B distribution, and the item is judged to be from Source B if the value exceeds the decision criterion between the two source distributions. Holding the two decision criteria constant, it is apparent that the accuracy of item memory and the accuracy of source memory both depend on the strength of the Source B distribution, and hence, increases in one will be associated with increases in the other.

Analysis of this model reveals that it is possible for some of the values that are sampled for Source B items (or Source A items) to fall below the old/new criterion—and be erroneously judged to be new—but to fall above the source criterion—and be correctly judged to have been presented in context B. This is recollection without remembering. Naturally, both decision criteria may vary in stringency, and according to the model, that affects the proportion of items that will exceed the source criterion but not the old/new criterion. In particular, this proportion will increase as the old/new criterion becomes more stringent, relative to the source criterion (Starns et al., 2008).

That is a testable prediction because the amount of recollection without remembering should increase if the old/new criterion becomes more stringent while the source criterion is held constant. A familiar method of doing that is to provide subjects with information about the percentage of old items versus distractors on item tests only. Here, Starns et al. (2008) informed some subjects that 25% of the test cues would be old (stringent) and informed others that 75% would be old (liberal). As traditionally happens with this manipulation, hit rates were higher in the 75% condition. The key finding, however, was that recollection without remembering was more pronounced when the item criterion was stringent (25%) than when it was liberal (75%), and indeed, the phenomenon was not reliable when it was liberal.

Another method of testing the same prediction is to administer the usual Source A versus Source B recognition tests to measure source memory, but to administer recall tests to measure item memory. The idea is that decision criteria are far more stringent for recall than for recognition, and hence, recall should produce more robust evidence of recollection without remembering. Cook et al. (2006) reported five experiments of that sort, which produced consistent evidence of this phenomenon: Over their experiments, the probability of correct source recognition for unrecalled items averaged .66. Kurilla and Westerman (2010) obtained analogous results with item tests that involved recall and source tests that involved recognition.

Artifactual Explanation

According to this explanation, the assumption that item memory is necessary for source memory is not actually violated when subjects respond accurately to source memory tests for items that they think are new. Here, Kellen, Singmann, and Klauer (2014) and Malejka and Bröder (2016) argued that recollection without remembering may be an artifact of a shared design feature of prior studies—namely, that item tests were always administered before source tests. That is the standard procedure in source-monitoring research (see Johnson, Hashtroudi, & Lindsay, 1993), and actually, it is predicated on the assumption that source memory without item is a null state (Brainerd et al., 2015). However, this procedure creates an interpretive difficulty in studies of recollection without remembering: If item memory is necessary for source memory, an initial item test provides practice at retrieving information that is essential to item memory, and hence, such practice may improve item memory on later source tests.

The denouement is that when item memory fails on an initial item memory test, causing an old item to be remembered as new, it may be more likely to succeed on a subsequent source test, thanks to the beneficial effects of retrieval practice on the earlier item test. Under that hypothesis, the naïve interpretation of recollection without remembering—that accurate source memory occurs in the absence of item memory—is wrong. Rather, inaccurate item memory on earlier item tests has been transformed into accurate item memory on later source tests, by virtue of retrieval practice. In our experiments, we eliminated this interpretive difficulty by counterbalancing the order of source and item tests for individual items.

Overview of the Research

We report two experiments in which we compared the three explanations of recollection without remembering by testing different predictions that they make about the phenomenon. The overall list structure that was used in these experiments is illustrated in Figure 1. In both instances, subjects studied two word lists (A and B), each of which was accompanied by a distinct auditory context (e.g., list A was read in a voice of one gender and list B was read in a voice of the other gender). Each list was composed of words belonging to familiar taxonomic categories. Each list had its own set of categories (e.g., list A might contain multiple exemplars of fabrics and multiple exemplars of kitchen utensils presented in blocks, whereas list B might contain multiple exemplars of animals and multiple exemplars of colors presented in blocks; cf. Figure 1). Within each category, two types of exemplars were presented on the lists, which we will call context-consistent targets (CT) and context-inconsistent (IT) targets. For CT targets (e.g., pan), the exemplar appeared on the same list and in the same voice as six other blocked exemplars of its category (e.g., pan, pot, spatula, tongs, colander, cup, and stove appeared in consecutive positions on list A in a male voice). For IT targets (e.g., fork), the exemplar appeared on a different list and in a different voice than all the other exemplars of its category (e.g., fork appeared on list B in a female voice). Thus, the correct source details for CT targets were the same as those for nearly all of the other exemplars of its category, but for IT targets, the correct source details were different than those for nearly all of the other exemplars of its category.

