Individual differences in value-directed remembering

Abstract

Capacity limits in cognition require that valuable information be prioritized for encoding and retrieval. Individual differences in prioritized value-directed encoding may derive from differences in the general ability to encode memories, or from differences in how strategies are altered for different stimuli to modulate maintenance in working memory. We collected multiple cognitive ability measures to test whether variation in episodic memory, working memory capacity, or both predict differences in value-directed remembering among a large sample of participants (n = 205). Confirmatory factor analysis and structural equation modeling was used to assess the contributions of episodic and working memory to value sensitivity in value-directed remembering tasks. Episodic memory ability, but not working memory capacity, was predictive of value-directed remembering. These results suggest that the ability to prioritize memory derives principally from episodic memory ability overall, so that greater capacity also permits greater flexibility.

1. Introduction

The central nervous system is constantly bombarded with an overload of sensory input. Given well-known capacity limits on attention and memory, important information must be selectively attended to, encoded, maintained, and retrieved from memory to support effective goal-directed behavior. In naturalistic contexts, our memory systems may have an even more limited capacity than estimated in laboratory settings. A recent study revealed that when tested for information of real-life events, participants were only minimally above chance at discriminating encountered information from foils (Misra, Marconi, Peterson, & Kreiman, 2018). To compensate for capacity limits of the memory system, important information is prioritized over less important information (Broadbent, 1958; Cowan, 2000). In this study we investigated how an individual’s ability to prioritize information in recognition memory differs with respect to underlying cognitive processes of working memory capacity and long-term memory abilities.

A prominent method used to study the effects of importance on memory is with value-directed remembering (VDR) paradigms. VDR paradigms assign a point (or monetary) value to each stimulus at encoding. Participants are instructed that their goal is to earn as many points (or as much money) as possible by correctly remembering the stimuli. Results from VDR paradigms with free recall typically find that stimuli associated with higher point values are remembered better than stimuli associated with lower point values (Castel, Benjamin, Craik, & Watkins, 2002; Stefanidi, Ellis, & Brewer, 2018; Watkins & Bloom, 1999).

VDR paradigms have also been implemented within recognition memory tasks. It has been hypothesized that two distinct processes contribute to recognition memory discriminations, recollection and familiarity (Mandler, 1980; Wixted, 2007; Yonelinas, 2002). Recollective memories allow conscious retrieval of the item and detailed associative information from the study episode. Familiarity-based memory discrimination is not accompanied by retrieval of associative information, but instead is composed of a more automatic and superficial feeling of prior exposure. Results from VDR recognition memory tasks have shown that value enhances recognition judgments based upon recollection with little impact on judgments based upon familiarity. The specificity of value-related effects on recognition memory has been demonstrated in “remember” judgments in the remember-know paradigm (Elliott & Brewer, 2019; Gruber & Otten, 2010; Hennessee, Castel, & Knowlton, 2017; Wittmann et al., 2005), accurate source memory (Shigemune, Tsukiura, Kambara, & Kawashima, 2014), and the parietal old-new ERP retrieval component in recognition (Elliott, Blais, McClure, & Brewer, 2020).

It is unclear how specific cognitive processes are modulated to allow important information to be encoded and retrieved over less important information. One mechanism by which mnemonic prioritization is accomplished is through interactions between the dopaminergic reward system and episodic memory regions in the medial temporal lobe (MTL). Results from functional magnetic resonance imaging (fMRI) studies have shown increased blood-oxygen-level-dependent (BOLD) signals in the dopaminergic reward system, including the substantia nigra/ventral tegmental area of the midbrain (SN/VTA), nucleus accumbens (NAcc) and the MTL are predictive of better memory for higher-valued stimuli (Adcock, Thangavel, Whitfield-Gabrieli, Knutson, & Gabrieli, 2006; Shigemune et al., 2014; Wittmann et al., 2005). Individual differences in functioning of these systems have been related to greater VDR effects. Adcock et al. (2006) showed that increased activation in the reward system and MTL during encoding of high value information predicted differential memory performance for higher-valued stimuli. Another study found that increased functional connectivity between midbrain and MTL during encoding of higher-valued information was associated with greater VDR effects on memory performance (Wolosin, Zeithamova, & Preston, 2012). It has also been shown that increased functional connectivity between midbrain and MTL during a post-encoding rest period before retrieval predicted VDR effects on memory performance (Gruber, Ritchey, Wang, Doss, & Ranganath, 2016). Together, these studies suggest that one important mechanism for remembering important information may be the midbrain-MTL loop, where dopaminergic signals from the midbrain increase hippocampal plasticity, thereby enhancing storage of important information in long-term memory (Lisman & Grace, 2005; Shohamy & Adcock, 2010). However, the aforementioned neuroimaging studies employed relatively small sample sizes for investigating individual differences (Schönbrodt & Perugini, 2013). To our knowledge, no known studies have explicitly tested the relation between episodic memory abilities and value-directed remembering effects using multiple behavioral assessments in an appropriately large sample of participants.

