Explaining The Association Between Repetition Priming And Source Memory: No Evidence For A Contribution Of Recognition Or Fluency
Mar 19, 2022
Contact: joanna.jia@wecistanche.com / WhatsApp: 008618081934791
Abstract In a conjoint memory task (measuring repetition priming, recognition memory, and source memory), items recognized as previously studied and receiving correct source decisions also tend to show a greater magnitude of the repetition priming effect. These associations have been explained as arising from a single memory system or signal, rather than multiple distinct ones. In the present work, we examine whether the association between priming and source memory can alternatively be explained as being driven by recognition or fluency. We first reproduced the basic priming-source association (Experiment 1). In Experiments 2 and 3, we found that the association persisted even when the task was modified so that overt and covert recognition judgments were precluded. In Experiment 4, the association was again present even though fluency (as measured by identification response time) could not influence the source decision, although the association was notably weaker. These findings suggest that the association between priming and source memory is not attributable to a contribution of recognition or fluency; instead, the findings are consistent with a single-system account in which a common memory signal drives responding.
Keywords Source memory; repetition priming; recognition memory
Memory can be expressed in a variety of ways, such as a change in identification or detection of an item due to previous exposure to the item (long-term repetition priming) or the ability to determine whether an item had been encountered before in a particular context (recognition memory). Prominent theories explain these particular phenomena as being driven by distinct memory systems, signals, or processes. Under some theoretical accounts, priming is driven by an implicit (unconscious or non-declarative) memory system, whereas recognition memory is driven by a functionally and neurally distinct explicit (conscious or declarative) memory system (e.g., Squire, 1994, 2004, 2009; Squire & Dede, 2015; Tulving & Schacter, 1990).
These multiple systems account of memory is pervasive in psychology textbooks as the default model of memory (e.g., Baddeley et al., 2014), and independent memory systems are still used to explain differential memory performance (e.g., Henson et al., 2016). Evidence for a multiple systems theory of memory is based on functional and neural dissociations between tasks (e.g., Craik et al., 1994; Jacoby & Dallas, 1981; Schacter et al., 2007; Squire, 2009; Staresina et al., 2011), though there is evidence challenging these findings and/or inferences (e.g., Addante, 2015; Berry et al., 2014; Buchner & Wippich, 2000; Dunn, 2003; Lukatela et al., 2007; Meier et al., 2009; Mulligan & Osborn, 2009; Ostergaard, 1992; Poldrack, 1996; Thakral et al., 2016).
The counterview to the multiple systems model of memory is that memory expression in different tasks, such as priming and recognition, is based on the same underlying memory signal. Under such an account, higher memory strength for an item should be simultaneously associated with greater priming and higher recognition memory. Berry et al. (2012) tested this account using a conjoint priming and recognition memory paradigm where, for each item at the test, participants were asked to identify a word as it clarified over a mask (to provide a measure of priming) and give a recognition judgment on a scale of certain-new to certain-old. In line with a single-system model, they found that identification for items judged old was faster than that of items judged new; the priming effect, as measured across all studied items, was greater than the priming effect for items not recognized, and identification RTs (response times) tended to decrease as recognition confidence increased. This has since been replicated many times and confirmed in formal modeling (e.g., Berry et al., 2006, 2008a, 2008b, 2010; 2014; 2017; Mazancieux et al., 2019; Ward et al., 2013; see Shanks & Berry, 2012, for a review).
Nicholas Lange and Christopher J Berry
1 Department of Psychology, University of Warwick, Coventry, UK
2 School of Psychology, University of Plymouth, Plymouth, UK
However, under some accounts of recognition memory, recognition memory itself is driven by two processes: recollection and familiarity (e.g., Yonelinas, 2002). While recollection relies on explicit retrieval of memory, familiarity is often argued to be driven by repetition priming (e.g., Jacoby & Dallas, 1981; Mandler, 1980). This means that the association of priming and recognition memory could be driven by this shared, implicit component and leaves the question of whether the same memory signal can drive performance in priming and a memory task that is traditionally seen as reliant on explicit memory.