Figure 1.

An example of List A and List B structure using four categories. IT words are context-inconsistent targets, and CT words are context-consistent targets. During the study phase, IT words were inserted between blocks of CT words. The order of the blocks of CT words and the interspersed IT studied words were randomized for each subject, as was the order of the test probes during the test phase.

After studying the lists, subjects responded to item- and source-recognition tests for all words in counterbalanced order—item tests came first for half of the words and source tests came first for the other half. We saw earlier that the dual-recollection account predicts that there are manipulations that will affect target recollection and context recollection in opposite ways and, hence, drive the accuracy of item and source memory in opposite directions (Brainerd et al., 2014). Context consistency is a case in point. On the one hand, IT targets should enhance target recollection, relative to CT targets, leading to more accurate item memory: IT targets are instances of what are traditionally called list isolates, items that stand out against a background of blocks of exemplars of other categories. List isolates display a classic target recollection phenomenon, the von Restorff effect (e.g., Schmidt & Schmidt, 2017). On the other hand, IT targets should simultaneously impair context recollection, reducing the accuracy of source memory. Remember that for IT targets, their correct source details (voice and list) are the opposite of those for all of the other exemplars of its category. Importantly, note that these opposing effects of the category-consistency manipulation on target and context recollection are memory effects rather than criterion effects.

Another key dual-recollection prediction is that the recollection without remembering effect should be stronger for CT items than for IT items, for reasons that are unconnected to criterion stringency. This is clear from the fact that the lower levels of target recollection for CT items should simultaneously increase accuracy on source tests and decrease it on item tests. Obviously, this prediction does not follow from the multivariate signal detection explanation because decision criteria are normally more stringent on source tests than on item tests (see estimates for counterbalanced item and source tests in Brainerd et al., 2012), which would work against recollection without remembering in that explanation. The prediction does not follow from the artifactual explanation either, as long as the order of the two types of tests is counterbalanced.

Finally, notice that the artifactual explanation makes simple, testable predictions about the test-ordering manipulation. If it is true that recollection without remembering is a by-product of retrieval practice on an initial item-memory test, it should only be observed for items whose test order is item recognition → source recognition. At the least, recollection without remembering should be stronger with the item recognition → source recognition order than with the source recognition → item recognition order. The second result would show that the effect is partly but not entirely a consequence of item-memory retrieval practice.

Experiment 1

Method

Subjects

The subjects were 44 undergraduates (32 female, M_age = 20 years) who participated in the experiment to fulfill course requirements. Subjects were randomly assigned to experimental conditions.

Design and Materials

The recognition test consisted of 56 pairs of tests (1 old/new and 1 source) randomized for each subject. 25% of probes were CT targets, 25% were IT targets, 25% were related distractors (RD; unpresented exemplars of presented categories), and 25% unrelated distractors (UD). Sixteen taxonomic categories composed of eight exemplars each were selected from Overschelde’s category norms (M_frequency = 0.45, where frequency is the average proportion of times a word was given as a category exemplar by subjects) (Overschelde et al., 2004). Half of the categories were randomly assigned to each list, independent of list presentation order. Selected exemplars were three to twelve letters long and were concrete nouns (e.g., apple, fork). To avoid potential item confounds, exemplars were counterbalanced over CT, IT, and RD words. All material in this experiment was presented on a computer screen, centered against a white background and printed in a black Arial font.

Procedure

Prior to the study phase, subjects received general instructions, informing them that they would be tested on their memory for the presented words and the gender of the voice that they were read in. Subjects then studied each word list, with words remaining on the screen for 1.8 sec followed by a cross hair for 0.2s. A pre-recorded audio file pronouncing each word was played using speakers. Each list was randomly read in either a male or a female voice. CT targets for each category appeared in sequence along with the other words in the same category (i.e., exemplar presentation was blocked by category). The IT targets were presented on the list opposite to the other exemplars of its category, and consequently, were read in a different voice. The order of IT targets, categories, and CT targets within each category were randomized for each subject. Following the study phase, subjects completed a self-paced distractor task, which involved providing solutions to 150 arithmetic problems.