A separate line of research has shown that important information encourages recruitment of top-down executive control processes. After recognizing that a stimulus is valuable, participants often engage in more elaborative or deep semantic encoding strategies. One relevant study conducted by Cohen and colleagues employed a VDR free recall paradigm and fMRI to investigate the neural underpinnings of value-driven encoding during free recall (Cohen, Rissman, Suthana, Castel, & Knowlton, 2014). During encoding, participants were presented with either a high or low value cue followed by a to-be-remembered word. After the encoding period, participants recalled as many words as possible from the word list. Correlations between value-modulated brain activity and participant’s selectivity index (which estimates a participant’s sensitivity to value; Castel et al., 2002) indicated that the left inferior frontal gyrus and left posterior lateral temporal cortex was predictive of how sensitive participants were to value, in addition to dopaminergic reward processing regions. The authors concluded that a frontotemporal semantic processing network supports elaborative rehearsal and executive processing to produce VDR effects on memory. Our recent work used divided attention manipulations during VDR encoding to argue for the importance of executive resources during encoding in supporting VDR effects (Elliott & Brewer, 2019). These studies suggest that, in addition to the dopaminergic reward system, executive control processes that support strategy selection and working memory may be another mechanism involved in VDR memory effects.

Individual differences in working memory capacity are driven by variability in the number of items that can be held in primary memory, attention control, and controlled retrieval from secondary memory (Unsworth, 2016). Few studies have investigated individual differences in VDR effects and its relation executive control processes (such as working memory capacity) with two notable exceptions (Griffin, Benjamin, Sahakyan, & Stanley, 2019; Robison & Unsworth, 2017). In two large-scale individual differences experiments (n = 122, n = 200) Robison and Unsworth implemented a self-regulated VDR paradigm where participants were presented with a list of words varying in value. As opposed to typical VDR paradigms in which words are presented sequentially, the words were presented simultaneously, allowing participants to choose how to allocate their study time. To further promote strategy usage, the participants were shown a list of only point values and had to click to see the word associated with each point value. After a two minute encoding period, the participants had 2 min to freely recall as many words as possible from the previous list. The authors found that individuals with higher working memory capacity showed greater VDR effects. Furthermore, equating participants on the use of an effective strategy (i.e., allocating less study time to lower valued items) at the beginning of the task attenuated the working memory capacity and VDR correlations. The authors concluded that individuals with higher working memory capacity are more likely to engage selective encoding strategies to improve VDR performance.

Another individual differences examination of working memory capacity and VDR was performed by Griffin and colleagues (Griffin et al., 2019). They conducted four experiments (n = 101, n = 95, n = 100, n = 97, respectively) examining memory prioritization in a value-based cued recall paradigm. The authors implemented a cued recall task instead of free recall in order to mitigate potential individual differences in cue-dependent retrieval processes. The authors found that working memory capacity was only weakly correlated with VDR performance. This finding suggests that working memory capacity may enhance VDR effect only in contexts for which effortful retrieval strategies are encouraged.

Variability in working memory may account for at least some of the variance in VDR abilities. However, individual differences in VDR may also derive from variability in abilities related to long-term memory encoding and retrieval functions. Greater overall abilities in memory formation and recall should allow for greater variance across items, particularly when value prioritizes some items over others. Individual differences in long-term memory can manifest from variability in encoding processes, implementation of effective encoding strategies, cue-dependent retrieval, and metacognitive monitoring (Unsworth, 2019). No prior work has tested whether individual differences in memory abilities predict the size of VDR effects. We reasoned that VDR may be related to both long-term memory ability and working memory capacity through either unique or shared mechanisms. Unique variance in long-term memory abilities predicting VDR may suggest variability in long-term memory encoding processes. Unique variance in working memory capacity predicting VDR may suggest the implementation of effective executive functions. Shared relations between VDR, long-term memory, and working memory may reflect individual differences in strategic encoding and retrieval processes inherent in all three domains.

The relation between working memory and episodic memory abilities is well established (Rose & Craik, 2012; Unsworth & Brewer, 2009; Unsworth & Engle, 2007; Unsworth, Spillers, & Brewer, 2010). The correlation between working memory and episodic memory typically ranges between 0.33 and 0.79 at the latent level (Unsworth, 2019). There are several factors that account for the covariation between working memory and episodic memory abilities, including selection and use of effective encoding and retrieval strategies, search efficiency, and metacognitive monitoring abilities (Unsworth, 2016). It is possible that overlapping variance between these constructs will be related to VDR. Prior individual differences research using the VDR task has not evaluated relations with long-term memory ability and working memory capacity in tandem.

The goal of the current study was to investigate whether memory prioritization ability is best predicted by episodic memory ability, working memory capacity, or some combination of both. Participants completed multiple measures of episodic memory ability (EPI), working memory capacity (WMC), and value-based recognition memory (VDR). A latent variable approach was used to examine the relations among EPI, WMC, and VDR. First, confirmatory factor analysis (CFA) was used to determine the structure of the data. The CFA results suggest that EPI, WMC, and VDR can be considered as three distinct factors. Therefore, structural equation modeling was used to determine what combination of EPI and WMC best explain VDR.

2. Method

2.1. Participants and design

Two-hundred thirty-three participants were recruited from the Arizona State University research participation pool (average age 18.84 years, SD = 1.33, min = 18 years, max = 29 years, 111 males, 117 females, 5 that did not report gender). According to research by Schönbrodt and Perugini (2013), assuming an average correlation of r = 0.21, and a confidence level of 80%, 238 participants are required for stable correlations. We set our minimum number or participants at 200 and collected data until the end of an academic term to get as close to 238 participants as possible. All participants were native English speakers and were compensated with course credit upon completion of the experiment. Participants were tested in group laboratory sessions lasting approximately 2 h in which they completed three working memory tasks (operation span, symmetry span, reading span), three episodic memory tasks (delayed free recall, cued recall, picture source recognition), two value-based recognition memory tasks (“remember” “know” VDR, VDR source memory), and a final post-experimental questionnaire probing strategy use during the VDR tasks. The questionnaire asked participants whether or not they implemented a different encoding strategy for high and low value words. All tasks were administered in the order listed above.