In Lange et al.’s (2019) study, we, therefore, extended the behavioral and modeling work of Berry et al. (2012) to source memory. In source memory tasks, participants are asked to retrieve the exact context an item was studied in, such as whether it was shown in red or blue font, on the top or the bottom of the screen, or on a beach or woods background. These tasks cannot be solved by relying on familiarity but require the explicit retrieval of memorial information (but see Diana et al., 2008; Taylor & Henson, 2012). In this extended task, at study participants were shown words at the top or the bottom of the screen. Attest, participants first identified an item as it was clarified across a mask, then gave a recognition confidence rating, followed by a source confidence rating. We replicated findings of the association of priming and recognition memory and observed the analogous association of priming and source memory: items with correct source decisions tended to also have faster identification RTs (for similar findings using a recall task as the source memory task, see Mazancieux et al., 2019, Exp 1).
These results are consistent with a single memory signal underlying responding where greater memory strength of an item is more likely associated with greater priming, correct “old” recognition judgments, and correct source judgments. While the core assumption of a single memory signal or multiple independent memory signals is central to the predictions about the association of those memory tasks, auxiliary assumptions about the response mapping describe how responding in one task changes with responding in another. In the standard response mapping, responses are assumed to be made independently of one another. For the association of priming and source memory, for example, this means that the magnitude of the priming effect should monotonically increase from “sure-(incorrect source decision)” to “sure-(correct source decision).”1 However, in all Lange et al.’s (2019) experiments, priming tended to be highest at both endpoints of the rating scale and lowest at the mid-point of the scale. In other words, priming increased with increasing confidence in the source decision, regardless of whether that decision was correct or incorrect.
cistanche for sale
Given source memory ratings followed recognition ratings in our task, we considered whether this unexpected pattern in the association of priming and source memory was due to the recognition ratings that preceded the source confidence ratings, that is, that recognition and source memory responses were not made independently. It is well-established that there is some dependency between source- and recognition-ratings, such that source decisions made with high confidence are more likely when recognition decisions are made with high confidence (e.g., Hautus et al., 2008; Starns et al., 2013) and that this is a consequence of more than just a shared memory signal (Starns & Ksander, 2016). Models of recognition and source memory incorporate this by allowing source decision criteria or response mapping to change with the recognition rating (e.g., Hautus et al., 2008; Klauer & Kellen, 2010; Onyper et al., 2010). When we adapted the response mapping to include the dependency between these responses, the single-system model of our conjoint memory tasks captured the finding that correct source decisions are associated with greater priming than incorrect source decisions overall, and that priming increases with source confidence regardless of whether the source response was correct.
One possibility is that the better prediction of the model with changed response mapping is evidence that the underlying process that gives rise to the specific characteristics of the association between priming and source memory is the decisional dependence of source memory ratings on preceding recognition memory ratings. In this article, we sought to test this empirically. If source memory confidence ratings change with recognition confidence ratings, removing recognition confidence ratings should remove that decisional bias. Then, overall, correct source decisions should still be associated with greater priming than incorrect source decisions (in line with the core assumption of the single-system model), but priming should now gradually increase with increasing confidence in the correct source decision. Experiment 1 is a replication of Experiment 2 by Lange et al. (2019) to re-establish the previously observed pattern of the association of priming and source memory. We then sought to determine whether the association would persist even when overt (Experiment 2) and covert (Experiment 3) recognition judgments were precluded. In Experiment 4, by measuring priming and source decisions in separate, rather than interleaved phases, we tested whether the priming-source association would persist under conditions where other factors, like the fluency of identification, would not influence the source decision.
Experiment 1
Method Participants.
Thirty-six individuals (7 male; M age=24.20, SD=9.52) took part in the experiment for payment of £8. This sample size provided a power of 0.8 to detect a medium-sized effect in a repeated measures design with two levels (i.e., a Cohen’s DZ approximately equal to 0.48) based on calculations for a pilot study. We used the same sample size in each subsequent experiment. Participants in each experiment were recruited using a University of Plymouth participation pool. Ethics were approved by the University of Plymouth ethics board. All participants provided informed consent prior to participating in the experiment.
Materials. The stimulus pool consisted of 384 four-letter low-frequency words, selected from the Medical Research Council psycholinguistic database (Coltheart, 1981). The frequency of occurrence ranged from 1 to 13 per million, and there were no concreteness or imageability constraints. Archaic and colloquial terms were excluded. For each participant, 176 words were randomly assigned to be the old stimuli, another 176 words were selected to be the new stimuli, and a further 32 words were selected to be the stimuli appearing on primacy and recency buffer trials in the study phase.