Prior to the test phase, subjects were given instructions for answering old/new (item) and source tests. For old/new tests, subjects were instructed to press the YES sticker on the “Z” key if they believed an item was presented on one of the lists or to press the NO sticker on the “M” key if they believed that it had not been presented. Source questions asked whether the probe had been presented in a male or female voice. Subjects were instructed to press either the MALE sticker on the “K” key or the FEMALE sticker on the “A” key. Subjects were also instructed that sometimes, they would be required to make a source judgment for a distractor item, in which case they should provide their best guess.

Results

Item and source memory

For all statistical tests in this paper, we used the .05 level of confidence. Table 1 shows the raw and bias-corrected acceptance probabilities on item recognition tests and source recognition tests for CT, IT, and RD items. For CTs and ITs, these are probabilities of accepting the correct alternative—old on item tests and the correct voice on source tests. For RDs, the acceptance probabilities for source tests are “correct” (i.e., accepting the voice in which most of the other exemplars of the category were presented), but those for item tests are incorrect (i.e., RDs are not old). The results in Table 1 and in the remainder of this article are based on the widely used two-high-threshold (2HT) method of correcting recognition data for response bias, in which UD acceptance rates for item and source tests are subtracted from the corresponding acceptance rates for CT, IT, and RD words (see Bayen, Murnane, & Erdfelder, 1996; Snodgrass and Corwin, 1988). It has been suggested that 2HT may sometimes produce different results than some other familiar correction methods—in particular, signal detection theory statistics such as d′ and A′ (Healy & Kubovy, 1978). In the present case, however, all of the empirical patterns are the same for 2HT, d′, and A′. Hence, the analyses that are reported below and in Experiment 2 relied on the 2HT method.

Table 1.

Mean Acceptance Probabilities on Item Recognition Tests and Source Recognition Tests

Probe type	Mean acceptance rate (bias corrected)	S.D.	One-sample t (test value = 0)	df	p (2-tailed)	95% CI of Difference
Experiment 1: Item Recognition
IT	.81(.66)	.20	21.89	43	< .0001	[.60, .73]
CT	.71(.56)	.20	18.81	43	< .0001	[.50, .62]
RD	.29(.14)	.20	4.88	43	< .0001	[.08, .20]
Experiment 1: Source Recognition
IT	.44(−.03)	.22	.95	43	> .05	[−.10, .04]
CT	.69(.20)	.13	10.52	43	< .0001	[.16, .24]
RD	.29(.01)	.32	.16	43	> .05	[−.10, .11]
Experiment 2: Item Recognition
IT	.81(.63)	.26	29.11	143	< .0001	[.59, .67]
CT	.72(.54)	.27	23.95	143	< .0001	[.50, .59]
RD	.26(.08)	.23	3.98	143	< .0001	[.04, .12]
Experiment 2: Source Recognition
IT	.41(.04)	.27	1.92	143	> .05	[.00, .09]
CT	.70(.31)	.21	17.72	143	< .0001	[.27, .34]
RD	.38(.06)	.28	2.51	143	< .05	[.01, .11]

Note. CT = category-consistent targets, IT = category-inconsistent targets, and RD = related distractors. Item and source acceptance rates for IT, CT, and RD were corrected for bias with the two-high-threshold statistic P_r. Source “correct” acceptance rates for RD were the probabilities of selecting the voice in which most of the exemplars of an RD were presented. t tests for bias-corrected data for IT, CT, and RD evaluated whether observed means were reliably > 0. The 95% confidence intervals represent the range of data that contains the difference between the observed means and the test value. Because difference scores are involved, the left hand value of a confidence interval can be negative.