Of the two-hundred thirty-three total participants, twenty-eight were excluded from the final analyses. Six participants were excluded because of computer errors that did not allow them to complete the study. Eight participants were removed from the final data set for scoring below chance on the picture source memory task (25%). Three were removed for having total working memory errors greater than three standard deviations above the mean (20). Eleven participants were excluded after we discovered that they wrote down the to-be remembered words for the memory tasks on their scratch piece of paper used for the distractor task. Only participants with complete data sets for all tasks were included in the final models (n = 205).

We collected multiple measures of each construct and used latent variable analysis because, in general, between-person variability in performance in any cognitive task reflects both domain-general and domain-specific cognitive abilities. For example, an operation span task measures not only active maintenance of memoranda, but also math ability because math problems are used to provide distraction. A reading span task measures not only active maintenance of memoranda, but also verbal ability because sentence verification problems are used to provide distraction. Therefore, we can use the latent variable approach to extract variance that is common among operation and reading span tasks. This tactic provides a more process-pure measurement of an individual’s ability to maintain information in working memory (Conway et al., 2005).

2.2. Materials

2.2.1. Operation span (Ospan)

Participants solved a series of math operations while trying to remember a set of unrelated letters (F, H, J, K, L, N, P, Q, R, S, T, Y). Participants were required to solve a math operation, and after solving the operation they were presented with a letter for 1 s. Immediately after the letter was presented the next operation was presented. Two trials of each list length (4–6) were presented; the order of list-lengths was randomly varied. At recall, letters from the current set were recalled in the correct order by clicking on the appropriate letters (see Unsworth, Heitz, Schrock, & Engle, 2005 for more details). Participants completed and initial three sets of list-length two to practice and become familiar with the procedures. For all of the span measures, items were scored correct only if the correct item was selected in the correct ordinal position. Reliability for this task was computed as the Cronbach’s alpha using the scores from each trial. The score was the total number of items correctly recalled in the correct serial position.

2.2.2. Reading span (Rspan)

Participants were required to read sentences while trying to remember the same set of unrelated letters as in Ospan. For this task, participants read a sentence and determined whether the sentence made sense or not (e.g. “The prosecutor’s dish was lost because it was not based on fact.”). Half of the sentences made sense while the other half did not. Nonsense sentences were made by simply changing one word (e.g. “dish” from “case”) from an otherwise normal sentence. Participants were required to read the sentence and to indicate whether it made sense or not. After participants gave their response they were presented with a letter for 1 s. At recall, letters from the current set were recalled in the correct order by clicking on the appropriate letters. There were two trials of each list-length with list-length ranging from 4 to 6. The same scoring procedure and reliability estimation as Ospan was used.

2.2.3. Symmetry span (SymSpan)

In this task participants were required to recall sequences of red squares within a matrix while performing a symmetry-judgment task. In the symmetry-judgment task participants were shown an 8 × 8 matrix with some squares filled in black. Participants decided whether the design was symmetrical about its vertical axis. The pattern was symmetric half of the time. Immediately after determining whether the pattern was symmetrical, participants were presented with a 4 × 4 matrix with one of the cells filled in red for 650 ms. At recall, participants recalled the sequence of red-square locations in the preceding displays, in the order they appeared, by clicking on the cells of an empty matrix. There were two trials of each list-length with list-length ranging from 3 to 5. The same scoring procedure and reliability estimation as Ospan was used.

2.2.4. Picture source recognition

During the encoding phase, participants were presented with a picture (30 total pictures) in one of four different quadrants on-screen for 3 s. Participants were explicitly instructed to pay attention to both the picture (item) and the quadrant it was located in (source). Immediately following the encoding phase, participants were instructed to indicate whether the picture was new or old and, if old, in what quadrant it had been presented, via keypress. Participants had 5 s to press the appropriate key to enter their responses. The instructions took approximately 10 s. At test, participants were presented with 30 old and 30 new pictures in the center of the screen. Reliability for this task was computed as the correlation between the even and odd trials with correct accuracy. A participant’s score was the proportion of correct quadrant responses.

2.2.5. Cued recall

Participants were given three lists of 10 word pairs each. All words were common nouns, and the word pairs were presented vertically for 2 s each. Participants were told that the cue would always be the word on top and that the target would be on bottom. Immediately after the presentation of the last word, participants saw the cue word and “???” in place of the target word. Participants were instructed to type in the target word from the current list that matched the cue. Cues were randomly mixed so that the corresponding target words were not recalled in the same order as that in which they had been presented. Participants had 5 s to type in the corresponding word. Reliability for this task was computed as Cronbach’s alpha for the recall performance in each of three lists. A participant’s score was the proportion of items recalled correctly.

2.2.6. Delayed free recall

Participants recalled six lists of 10 words each. All of the words were common nouns that were presented for 2 s each. After the list presentation, participants engaged in a 16 s distractor task before recall: Participants saw 8 three-digit numbers appear for 2 s each, and they were required to write the digits in ascending order. After the distractor task, participants typed as many words as they could remember from the current list in any order they wished. Participants had 45 s for recall. Recall occurred after each study block, and participants were instructed to only recall words from the study list immediately preceding the test phase. Reliability for this task was computed as Cronbach’s alpha for recall performance in each of the six lists. A participant’s score was the total number of items recalled correctly.