Procedure. At the beginning of the experiment, participants completed six practice trials of the continuous identification task (CID; Berry et al., 2012; Feustel et al., 1983; Lange et al., 2019; Stark & McClelland, 2000) to familiarise themselves with the task prior to the experimental trials. The CID procedure was the same as that of Lange et al. (2019). On each CID trial, a single word was flashed for longer and longer durations, becoming clearer over time. Participants were instructed to press the Enter key as soon as they were sure that they could identify the word correctly. Accuracy and speed were emphasized in the task instructions. At the start of each trial, a fixation mask “####” was presented in 24-point Courier font for 1,000ms. Next, the word was presented in 20-point Courier font for 16.7ms (one screen refresh at 60Hz). The mask was then presented for 233.3ms, forming a 250ms presentation block. There were thirty 250ms presentation blocks. The stimulus duration increased by 16.7ms on each alternate block, and the mask was always presented for the remainder of the 250ms block. Thus, each CID trial was potentially 7,500ms long but could be terminated prematurely by the participant pressing the Enter key. When the Enter key was pressed, the mask was then re-presented for 16.7ms. Next, a white outlined box was presented that indicated to the participant that he or she must type the word on the keyboard. Key presses were displayed in the box. Participants were told to press Enter after typing the word to advance to the next trial.
Study phase. Participants were told that they would see words presented below or above the center of the screen for a brief duration and that their task was to remember the location of each word for a later test. Participants completed eight study-test blocks, which were identical except that the stimuli in each block were unique. At the start of each study block a “+”-fixation was presented for 500ms in the center of the screen. The words were presented for 2 s each, with half of them presented 0.9 cm below the central fixation point (i.e., subtending a vertical visual angle of approximately 0.69°, from a viewing distance of approximately 75 cm) and the other half 0.9cm above the fixation point. The inter-stimulus interval was 100ms. The assignment of words to the location and the order of presentation was randomized across participants. Participants completed 26 study trials per block, with the first and last two trials in each block designated as primacy and recency buffer trials. The buffer stimuli were not presented in the experiment again.
Test phase. Next, instructions were presented for the first CID-RS (i.e., CID with Recognition and Source judgments) test phase. Participants were told that they would again complete identification trials and that some of the words were from the previous study block and some were novel. They were told that they must decide whether they thought the word was new (i.e., not shown previously) or old (i.e., studied) after each identification, and to indicate whether it was previously shown at the bottom or the top of the screen. They were informed to make that location judgment even for items they indicated were new and to guess if unsure. Participants were told that half of the words would be new and half would be old, that half of the old words were presented at the bottom of the screen and half were presented at the top. There were 44 trials in each test block, composed of 22 old and 22 new items. On each trial, a word was presented in the center of the screen using the same CID procedure as in the practice trials. After participants made their identification, the word was replaced by a recognition probe (“Is the word New or Old?”) and a rating scale (“1=sure new, 2=probably new, 3=guess new, 4=guess old, 5=probably old, 6=sure old”). After participants made their recognition judgment, a source memory probe was presented (“Was the word presented at the bottom or top?”) with a rating scale (“1=sure bottom, 2=probably bottom, 3=guess bottom, 4=guess top, 5=probably top, 6=sure top”). Participants used the number keys 1 through 6 on the main part of a QWERTY keyboard for the recognition judgments and the number keys on the number pad for the source memory judgment. Stickers were added to the number pad with up arrows indicating the “top” response and down arrows indicating the “bottom” response. After making their source memory judgment, a prompt was presented instructing participants to press the Enter key to start the next trial. On completion of the test block, participants were presented with the next study block. On completion of the final test block, the experiment terminated.
Initial screening of identification trials. In this experiment and subsequent ones, a trial was not included in the analysis if a word was misidentified during the identification phase of a trial or identification responses were too fast or too slow. Identification responses were corrected for minor typographical errors (e.g., where a number or a symbol was typed after the correctly typed word). One participant was excluded at this stage because they did not attempt to identify any words in the first study-test block. Overall, the proportion of misidentified trials after correction for typographical errors was low (M=3.05%, SD=2.58), as was the proportion of trials on which participants did not provide a response (M=0.19%, SD=0.78). The proportion of trials on which the identification RT was less than 200ms or greater than three standard deviations above the mean identification RT (within-participant) was also low (M=1.22% of trials, SD=0.49). Following Lange et al. (2019), these four types of trials were not analyzed further. This left a sufficient number of valid trials for all individuals (M=95.54, SD=2.52, Min=88.07%).