For bias-corrected data (i.e., excluding UDs), the factorial design of this experiment was 2 (word presentation order: first list vs. second list) X 2 (voice presentation order: first list vs. second list) X 2 (test presentation order: first vs. second) X 3 (item type: CT vs. IT vs. RD) X 2 (test type: item vs. source), with item type and test type manipulated within subjects. A series of preliminary analyses revealed that the between-subject factors of word presentation order, voice presentation order, and test presentation order produced no main effects and did not interact with any other factors in the design. Consequently the main statistical analysis for treatment effects was a 3 (item type: CT vs. IT vs. RD) X 2 (test type: item vs. source) repeated-measures analysis of variance (ANOVA) of the bias-corrected acceptance probabilities (target hits and RD false alarms). RDs were not presented on study lists, of course, and hence, acceptances were false alarms. For purposes of this ANOVA, source “hits” for RDs refer to source tests on which subjects identified the voice in which most of the exemplars of that category had been presented (e.g., identifying the voice of the distractor strawberry as male, when all but one of the fruit exemplars had been presented in a male voice). Importantly, two of the variables that failed to produce either main effects or interactions were test presentation order and test type—whether the item test preceded the source test or the source test preceded the item test. Recall that these variables provide a direct test of the artifactual explanation of recollection without remembering, which predicts that this phenomenon will only be detected when item tests precede source tests. That scenario obviously predicts a Test Presentation Order X Item Type X Test Type interaction, such that source accuracy for IT and CT targets is better when source tests are administered second than when they are administered first. There was no interaction of this sort, which means that the artefactual explanation was not supported.

There was a reliable item type main effect, F(2, 76) = 50.35, MSE = 04, η_p² = .57. There was also a reliable test type main effect, F(1, 38) = 153.21, MSE = .06, η_p² = .80. Finally, there was a reliable Item Type X Test Type interaction, F(2, 76) = 48.46, MSE = .03, η_p² = .56. Post hoc analyses (Tukey HSD) of this interaction revealed that the ordering of mean acceptance probabilities was IT > CT > RD on item tests but was CT > IT = RD on source tests.

Taken together, these results are congruent with the dual-recollection predictions about the category-consistency manipulation. According to those predictions, the von Restorff effect for ITs means that target recollection will be highest for such items. That, in turn, is expected to increase items hits for ITs relative to CTs, whereas the inconsistency between voice and other source details for IT items is expect to lower source hits for ITs relative to CTs. In addition, mean acceptance probabilities for ITs and CTs were higher than the mean acceptance probability for RDs on both item tests because, of course, RDs were not presented.

We also computed one-sample t tests of the bias-corrected item and source data in order to determine whether the absolute levels of acceptance of CTs, ITs, and RDs on item and source tests were reliably > 0. The complete results are reported in the fourth column of Table 1. In connection with those results, the item rates for CTs, ITs, and RDs were all reliably > 0, meaning that there was no doubt that the presentation of both types of targets could be remembered and that there was a semantic false memory effect. In contrast, only the source hit rate for CTs were reliable—so that although item memory was better for IT than for CT items, IT source memory was not even reliable.

Recollection without remembering

Turning to direct statistical measures of recollection without remembering, we analyzed the conditional probability of a correct source judgment given an incorrect item judgment (CP0) and of a correct source judgment given a correct item judgment (CP1). The former is the measure of recollection without remembering, and the latter is the measure of recollection with remembering. Then, we computed one-sample t tests of these conditional probabilities in order to determine if they were reliably > .5 (the guessing probability with two-alternative source tests).

The complete results—the values of CP0 and CP1 for CTs and ITs, as well t values and significance levels—are reported in Table 2. There are two key findings. First, for CTs, both the value of CP0 (.64) and the value of CP1 (.74) were reliably > .5. Thus, for targets that had been presented on the same list and in the same voice as most of the other exemplars of their category, it was not necessary to be able to remember that an item has been presented in order to be able to remember its voice, although source memory was better when CT targets could be remembered as old. Second, for ITs, neither the value of CP0 (.49) nor the value of CP1 (.40) were reliably above chance, and in fact, the value of CP1 was reliably below chance. Obviously, these findings are startling from the perspective of the traditional assumption that item memory is a necessary precondition for accurate source memory. That would lead one to expect that source memory would be better for ITs than for CTs, because item memory is better for ITs, especially for items that were remembered as old on item tests. Instead, the opposite was true, and accurate item memory actually impaired source memory, relative to inaccurate item memory. With data such as these, it is difficult to see how accurate memory for a specific item’s source could be said to depend on remembering that the item is old.

Table 2.