2.2.7. “Remember” “know” VDR (VDR_RK)

The paradigm implemented is the same paradigm from Elliott and Brewer (2019) Experiment 1. For clarity and consistency across manuscripts, we largely replicate our description here. Stimuli consisted of 320 nouns from the Toronto noun pool (Friendly, Franklin, Hoffman, & Rubin, 1982). Participants completed four study-test blocks. The study phases consisted of 40 words each randomly assigned either a high (7,9) or low (1,3) point value (i.e., 20 high value and 20 low value, 10 of each specific value). The point value was displayed above the word in center of the screen, simultaneously with the appearance of the word. The test phases consisted of 80 words (40 from the most recent list and 40 new nouns, randomly intermixed) presented one at a time without point values. Participants classified these old and new items at test and made judgments on subjective states of recollection (i.e. “remember”) and familiarity (i.e., “know”, Tulving, 1985). Participants were instructed that they would earn the point value previously associated with word during the study phase if they correctly recognized the word at test (regardless of whether the classified the word as “remember” or “know”). It was emphasized to the participants that they should treat the task like a game and try to maximize the number of points they earn. Each study-test phase was considered 1 block and 4 blocks were completed by each participant.

Before the experiment began, participants were briefed on the difference between remembering and knowing with the following instructions (adapted from Herzmann & Curran, 2011):

“Make a remember judgment if you not only remember the word, but also consciously remember the experience of studying the word. For example, perhaps you remember the specific value of the word, something else that happened in the room while you were studying it (like a cough or sneeze), an association that came to mind, or what came just before or after the word in the study phase. To give you a real world example, imagine you are walking across campus and recognize someone, but cannot recall their name or where you have met them. You are certain you have seen this person before, but do not remember anything specifically about them or where you met them. This would be “knowing”. If you recognize this person and remember that it is John whom you met in Biology class, this would be “remembering”.”

The effect of value on memory performance was quantified as the difference between pooled high value (7,9) “remember” responses and pooled low value (1,3) “remember” responses for the latent variable analysis. We chose to use only “remember” responses as this is where the effect was isolated in previous experiments with this task (Elliott et al., 2020; Elliott & Brewer, 2019). This difference score is thought to reflect memory sensitivity and prioritization of memory encoding (i.e., greater performance on high compared with low value items). Reliability for this task was computed as Cronbach’s alpha for the average of the first two lists and the average of the last two lists.

2.2.8. VDR source memory (VDR_Source)

Stimuli consisted of 320 new nouns from the Toronto noun pool (i.e. no nouns overlapped with nouns presented in the VDR_RK task; Friendly et al., 1982). Participants completed four study-test blocks. The study phases consisted of 40 words each randomly assigned either a high (10) or low (1) point value (i.e., 20 high value and 20 low value, 10 of each specific value). Point value was denoted by the font color of the word (i.e. green = ten points, red = one point). During the study phase, words were shown on either the left or right side of the screen. The test phases consisted of 80 words (40 from the most recent list and 40 new nouns, randomly intermixed) presented one at a time. Participants classified these old and new items at test and made judgments of which side of the screen the word appeared for items classified as old. Participants were instructed that they would earn the point value previously associated with the word during the study phase if they correctly recognized the word and location at test. It was emphasized to the participants that they should treat the task like a game and that they should try to maximize the number of points they earned. Each study-test phase was considered 1 block and 4 blocks were completed by each participant.

The effect of value on memory performance was quantified as the difference between pooled high value and low value average conditional source identification measure (ACSIM). ACSIM reflects the proportion of correct source judgments only among the “old” responses to target stimuli and is less prone to item memory influences (Bröder & Meiser, 2007). ACSIM was calculated with the following formula:

$$ACSIM=\frac{1}{2}\left(\frac{f_{LL}}{f_{LL+}{f}_{LR}}+\frac{f_{RR}}{f_{RL+}{f}_{RR}}\right)$$

where f_LL is the frequency of words presented on the left side of the screen that participants identified as being shown on the left, f_LR is the frequency of words presented on the left side of the screen that participants identified as being shown on the opposite side of the screen (i.e., the right side of the screen), f_RR is the frequency source of words presented on the right side of the screen that participants identified as being shown on the right, and f_RL is the frequency of words presented on the right side of the screen that participants identified as being shown on the opposite side of the screen (i.e., the left side of the screen). This measure was calculated separately for high and low value words and a difference score between these measures quantified value-sensitivity in the source monitoring task. Reliability for this task was computed as Cronbach’s alpha for the average of the first two lists and the average of the last two lists.

3. Results

3.1. VDR effects on memory

Memory performance as a function of value in the VDR_RK task was tested with three separate paired samples t-tests on total hit rates, “remember”, and “know” responses. The results replicate Elliott and Brewer (2019). Higher recognition accuracy was found for high value (M = 0.78, SD = 0.12) compared with low value (M = 0.65, SD = 0.15) words (t₍₂₀₄₎ = 13.917, p < 0.001, Cohen’s d = 0.97). This effect was specific to “remember” responses, for which high value words (M = 0.36, SD = 0.24) were more often classified as “remember” than low value (M = 0.20, SD = 0.19) words (t(204) = 12.968, p < 0.001, Cohen’s d = 0.91). Similar to Elliott and Brewer (2019), there were slightly more “know” responses for low value (M = 0.45, SD = 0.19) versus high value (M = 0.42, SD = 0.22) words, although this effect is small (t₍₂₀₄₎ = −2.773, p = 0.006, Cohen’s d = 0.19).

