1. Neuroscience
Download icon

Developmental change in prefrontal cortex recruitment supports the emergence of value-guided memory

  1. Kate Nussenbaum
  2. Catherine A Hartley  Is a corresponding author
  1. New York University, United States
Research Article
  • Cited 0
  • Views 290
  • Annotations
Cite this article as: eLife 2021;10:e69796 doi: 10.7554/eLife.69796

Abstract

Prioritizing memory for valuable information can promote adaptive behavior across the lifespan, but it is unclear how the neurocognitive mechanisms that enable the selective acquisition of useful knowledge develop. Here, using a novel task coupled with functional magnetic resonance imaging, we examined how children, adolescents, and adults (N = 90) learn from experience what information is likely to be rewarding, and modulate encoding and retrieval processes accordingly. We found that the ability to use learned value signals to selectively enhance memory for useful information strengthened throughout childhood and into adolescence. Encoding and retrieval of high- vs. low-value information was associated with increased activation in striatal and prefrontal regions implicated in value processing and cognitive control. Age-related increases in value-based lateral prefrontal cortex modulation mediated the relation between age and memory selectivity. Our findings demonstrate that developmental increases in the strategic engagement of the prefrontal cortex support the emergence of adaptive memory.

Introduction

Memories of past experiences guide our behavior, promoting adaptive action selection throughout our lives (Biderman et al., 2020). But not all experiences are equally useful to remember — the information we encounter varies in its utility in helping us gain future reward. By adulthood, individuals demonstrate the ability to prioritize memory for information that is likely to be most rewarding in the future (Adcock et al., 2006; Cohen et al., 2019b; Cohen et al., 2014; Hennessee et al., 2019; Shigemune et al., 2014; Shohamy and Adcock, 2010; Wittmann et al., 2005). Children, however, demonstrate weaker memory selectivity, often remembering relatively inconsequential information at the expense of higher value items or associations (Castel et al., 2011; Hanten et al., 2007; Nussenbaum et al., 2020). Behavioral studies have found that the use of value to guide encoding and retrieval processes emerges and strengthens gradually throughout childhood and adolescence, promoting more efficient acquisition of useful knowledge with increasing age (Castel et al., 2011; Hanten et al., 2007; Nussenbaum et al., 2020). It is unclear, however, how changes in brain activity support this observed emergence of motivated memory. Although a large literature has examined developmental change in the neural mechanisms that support memory from early childhood to young adulthood (Ghetti and Fandakova, 2020; Ofen, 2012; Shing et al., 2010), no prior studies have investigated how the developing brain prioritizes memories based on their relative utility.

Prioritizing valuable information in memory requires both determining the value of information and strategically modulating encoding accordingly. The vast majority of adult studies have focused only on the strategic use of value to guide memory — in most studies of motivated memory, determining the value of information is trivial for participants because experimenters label to-be-remembered information with explicit value cues (e.g. dollar signs, stars, point amounts) (Adcock et al., 2006; Castel et al., 2011; Cohen et al., 2014; Murty et al., 2017). However, in real-world contexts, individuals must derive the value of information from the statistics of their environments. In a recent behavioral study (Nussenbaum et al., 2020), we demonstrated that young adults could use naturalistic value signals to prioritize memory for useful information. Specifically, we manipulated information value via item frequency. Across many environments, the frequency of encountering something in the past predicts the frequency of encountering it in the future. In this way, frequency can signal information utility (Anderson and Milson, 1989; Anderson and Schooler, 1991; Liu et al., 2021; Pachur et al., 2014; Rich and Gureckis, 2018; Stevens et al., 2016) — if individuals are likely to encounter something often in the future, encoding information about it is likely to be valuable. For example, remembering the supermarket aisle of an ingredient with which one often cooks is likely to facilitate greater reward gain than remembering the aisle of an ingredient one almost never uses. In our prior study, we translated this feature of real-world environments to a laboratory task in which individuals could learn the potential reward value of associative information by first learning the relative frequency of items in their environments. We found that individuals exploited these naturalistic value signals and demonstrated better memory for information associated with high- relative to low-frequency items. Critically, this pattern of results varied with age; the strategic prioritization of high-value information in memory increased from age 7 to age 25 (Nussenbaum et al., 2020). It is unclear, however, if developmental improvements in memory prioritization stemmed from differences in learning the relative value of information based on environmental statistics or using learned value signals to strategically prioritize memory. Each of these processes likely engages separable neural systems.

Deriving value from the structure of the environment first requires the learning of statistical regularities. In the case of learning the frequency with which one might need to use information, individuals must differentiate novel occurrences (e.g. cooking with a rare ingredient) from oft-repeated experiences (e.g. cooking with a common food). Neurally, medial temporal lobe regions may support sensitivity to item repetitions. The parahippocampal cortex in particular demonstrates reduced responsivity to repeated relative to novel presentations of items (i.e. ‘repetition suppression’) (Gonsalves et al., 2005; Kirchhoff et al., 2000; Köhler et al., 2005; O'Kane et al., 2005; Turk-Browne et al., 2006). Although some accounts of repetition suppression suggest that attenuated responses simply indicate neural ‘fatigue,’ the phenomenon has also been shown to be sensitive to the statistical context of the environment, suggesting that suppression may reflect stimulus expectation and index learning of environmental regularities (Auksztulewicz and Friston, 2016). Repetition suppression has also been shown to relate to implicit memory for repeated items (Ward et al., 2013). Paralleling their robust implicit learning abilities (Amso and Davidow, 2012; Finn et al., 2016; Meulemans et al., 1998), children and adolescents also demonstrate neural repetition suppression effects (Nordt et al., 2016; Scherf et al., 2011; Turi et al., 2015), although repetition suppression — and the ability to learn the statistical structure of the environment — may increase throughout childhood (Scherf et al., 2011). When individuals need to remember information associated with previously encountered stimuli (e.g. the grocery store aisle where an ingredient is located), frequency knowledge may be instantiated as value signals, engaging regions along the mesolimbic dopamine pathway that have been implicated in reward anticipation and the encoding of stimulus and action values. These areas include the ventral tegmental area (VTA) and the ventral and dorsal striatum (Adcock et al., 2006; Liljeholm and O'Doherty, 2012; Shigemune et al., 2014).