Conditional Probabilities of Correct Source Recognition Given Correct or Incorrect Item Recognition

Probe type	Question order	Conditional rate	S.D.	One-sample t (test value = .5)	p (2-tailed)	95% CI of Difference
	Experiment 1: correct source recognition given incorrect item recognition
IT	S -> I	.47	.33	.33	> .05	[−.19, .14]
	I -> S	.49	.37	.06	> .05	[−.19, .18]
CT	S -> I	.63	.24	2.45	< .02	[.02, .25]
	I -> S	.64	.27	2.47	< .02	[.02, .27]
	Experiment 1: correct source recognition given correct item recognition
IT	S -> I	.37	.18	3.18	< .001	[−.21, −.04]
	I -> S	.42	.15	2.58	< .02	[−.14, −.02]
CT	S -> I	.75	.09	12.95	< .001	[.21, .29]
	I -> S	.73	.08	12.82	< .001	[.20, .27]
	Experiment 2: correct source recognition given incorrect item recognition
IT	S -> I	.45	.38	1.02	> .05	[−.15, .05]
	I -> S	.59	.43	1.46	> .05	[−.04, .22]
CT	S -> I	.66	.33	3.76	< .001	[.08, .25]
	I -> S	.52	.39	.37	> .05	[−.08, .12]
	Experiment 2: correct source recognition given correct item recognition
IT	S -> I	.40	.23	3.78	< .001	[−.16, −.05]
	I -> S	.40	.23	3.72	< .001	[−.15, −.05]
CT	S -> I	.72	.27	7.07	< .001	[.16, .28]
	I -> S	.75	.20	10.75	< .001	[.21, .30]
	Pooled Data: correct source recognition given incorrect item recognition
IT	S -> I	.46	.36	1.06	> .05	[−.13, .04]
	I -> S	.57	.41	1.26	> .05	[−.04, .17]
CT	S -> I	.65	.31	4.44	< .001	[.09, .22]
	I -> S	.55	.37	1.28	> .05	[−.03, .13]
	Pooled Data: correct source recognition given correct item recognition
IT	S -> I	.39	.22	4.74	< .001	[−.15, −.06]
	I -> S	.41	.21	4.40	< .001	[−.14, −.05]
CT	S -> I	.73	.24	9.28	< .001	[.18, .28]
	I -> S	.75	.18	13.14	< .001	[.21, .29]

Note. CT = category-consistent targets, IT = category-inconsistent targets, and RD = related distractors. t tests for IT, CT, and RD conditional probabilities evaluated whether observed means were reliably > 5. The 95% confidence intervals represent the range of data that contains the difference between the observed means and the test value. Because difference scores are involved, the left hand value of a confidence interval can be negative.

Finally, the data of Experiment 1 rule out an alternative source-guessing explanation of our recollection without remembering results that may occur to readers, an explanation that does not run counter to the traditional hypothesis that accurate source memory requires accurate item memory. The targets that subjects studied were exemplars of familiar taxonomic categories, most of which (7 of 8 exemplars) were presented in the same voice on the same list. Suppose that when item memory fails for a presented exemplar, subjects use metacognitive knowledge of the category’s dominant voice (e.g., “animals are female” or “fabrics are male,” as in Figure 1) to guess its source. This would produce the above-chance levels of source accuracy that we observed for CP0 items that are CTs (e.g., cat), but so would recollection without remembering. There is another finding, however, that differentiates the source-guessing hypothesis from recollection without remembering. Notice that under this hypothesis, source guessing for CP0 items that are ITs would produce accuracy levels that are reliably below chance. For instance, if item memory fails for dog and cotton in Figure 1, the guessing hypothesis says that subjects will choose “female” for dog and “male” for cotton, both of which are incorrect. That did not happen, and instead, source accuracy for CP0 items that were IT (.49) was almost exactly the chance probability. Therefore, the source-guessing explanation of recollection without remembering fails.

Experiment 2

The results of Experiment 1 were consistent with the dual-recollection explanation of recollection without remembering and not with the multivariate signal detection or artifactual explanations. Neither of the latter accounts predicts that it is possible to simultaneously impair item memory and enhance source memory, but the category-consistency manipulation did precisely that. In addition, the artifactual explanation predicts that recollection without remembering will be observed when item tests precede source tests but not when source tests precede item tests. That prediction was disconfirmed. However, it sometimes happens that a memory manipulation has much more robust effects when it is varied within- rather than between-subjects, with word frequency being a classic example.

Consequently, we decided to conduct a second experiment in which the order of item and source tests was manipulated within-subjects. Our chief aim was to provide a second, more sensitive, test of the predicted test-order effect. A secondary aim was to establish the replicability of the key findings of Experiment 1 with this modified design.