Memory performance as function of value in the VDR_Source task was tested with three separate paired samples t-tests on total hit rates, correct item recognition without correct source (Item Only memory; IO), and average conditional source identification measure (ACSIM, Bröder & Meiser, 2007). The results showed greater total hit rates for high (M = 0.67, SD = 0.17) versus low (M = 0.57, SD = 0.18) valued words (t₍₂₀₄₎ = 9.706, p < 0.001, Cohen’s d = 0.72). The results demonstrate greater source memory for high (M = 0.69, SD = 0.16) compared with low (M = 0.64, SD = 0.18) value words (t₍₂₀₄₎ = 6.633, p < 0.001, Cohen’s d = 0.33). No effect of value was found on item only memory for high (M = 0.35, SD = 0.14) compared with low (M = 0.36, SD = 0.14) valued words (t₍₂₀₄₎ = −1.559, p = 0.126, N.S.). Furthermore, the value effects on memory between VDR_Source and VDR_RK were positively correlated (r (204) = 0.320, p < 0.001) suggesting that similar encoding and retrieval mechanisms are utilized across the two tasks. These studies replicate previous findings that VDR effects are specific to recollection in recognition memory tasks.

Descriptive statistics and correlations for all of the measures are shown in Table 1 and Supplemental Fig. 1. The measures had generally acceptable values of internal consistency and most of the measures were approximately normally distributed with values of skewness and kurtosis under the generally accepted values for latent factor modeling (i.e., skewness < 2 and kurtosis < 4; see Kline, 1998).

Table 1.

Correlations and descriptive statistics for all tasks. Ospan = operation span; Symspan = symmetry span; Rspan = reading span; PicSource = picture source recognition; CR = cued recall; DFR = delayed free recall; VDR_RK = “Remember” “Know” value-directed recognition memory; VDR_Source = value-directed source memory. Gray shaded regions show the correlations between defined factors (i.e. WMC tasks, EPI tasks, and VDR tasks).

Means, standard deviations, and correlations for all measures
Variable	PicSource	CR	DFR	Ospan	Rspan	SymSpan	VDR_RK	VDR_Source
PicSource	–
CR	0.376**	–
DFR	0.269**	0.459**	–
Ospan	0.178**	0.083	0.166*	–
Rspan	0.192**	0.150*	0.192**	0.467**	–
SymSpan	0.193**	0.066	0.232**	0.385**	0.378**	–
VDR_RK	0.205**	0.271**	0.233**	−0.005	0.048	−0.036	–
VDR_Source	0.152*	0.250**	0.105	0.023	−0.014	−0.032	0.320**	–
M	0.740	0.358	0.461	23.090	22.030	15.200	0.078	0.094
SD	0.145	0.202	0.151	5.558	5.846	4.518	0.086	0.141
Skew	−0.831	0.777	0.442	−0.707	−0.993	−0.248	1.213	0.873
Kurtosis	0.113	0.091	0.421	−0.119	0.720	−0.395	1.397	0.778
Reliability	0.852	0.846	0.830	0.596	0.668	0.543	0.777	0.425

Note.

^**Correlation is significant at the 0.01 level (2-tailed).

^*Correlation is significant at the 0.05 level (2-tailed). Not corrected for multiple comparisons.

3.2. Confirmatory factor analysis

We used confirmatory factor analysis (CFA) to test a measurement model and determine the latent structure of the data. Specifically, the CFA Measurement Model tested whether EPI, WMC, and VDR are best conceptualized as distinct, but related constructs when compared with a baseline null model. In the model all measures of EPI loaded onto an EPI factor, all measure of WMC loaded onto a WMC factor, and all measures of VDR loaded onto a VDR factor. The three factors were allowed to be correlated.

Model fits were assessed via the combination of multiple fit statistics. These include chi-square, root mean square error of approximation (RMSEA), standardized root mean square residual (SNMR), the comparative fit index (CFI), and the Tucker-Lewis Index (TLI). The chi-square statistic measures the difference between the observed and reproduced covariance matrices. Therefore, nonsignificant values are desirable. The root mean square error of approximation (RMSEA) is a measure of model misfit due to model misspecification. The standardized root mean square residual (SRMR) represents the average squared deviation between the observed and reproduced covariances. In addition, the comparative fit index (CFI) and the Tucker-Lewis Index (TLI), which compare the fit of the specified model to a baseline null model were calculated. CFI and TLI values > 0.90 and RMSEA and SRMR values < 0.08 are indicative of acceptable fit (Kline, 1998).

The estimates for the CFA model are shown in Fig. 1. The fit of the model was good, χ²(17) = 18.245, p = 0.374, RMSEA = 0.019, SNMR = 0.040, CFI = 0.995, TLI = 0.991. Replicating previous research, EPI and WMC were significantly correlated. EPI was also correlated with VDR, but WMC was not. Thus, it appears that the theoretically driven model with three distinct factors fits the data well, replicates prior research, and suggests that individuals with higher performance in the episodic memory tasks tended to show greater value-directed remembering effects.

Fig. 1.