Measures. All analyses were conducted in R (R Core Team, 2019). For all relevant statistical comparisons, we excluded participants listwise if they had missing data in any cell of that analysis. Analysis of variance (ANOVAs) was calculated using the aov_car function in the apex package (Singmann et al., 2020), with post hoc contrasts calculated with means (Lenth, 2020). Degrees of freedom were corrected for violation of sphericity where necessary using the Greenhouse-Geisser correction. An alpha level of .05 was used for all statistical analyses and all t-tests were two-tailed. We also conducted equivalent Bayesian analyses, and report Bayes factors (BF) for all reported frequentist tests, using the BayesFactor package (Morey & Rouder, 2018), with the package’s default priors for all tests. We report the following effect sizes: ηP2 for ANOVAs, Cohen’s DZ (DZ; mean difference of two dependent measures, divided by the average standard deviation of the difference of the two measures) for t-tests. Trials were aggregated across study-test blocks for all analyses.
The priming effect was calculated as the mean identification RT for new items minus the mean identification RT for old items. Recognition discrimination was measured with d′ (henceforth referred to as recognition d′), which is calculated as z(p[“old”| old])—z(p[“old”| new]), where p(“old”| old)=(number of hits+0.5)/(number of old items+1) and p(“old”| new)=(number of false alarms+0.5)/(number of new items+1), following Snodgrass and Corwin (1988). The pattern of results for Pr, which is the measure of discriminability in the two-high threshold model and is calculated as p(“old”| old)—p(“old”| new), was the same, so we only report recognition d′ throughout. Recognition response bias was measured with c (henceforth referred to as recognition c), which is calculated as −0.5 * (z(p[“old”| old])+z(p[“old”| new])). Source discrimination was measured with d′ (henceforth referred to as source d′). For this measure, source-top items were arbitrarily designated as targets and source-bottom items as nontargets; thus, source d′=z(p[“top”| top])—z(p[“top”| bottom]), where p(“top”| top)=(number of correct top responses+0.5)/(number of source-top items+1) and p(“top”| bottom)=(number of incorrect top responses+0.5)/(number of source-bottom items+1). The pattern of results for source accuracy—calculated as (number of “top”|top items+number of “bottom”|bottom items)/number of old items—was the same, so only the former is reported. Source bias was measured with c (henceforth referred to as source c) and calculated as −0.5* ( z(p[“top”| top])+z(p[“top”| bottom])).
For the analysis of identification RTs classified according to source confidence ratings, responses were collapsed across source-top and source-bottom items. Source ratings 3, 2, and 1 for source-bottom items and 4, 5, and 6 for source-top items constituted correct source decisions with increasing certainty of response, while sourcing ratings 4, 5, and 6 for source-bottom items and 3, 2, and 1 for source-top items constituted incorrect source decisions. Reliability of measures. Prior research has shown that it is important to consider the relative reliabilities of direct and indirect memory tasks when comparing task performance (Buchner & Wippich, 2000). Accordingly, split-half correlations were used to determine the reliability of the priming, recognition, and source measures in all experiments. To calculate these, we first split the data from each participant into odd- and even-numbered trials and then calculated the priming effect, recognition d′, and source d′ in each half. The split-half correlations were then given as the Pearson correlation between performance in each half across participants. In Experiment 1, these were large and significant, priming, r(33)=.90, p<.001, BF=1.94 × 109 ; recognition d′, r(33)=.90, p<.001, BF=3.55 × 109 ; source d′, r(33)=.81, p<.001, BF=7.70 × 105.
cistanche for sale
Results
Considering first overall levels of memory performance, the priming effect, recognition d′, and source d′ all exceeded chance (0): M priming=247ms, SE=34, t(34)=7.17, p<.001, d=1.22, BF=5.11 × 105; M recognition d′=1.23, SE=0.10, t(34)=12.02, p<.001, d=2.03, BF=8.61 × 1010; M source d′=0.80, SE=0.11, t(34)=7.48, p<.001, d=1.26, BF=1.16 × 106 . Table 1 shows the mean identification RT for new and old items, and also the mean hit rate and false alarm rate for recognition and source decisions. Neither recognition nor source responding was biased overall (recognition c=−0.04, SE=0.04, t(34)=0.98, p=.33, d=0.17, BF=0.28; source c=0.01, SE=0.05, t(34)=0.11, p=.91, d=0.02, BF=0.18).