Method

Subjects

The subjects were 75 undergraduates (59 female, M_age = 19.87 years) who participated in the experiment to fulfill course requirements. Subjects were randomly assigned to experimental conditions.

Design, Materials and Procedure

The materials and experimental procedure were identical to Experiment 1, except that question order was manipulated within subjects, rather than between subjects. Recall that in Experiment1, the recognition test consisted of 56 pairs of tests for each word—an old/new item test and a male/female source test—with half of the subjects receiving the 56 item tests followed by the 56 source tests and the other half receiving the reverse order. In this experiment, subjects received a single sequence of tests that consisted of the 56 pairs being presented in random order, with the item test preceding the source test for 28 pairs (7 CT, 7 IT, 7 RD, and 7 UD) and the source test preceding the item test for the other 28 pairs. The item and source tests for individual words were separated by at least six tests for other words.

Results

Item and source memory

Table 1 shows the same types of item and source memory descriptive results as were previously reported in Experiment 1, for CTs, ITs, and RDs. For CTs and ITs, these are probabilities of accepting the correct alternative—old on item tests and the correct voice on source tests. For RDs, acceptance probabilities for source tests are “correct” (i.e., the voice in which most of the other exemplars of the category were presented), but acceptance probabilities for item tests are incorrect (i.e., RDs are not old).

It will be recalled that there were three ordering factors in our design—namely, whether words were presented on the first or the second list, whether a given voice occurred first or second, and whether a given type of test was administered first or second. As in Experiment 1, a series of preliminary analyses revealed that none of these ordering factors produced a main effect, and that none interacted with any other factor in the design. Therefore, the ordering factors were not included in subsequent analyses. The main analysis for treatment effects was a 3 (item type: CT vs. IT vs. RD) x 2 (test type: item vs. source) repeated-measures ANOVA of the bias-corrected acceptance probabilities. As in Experiment 1, we excluded the other design factors because we found no effects of these variables in preliminary analyses. The 3 X 2 ANOVA produced a reliable item type main effect, F(2, 274) = 178.12, MSE = .05, η_p² = .57, and a reliable test type main effect, F(1, 137) = 175.26, MSE = 09, η_p² = .56. As in Experiment 1, the ANOVA also produce a reliable Item Type X Test Type interaction, F(2, 274) = 126.26, MSE = .05, η_p² = .48. Post hoc analysis of this interaction revealed the same pattern as in Experiment 1—namely, that the order of accuracy was IT > CT > RD for item tests but was CT > IT = RD for source tests. Once again, then, the results for the category-consistency manipulation conformed to the predictions of the dual-recollection model inasmuch as the manipulation drove item and source memory in opposite directions.

We again computed one-sample t tests of bias-corrected probabilities, in order to determine whether the absolute levels of acceptance of CTs, ITs, and RDs on item and source tests were reliably > 0. Those statistics are reported separately for the ITs, CTs, and RDs in Table 1. It can be seen that the overall picture was the same as in Experiment 1. For item tests, the acceptance probabilities for CTs, ITs, and RDs were all reliably > 0, so that the presentation of both types of targets could be remembered and there was a semantic false memory effect. In contrast, only the source hit rate for CTs (.31) was substantially > 0. The IT source hit rate was not reliably > 0, and although we saw that the RD source hit rate did differ significantly from the IT source hit rate and both were close to 0 (.06 vs. .04), the RD value was reliably greater than 0. The latter result must be very cautiously interpreted, however, considering how small the RD value was and the fact that it was not reliably > 0 in Experiment 1.

Recollection without remembering

In order to generate direct measures of recollection without remembering, we again analyzed the conditional probability of a correct source judgment given an incorrect item judgment (CP0) and of a correct source judgment given a correct item judgment (CP1). Also as before, we computed one-sample t tests of these conditional probabilities, which determined if they were reliably > 0.5 (the guessing probability with two-alternative source tests).