Confirmatory factor model for working memory capacity (WMC), episodic memory (EPI), and value-directed remembering (VDR). Paths connecting latent variables (circles) to each other represent the correlations between the constructs, the numbers from the latent variables to the manifest variables (squares) represent the standardized loadings of each task onto the latent variable. Ospan = operation span; SymSpan = symmetry span; Rspan = reading span; PicSource = picture source recognition; CR = cued recall; DFR = delayed free recall; VDR_RK = “Remember” “Know” value-directed recognition memory; VDR_Source = value-directed source memory. All loadings and paths are significant at the p < 0.001 level with the exception of WMC to VDR (p = 0.913, N.S.).

3.3. Structural equation model

Considering the CFA model suggested that EPI, WMC, and VDR could all be considered distinct but correlated factors, we used structural equation modeling to examine the relative contributions of EPI and WMC to VDR abilities. If theories suggesting midbrain-hippocampal contributions or episodic memory encoding abilities are important for VDR, we expect EPI alone to predict VDR. If dlPFC-driven executive resources or retrieval processes are necessary, we expect WMC to predict VDR. To the degree that both of these systems may contribute to individual differences in VDR, neither EPI nor WMC would have independent prediction of VDR effects. Thus, we tested a model were both EPI and WMC were allowed to predict VDR while accounting for their shared relation. The estimates from the SEM model are shown in Fig. 2. The fit of the model was good, χ² (17) = 18.245, p = .374, RMSEA = 0.019, SNMR = 0.040, CFI = 0.995, TLI = 0.991. Only the regression path from EPI to VDR was statistically significant (0.86, Fig. 2). These results suggest that, although there is a relation between EPI and WMC, EPI uniquely predicts VDR.

Fig. 2.

Structural equation model for working memory capacity (WMC) and episodic memory (EPI) predicting value-directed remembering (VDR). Single-headed arrows connecting latent variables (circles) to each other represent standardized path coefficients indicating the unique contribution of the latent variable. The double headed arrow connecting WMC and EPI represents the correlation between the two factors. Solid paths are significant at the p < 0.01 level, whereas dashed paths are not significant.

3.4. Strategy use

It is important to note that the post-experimental questionnaire was not intended to address specific hypotheses, and thus, analyses about strategy use were exploratory in nature. However, it is crucial to test if the observed relations among the factors holds when self-reported strategy use is taken into account, as WMC has been related to strategy use and selection (Bailey, Dunlosky, & Kane, 2008; McNamara & Scott, 2001; Turley-Ames & Whitfield, 2003). We found that differential strategy usage for high and low value words led to greater memory selectivity. An independent samples t-test was conducted on VDR_RK (the difference score of high and low value “remember” responses) and VDR_Source (the difference score of high and low value ACSIM scores) for the two strategy groups (same strategy for high and low value words vs. different strategy for high and low value words). For both the VDR_RK task and the VDR_Source task, Levene’s test for equality of variances was significant, and appropriate corrections to the degrees of freedom were applied. There was a significant difference between the same strategy group (M = 0.106, SD = 0.127) and the different strategy group (M = 0.196, SD = 0.191) on the VDR_RK task, t_(197.044) = −4.038, p < 0.001, Cohen’s d = 0.55. Likewise, there was a significant difference between the same strategy group (M = 0.074, SD = 0.108) and the different strategy group (M = 0.114, SD = 0.160) on the VDR_Source task, t_(200.603) = −2.160, p < 0.05, Cohen’s d = 0.29.

Overall, participants who reported using a different strategy for high and low value words increased their memory selectivity. However, both strategy groups still showed an effect of value on both tasks. A one sample t-test revealed that the same strategy group in the VDR_RK task (M = 0.11, SD = 0.13) was statistically > 0 (t₍₈₆₎ = 7.764, p < 0.001, Cohen’s d = 0.84). Likewise, a one sample t-test revealed that the same strategy group in the VDR_Source task (M = 0.07, SD = 0.11) was statistically > 0 (t₍₉₀₎ = 6.462, p < 0.001, Cohen’s d = 0.68, see Supplementary Fig. 2). The effect of strategy use on the WMC, EPI, and VDR relationship was also investigated. Neither WMC nor EPI were correlated with self-reported strategy use (Supplementary Table 1).

We were concerned that individual differences in strategy usage may serve as a third variable explanation for the lack of a relation between WMC and VDR and the positive relation for EPI and VDR. We therefore conducted partial correlation analyses controlling for strategy use. The null relation between WMC and VDR_RK (r = −0.016, p = 0.821) and VDR_Source (r = −0.013, p = 0.853) held while controlling for individual differences in strategy usage. The positive relationship between EPI and VDR_RK (r = 0.287, p < 0.001) and VDR_Source (r = 0.214, p = 0.003) held while controlling for individual differences in strategy usage. These results provide further support for our conclusion that strategy use had a positive but independent influence on VDR performance compared with EPI. Episodic memory abilities were the best predictor of differences in the ability to prioritize memory based on value in this study.