There was evidence for correlations between these overall measures, though this was only substantial for the association of recognition and source memory (priming and recognition d′, r(34)=.35, p=.041, BF=2.32; priming and source d′, r(34)=.33, p=.056, BF=1.84; recognition d′ and source d′, r(34)=.82, p<.001, BF=1.82 × 106 ). As in Lange et al.’s (2019) study, we expected associations between priming and source memory to be evident when broken down according to the source decision. We consider two aspects of the data: (a) the difference in the magnitude of the priming effect for items with correct and incorrect source decisions, and (b) how the priming effect varies with participants’ confidence in their source decision.
First, the priming effect for items with correct source decisions was significantly greater than for items with incorrect source decisions (M difference = 71ms, SE= 24), t(34) = 3.00, p < .005, d = 0.51, BF= 7.76), see the left-hand side of Figure 1a). This difference was consistent across individuals, being present in 69% of participants.
Second, we examined identification RTs for correct and incorrect source decisions across participants’ confidence. This analysis is limited to studied items, i.e., items that can be associated with a correct and incorrect source decision. Table 2 shows the mean number of items at each level of this analysis. Please see the supplementary material for the analysis of the relationship of Identification RT and source confidence for new items. Identification RTs tended to decrease (i.e., the priming effect was greater) as confidence in the source decision increased, as is shown on the right-hand side of Figure 1a. This trend was confirmed in a 3 (source confidence: guess, probably, sure)×2 (source decision: correct, incorrect) repeated measures ANOVA, which yielded a significant main effect of source confidence, F(1.63, 48.77)=11.62, MSE=70,424, p<.001, ηP2=.28, BF=9.79×102. Four participants could not be included in this ANOVA because they had zero responses for particular cells of the analysis (hence N=31 for this analysis). Post hoc analyses confirmed a significant linear trend, t(43)=4.82, p<.001, with higher-level trends not significant (p>.89). Source decisions given a high confidence rating were associated with a faster identification than source decisions given a low confidence rating, p<.001 (the remaining comparisons, Bonferroni-adjusted, p>.043). There was no main effect of source decision, F(1, 30)=1.14, MSE=32431, p=.29, ηP2=.04, BF=0.22, or interaction, F(2, 60)=1.16, MSE=40521, p=.32, ηP2=.04, BF=0.23.
Briefly, in this experiment, we also replicated the association of priming and recognition memory shown by Berry et al. (2012) and Lange et al. (2019). For old items, identification was faster for items judged old than items judged new, M difference=210ms, SE=51, t(34)=4.15, p<.001, dz=0.70, BF=127, and identification RTs decreased with increasing recognition confidence (p<.001, though p<.015 for quadratic and cubic trends). For new items, there was no clear evidence for an effect of fluency, that is, M difference in identification RT to new items judged old and new=48ms, SE=25, t(34)=1.95, p=.060, dz=0.33, BF=0.98, though overall identification RTs decreased with increasing recognition confidence (p<.001, all higher-level contrasts: p>.050).
Discussion
These results are consistent with those of Lange et al. (2019), showing greater priming for correct than incorrect source decisions, and greater priming with increasing confidence regardless of the source decision. We also replicated the now well-established association of priming and recognition memory in this paradigm (e.g., Berry et al., 2012). Having established the association between priming and source, we now turn to test if recognition confidence ratings are central to the nature of this association. This is the theoretical assumption underlying the adapted response mapping in the single-system model by Lange et al. (2019). In all following experiments, we will not elicit overt recognition ratings from participants. In addition, in Experiments 3 and 4, we will also limit covert recognition judgments, that is, judgments of an item’s oldness in the absence of an instruction to do so, by only showing old items at test.
cistanche for sale