The complete results for these tests—the values of CP0 and CP1 for CTs and ITs, as well t values and significance levels—are reported in Table 2. Note that as in Experiment 1, the findings were different for CTs than for ITs. For CTs, both the value of CP0 (.60) and the value of CP1 (.74) were reliably > 0. Once again, for these items it was not necessary to be able to remember that an item was old in order to be able to remember its source, although source memory was more accurate for items that were remembered as old. Second, also as before, there was no evidence of accurate source memory for ITs, as neither the value of CP0 (.52) nor the value of CP1 (.40) was reliably above chance. Like the corresponding results of Experiment 1, these IT findings are counterintuitive from the perspective of the traditional assumption that accurate source memory attaches to accurate item memory. If that assumption were true, source memory should have been better for ITs than for CTs because item memory was better.

Finally, as in Experiment 1, the IT source accuracy data when item memory failed rules out the source-guessing explanation of recollection without remembering for CT items. Remember that this explanation predicts that IT source accuracy will be reliably below chance expectations when item memory fails. However, the observed accuracy level (.52) was slightly above chance.

General Discussion

Considering that demonstrations of recollection without remembering are as yet rather sparse in the literature, an instructive outcome of our experiments is that they provided clear evidence of this phenomenon under theoretically-specified conditions. Overall, the probabilities of recollection without remembering in the conditions in which it was expected to occur were .64 when the order of item and source tests was varied between subjects and .59 when it was varied within subjects. These values are comparable to those in prior reports of this phenomenon.

Beyond this, the main purpose of our experiments was to advance theoretical understanding of recollection without remembering by pitting three accounts of the effect against each other—the dual-recollection explanation, the multivariate signal detection explanation, and the artifactual explanation. The featured manipulations were the consistency of presented items’ voice with the other source details for their taxonomic category and the order in which source and item tests are administered, the latter manipulation separating the artifactual account from the other two explanations and the former separating the dual-recollection account from the other two explanations. Apart from quantitative differences, our experiments produced the same general pattern with respect to the three explanations—namely, that the findings were more congruent with the dual-recollection explanation than with the other two accounts.

Weighing the Explanations

The most fundamental question is whether recollection without remembering is a real memory effect that requires theoretical explanation or is merely an epiphenomenon. The artifactual hypothesis says that it is an epiphenomenon of test order. Explicitly, recollection without remembering is a consequence of the traditional practice of administering item tests before source tests, which may produce higher levels of item memory on later source tests than on earlier item tests. The usual rationale for that practice is that if accurate item memory is a necessary precondition for source memory, it is nonsensical for subjects to make source judgments about items that they regard as new (see Johnson et al., 1993). However, the retrieval practice that is provided by earlier item tests could enhance item memory on later source tests, producing what only appears to be recollection without remembering. If so, failures of item memory on earlier tests will be transformed into successful item memory on later source tests, leading to accurate source memory.

That claim is easily tested because recollection without remembering should only occur for the item → source test order, or at least, it should be more pronounced for that order than for the source → item order. We counterbalanced test order, both between and within subjects, and the results failed to confirm these predictions. The test order manipulation never produced a main effect, and it failed to interact with any of the other design factors. In short, there was no support at all for the notion that recollection without remembering is wholly or partly a by-product of responding to item tests before responding to source tests.

Turning to the other two explanations, in order to compare the dual-recollection and multivariate signal detection accounts, we analyzed the effect of category consistency on item and source memory for presented words. Because the latter account posits that the same retrieval process underlies performance on both item and source tests (sampling values from a single source strength distribution), manipulations that improve item memory should also improve source memory. In contrast, the dual-recollection explanation does not impose that constraint because it assumes that some manipulations can have opposite effects on target and context recollection, which drives accuracy on item and source tests in opposite directions. We saw that category consistency is such a manipulation inasmuch as IT targets should exhibit enhanced target recollection but impaired context recollection, relative to CT targets. A robust pattern of that sort was observed in both experiments.

It should be added that there is a prior literature in which selected manipulations have been found to enhance item memory but impair source accuracy, with articles by Dodson and Shimamura (2000), Jurica and Shimamura (1999), and Lindsay and Johnson (1991) being cases in point. Upon first impression, the results of such studies may seem to support dual-recollection theory and disconfirm the traditional hypothesis that source memory requires accurate item memory. However, those results are subject to two interpretive problems (see Brainerd et al., 2015), both of which arise from the conventional procedure (see Johnson et al., 1993) of administering a target’s item test before its source test and only administering its source test if the response to the item test is “old.” It is not difficult to see that this procedure assumes the validity of the hypothesis that source memory requires accurate item memory. The two interpretative difficulties that it creates with respect to manipulations that doubly dissociate item and source performance are that we do not know if a dissociation will hold (a) when the administration of source tests does not depend on item performance and (b) when source tests precede item tests.