4. Discussion

The current study examined whether individual differences in value-directed remembering effects (VDR) are predicted by episodic memory abilities (EPI), working memory capacity (WMC), or both. In two VDR recognition tasks (“remember”-“know” and source memory) participants used value to guide encoding. The results replicated previous research showing that the value-based gain in memory performance is selective to recollective memories (i.e. “remember” judgments, accurate source hits, and retrieval ERPs; Elliott & Brewer, 2019; Gruber & Otten, 2010; Hennessee et al., 2017; Wittmann et al., 2005; Shigemune et al., 2014; Elliott et al., 2020). Additionally, the value-based effects on memory between the two tasks were positively correlated, suggesting that similar encoding and retrieval mechanisms are implemented across the two tasks. It is very clear that participants remembered high value words better than low value words in the current tasks, but what is not clear is if this is from high value words being remembered better, low value words being remembered worse, or both. Additionally, although the effect of value is specific to recollection when interpreted in a dual-process framework, these results may be due to overall strength at encoding as opposed to distinct processes (Heathcote, Raymond, & Dunn, 2006). Future research should investigate how strength theory models can incorporate results from VDR paradigms.

Confirmatory factor analysis suggested that episodic memory, working memory capacity, and VDR effects represented three distinct constructs. Episodic memory abilities were correlated with working memory capacity and VDR, whereas working memory capacity was not correlated with VDR. Structural equation modeling revealed that unique variance in episodic memory abilities was predictive of VDR effects, with no relationship observed between working memory capacity and VDR effects.

The current study extends previous research using individual differences methodology which has shown a positive relation between episodic memory abilities and working memory capacity (Unsworth & Brewer, 2009). It is hypothesized that this relation is at least partially explained by individual differences in the ability to implement an effective encoding strategy (Bailey et al., 2008; Robison & Unsworth, 2017; Turley-Ames & Whitfield, 2003; Unsworth, 2016; Unsworth & Spillers, 2010). Our results revealed that episodic memory abilities were predictive of VDR effects and working memory capacity was not. When strategy use was considered (i.e., whether or not individuals used different strategies for high versus low value information) the relation between EPI, WMC, and VDR did not change.

The unique variance in the relation between episodic memory abilities and VDR effects suggests effective encoding strategies may not underlie the observed effects. Instead, we interpret the current results to be consistent with presumed individual differences in the dopaminergic reward system and hippocampal memory processes (midbrain – MTL loop). This brain network is theorized to control the encoding of information into long-term memory (Lisman & Grace, 2005; Lisman, Grace, & Duzel, 2011; Rutishauser, 2019). A recent study using single unit recordings from the midbrain in humans during a recognition memory task found that increased activation of cells in the midbrain was predictive of recognition memory performance, providing further evidence that the midbrain — MTL loop underlies episodic memory encoding in humans (Kamiński et al., 2018). Furthermore, results from human neuroimaging studies using value-based memory paradigms have found that BOLD activity in the midbrain and hippocampus is predictive of successful memory for more valuable information (Adcock et al., 2006; Gruber et al., 2016; Shigemune et al., 2014; Wittmann et al., 2005; Wolosin et al., 2012).

Although we did not observe a relation between working memory capacity and VDR effects in the current study, previous findings have provided evidence that working memory capacity (Robison & Unsworth, 2017) and executive resources (Cohen et al., 2014; Cohen, Rissman, Suthana, Castel, & Knowlton, 2016) are related to VDR effects on memory. However, these studies implemented VDR paradigms with traditional free recall tasks (Cohen et al., 2014; Cohen et al., 2016) and modified free recall tasks (Robison & Unsworth, 2017). The current study used VDR recognition memory tasks opposed to free recall. Different task demands of recall versus recognition may underlie these seemingly contradictory findings. A recent study Middlebrooks, Murayama, and Castel (2017) suggests that participants expecting a VDR free recall task engage more strategic encoding compared to participants expecting a VDR recognition task. The current results are consistent with this hypothesis.

Previous findings from Griffin et al. (2019) also reported a relation between working memory capacity and recall performance (albeit a weak correlation). In contrast to Robison and Unsworth (2017), the authors implemented a cued recall instead of free recall paradigm. The purpose for this was to better isolate effects of value on encoding rather than encoding and retrieval. Recognition memory further constrains retrieval, as participants are forced to make an old/new item discrimination instead of recalling the item. Thus, the relation between working memory capacity and VDR effects could reflect retrieval strategies. Future studies could investigate these VDR free recall, cued recall, and recognition tasks in tandem to further elucidate individual differences in memory prioritization.

Although differential task demands in VDR free recall and VDR recognition tasks may be one explanation for the null relation between working memory capacity and VDR effects, we recently argued that executive resources are necessary for VDR effects on recognition memory (Elliott & Brewer, 2019). The results from Elliott & Brewer can be interpreted in a number of ways. One interpretation is that executive resources are necessary for selection and maintenance of items in working memory. It has previously been shown that applying a cognitive load during encoding removes the gain in recall performance observed for high-WM participants (Engle, Cantor, & Carullo, 1992). However, the results from the current study are inconsistent with that hypothesis. If working memory resources were responsible for the effects observed in Elliott and Brewer (2019) than using individual differences methodology as a “crucible for theory testing” (Underwood, 1975) should have found a correlation between WMC and VDR, which we did not.

An alternative interpretation of the results from Elliott and Brewer (2019) is that dividing attention at encoding disrupted the use of deeper, more elaborative and semantic processes. Engaging deeper, elaborative encoding processes is shown to have selective enhancements to recollection (Yonelinas, 2002) which is also where value effects are localized in many VDR studies. Another theory is that although dividing attention during encoding may disrupt executive resources related to deeper encoding strategies, it also disrupts early attention processes which can be modulated by value (Anderson, 2013; Ariel & Castel, 2014). A recent study in our laboratory set out to test these hypotheses using a VDR recognition memory task with electroencephalography (EEG). The results suggested that earlier attentional processes, and not elaborative rehearsal processes may underlie VDR effect on recognition memory performance (Elliott et al., 2020).

One way to interpret these seemingly disparate results is a model in which the prefrontal cortex interacts with the midbrain to initiate reward — influencing behavior on memory circuits. The frontal cortex and the midbrain have been shown to be functionally connected (Yoon, Minzenberg, Raouf, D’Esposito, & Carter, 2013; Tomasi & Volkow, 2014) and fMRI activation in the frontal cortex has been shown to predict successful memory encoding (Blumenfeld & Ranganath, 2007). Single unit and ECOG recordings in humans have shown that coherence between midbrain activity and theta oscillations in the frontal cortex predict memory performance (Kamiński et al., 2018). Furthermore, there is strong evidence that the prefrontal cortex may mediate phasic bursts in the midbrain (Gariano & Groves, 1988; Grace, Floresco, Goto, & Lodge, 2007; Jo, Lee, & Mizumori, 2013; Patton, Bizup, & Grace, 2013; Parker, Beutler, & Palmiter, 2011; Svensson & Tung, 1989; Takahashi et al., 2011). Previous studies using fMRI in humans have suggested that the prefrontal cortex may drive the midbrain to initiate goal-directed (Murty, Ballard, & Adcock, 2016) and reward-driven behavior (Ballard et al., 2011). These studies suggest that the prefrontal cortex, MTL, and midbrain may all be potentially significant and synergistic in reward-motivated memory encoding.

Although our results discuss the effects of reward (and presumably, the dopamine system) on immediate memory tests, it is important to note that some studies using rewards (Murayama & Kuhbandner, 2011), fMRI (Gruber et al., 2016; Murty, Tompary, Adcock, & Davachi, 2017; Patil, Murty, Dunsmoor, Phelps, & Davachi, 2017; Spaniol, Schain, & Bowen, 2014), and pharmacological manipulations (Murphy, Henry, & Weingartner, 1972; Knecht et al., 2004) traditionally have found effects of reward and dopamine on memory only after a consolidation period. Other studies using rewards (Castel et al., 2002; Stefanidi et al., 2018; Hennessee et al., 2017; Elliott & Brewer, 2019; Elliott et al., 2020), single unit recordings from the midbrain (Kamiński et al., 2018), and fMRI imaging (Cohen et al., 2014; Shigemune et al., 2014; Wolosin et al., 2012) have found effects on immediate memory tests. One possible explanation for these differences is that differential functioning of the midbrain (specifically tonic versus phasic firing) may affect memory on different timescale. Recent research has tried to disentangle the neural generators driving tonic and phasic firing of the midbrain in humans (Murty et al., 2016). The authors discovered that the prefrontal cortex predicted event-evoked phasic firing in the midbrain, whereas the hippocampus predicted tonic shifts in baseline firing. Phasic midbrain signals may be related to overall goal relevance and tonic signals to stimulus novelty (which are confounded in most VDR experiments). Perhaps the effect of reward (and presumably, dopamine) on memory may be multifaceted and confounded by other factors that influence the dopamine system like novelty and overall motivation. To our knowledge, no study has tried to directly address these observed differences. Individual differences methodology may be one approach which can illuminate the apparent differences of reward on immediate and delayed memory tests.

The dopamine system is sub-divided into parallel mesolimbic, mesocortical, and mesotemporal pathways (Haber & Knutson, 2010). These heterogenous functional circuits support different aspects of behavior, including memory formation (associated with midbrain — hippocampal circuitry; Lisman & Grace, 2005), reward learning, cognitive control, and working memory (associated with cortico-striatal circuitry; Montague, King-Casas, & Cohen, 2006; Schultz, Dayan, & Montague, 1997; McClure, Berns, & Montague, 2003; Braver & Cohen, 2000; Cools, Gibbs, Miyakawa, Jagust, & D’Esposito, 2008). Individual anatomical differences may exist across any of these pathways and account for individual differences in specific and non-overlapping tasks (e.g. episodic memory and working memory tasks). Our results suggest that variation in cognitive abilities related to episodic memory is most informative of a person’s ability to prioritize memory. It is exciting to think that a corresponding relation exists with the integrity of the mesotemporal dopamine pathway.

An important (and rarely discussed) limitation of individual differences research is the possibility that the order in which tasks are administered influences performance on the individual tasks and correlations among tasks. On the one hand, researcher could fix the task order and not introduce variance due to randomization. On the other hand, the researcher could randomize task order to ensure that they are not measuring individual differences in learning, fatigue or other contributors to order effects. In the current study, as in many other studies, we chose to use a fixed order, but we certainly acknowledge the limitation, as we would have to acknowledge the alternative limitation if we randomized the order.

Capacity limits of the central nervous system demand that important information be prioritized and encoded over less important information. The results of the current study suggest that episodic memory ability is predictive of a person’s ability to prioritize and selective encode important information, whereas working memory capacity is not. This finding expands our knowledge of VDR effects on memory, and how individual differences in other cognitive processes relate to VDR effects. These results also demonstrate the utility of employing a variety of research methodologies to test cognitive and behavioral phenomenon. Converging operations approach can provide seemingly disparate, but necessary, results that can only bolster further theories about value-directed remembering and cognitive psychology in general (Cronbach, 1957; Garner, Hake, & Eriksen, 1956).