Another finding that bears on the comparative validity of the dual-recollection and multivariate signal detection explanations is that the category-consistency manipulation affected the levels of recollection without remembering that were observed. On the one hand, one of the reasons for studying this manipulation is that the dual-recollection account predicted that recollection without remembering would be more robust for CT items than for IT items. That was the observed pattern, and further, recollection without remembering was reliable for CT items but not for IT items. On the other hand, it is difficult to reconcile those findings with the multivariate signal detection account because this explanation depends on the decision criterion on item tests being more stringent than the decision criterion on source tests. If we compute the usual signal detection criterion statistic, C, for the data of these experiments, the mean values for CT items are more stringent for item tests than for source tests, the mean values for CT items are also more stringent for item tests than for source tests, and the differences in criterion stringency CT vs. IT items are roughly comparable. Thus, the expectations are that (a) recollection without remembering should be observed for both CT and IT items because C was more stringent for item tests in both cases, and (b) the magnitude of the effect should be roughly comparable for CT and IT items because the differences in C for item and source tests were roughly comparable. Neither result was obtained, however.

Representation of Contextual Details

Although the results favored one of the explanations over the other two, a remaining question that will be crucial in explaining recollection without remembering is how the contextual details that support accurate responses to source tests are stored in memory. Consider this question in light of the familiar distinction between verbatim traces that represent item presentations per se—their exact surface form—and gist traces that represent the underlying meaning relations that subjects extract from those presentations. Because contextual details are a form of surface information, a natural hypothesis, one that is congruent with item memory being a necessary precondition for source memory, is that contextual details are stored in the verbatim traces that represent items themselves. However, our results, along with those of Ball et al. (2014), challenge that scenario in two ways. First, we now know that manipulations that should enhance verbatim traces of item presentations can impair performance on source tests. That was the pattern produced by the category-consistency manipulation. Remember that the von Restorff effect for IT items was expected to produce superior verbatim memory for particular targets, relative to the CT items, but source memory was not even reliable for IT items. The fact that subjects could not remember contextual details at all for words that produced superior item memory makes it seem unlikely that such details could be stored in verbatim traces of item presentations.

Second, the other challenge to that idea grows out of the fact that although verbatim traces of item presentations cannot be deposited for distractors, that does not prevent subjects from exhibiting “accurate” source memory for distractors. The subjects in our second experiment and in Ball et al.’s (2014) experiments displayed “accurate” source memory for distractors that were semantically related to list items—same-category exemplars in our case and associates in Ball et al.’s case. Based on these particular findings, a more likely scenario for the representation of contextual details is that they are stored with traces of the semantic information that subjects extract from item presentations. In our experiments, there was evidence that contextual details are stored with taxonomic information about list items. In that connection, recall that source memory was always better for CT items than for IT items. Indeed, source memory was not above chance for IT items, which suggests that contextual details were only stored with representations of taxonomic information.

In addition to taxonomic information, Ball et al.’s (2014) research points to associative information as a mechanism for storing contextual details. In their experiments, subjects displayed reliable source memory for distractors that were associates of list items. However, the mechanism was complex. Source memory was reliable when the association was backward from distractors to list items, but not when it was forward from list items to distractors. Thus, it may be that associative relations are a basis for retrieving contextual details but not for storing them during item presentations. Instead, contextual details might be stored with representations of other semantic properties (e.g., taxonomic relations), with their retrieval then being triggered by backward associations on memory tests.

Clearly, these findings about how the contextual details that are responsible for recollection without remembering are stored are far from definitive. At most, what can be said is that the evidence (a) runs against those contextual details being stored in verbatim traces of item presentations and (b) runs in favor them being stored in traces of items’ semantic content.

Concluding Comments

On the whole, the current situation is that accumulated data favor the view that recollection without remembering is a real memory effect, not an epiphenomenon. It is therefore correspondingly difficult to continue entertaining the assumption that remembering an item’s source details without remembering the item itself is a null state. Although the exact explanation of recollection without remembering remains to be determined, dual-recollection theory at least appears to be consistent with what we currently know about this effect.