Using these learned value signals to guide memory likely requires cognitive control (Castel et al., 2007; Cohen et al., 2014). Value responses in the striatum may signal the need for increased engagement of the dorsolateral prefrontal cortex (dlPFC) (Botvinick and Braver, 2015), which supports the implementation of strategic control. Enhanced recruitment of control processes promotes the use of deeper and more elaborative encoding strategies (Cohen et al., 2019b; Cohen et al., 2014; Miotto et al., 2006; Uncapher and Wagner, 2009) as well as the selection and maintenance of effective retrieval and post-retrieval monitoring strategies (Libby and Lipe, 1992; Scimeca and Badre, 2012), which may contribute to better memory for high-value information. The use of value to proactively upregulate cognitive control responses improves throughout development, though the specific trajectory of improvement may relate to the control demands of a given task (Davidow et al., 2018). Selectively enhancing the use of encoding and retrieval strategies requires not only tight coordination between subcortical regions involved in value processing and prefrontal areas implicated in control (Murty and Adcock, 2014), but also an available repertoire of memory strategies to implement. Even in the absence of value cues, children and adolescents demonstrate reduced use of strategic control (Bjorkland et al., 2009) and reduced lateral prefrontal engagement during encoding (Ghetti et al., 2010; Ghetti and Fandakova, 2020; Shing et al., 2016; Tang et al., 2018), suggesting that the availability of mnemonic control strategies may increase with age.

Taken together, prior work suggests that adaptive memory requires the recruitment and coordination of multiple neural systems, including mechanisms for learning environmental structure, representing value, and engaging strategic control, all of which may undergo marked changes from childhood to adulthood. Here, we examined how the development of these neurocognitive processes supports the emergence and strengthening of value-guided memory from childhood to young adulthood. To pinpoint loci of developmental differences in adaptive memory prioritization, we combined our novel motivated memory experiment (Nussenbaum et al., 2020) with functional neuroimaging. During the task, participants first learned the frequency of items in their environments, and then learned information associated with each item. Importantly, we structured our task such that the frequency with which participants first experienced each item indicated the frequency with which they would be asked to report the information associated with it, and therefore, the number of points they could earn by remembering the association. Immediately following encoding, we administered a memory test in which participants had to select each item’s correct associate. Because frequency of exposure to an item may facilitate subsequent associative memory even when it does not signal the value of information (Popov and Reder, 2020; Reder et al., 2016), in our prior behavioral study (Nussenbaum et al., 2020), we examined the effects of item frequency on subsequent associative memory in two contexts: one in which item frequency signaled information value and one in which it did not. Critically, we found that with our experimental design, frequency only facilitated memory when it signaled the value of remembering information — increased item exposure did not in and of itself enhance subsequent associative memory. Thus, in the present fMRI study, we focused only on the condition in which item frequency did indicate the potential reward that could be earned for remembering associations.

We examined neural activation during the learning of item frequency, and when participants were asked to encode and retrieve information associated with high- vs. low-frequency items. We hypothesized that while participants across our entire age range would demonstrate sensitivity to the frequency of items in their environments, with increasing age, participants would show improvements in transforming this experiential learning into value signals and modulating the engagement of strategic control processes during encoding. Neurally, we expected that at encoding and retrieval, the recognition of information value would be reflected in increased striatal activation in response to associations involving high- vs. low-frequency items, while the engagement of strategic control would be reflected in increased activation in lateral prefrontal cortex. Further, we hypothesized that increased recruitment of the striatum and prefrontal cortex during encoding and retrieval of high- vs. low-value information would underpin the strengthening of adaptive memory prioritization from childhood to early adulthood.

Results

Approach

Participants ages 8–25 years (N = 90; 30 children ages 8–12 years; 30 adolescents ages 13–17 years; 30 adults ages 18–25 years) completed two blocks of three tasks (Figure 1) while undergoing functional magnetic resonance imaging. In the first, frequency-learning task, participants viewed a continuous stream of 24 unique postcards, one at a time. Twelve of the postcards only appeared once, while 12 repeated five times. Participants indicated whether each postcard they viewed was old or new. In the second, associative encoding task, participants viewed the type of stamp that went on each type of postcard. Participants were instructed that in the subsequent task, they would have to stamp all of their postcards, earning one point for each postcard stamped correctly. Critically, in the associative encoding task, regardless of the number of each type of postcard that they had (i.e. 1 or 5), participants saw each type of postcard with its corresponding stamp only once. Thus, participants were informed that the prior frequency of each postcard indicated the value of encoding its associated stamp, but they had equal exposure to the to-be-encoded associations across frequency conditions. In the retrieval task, participants had to indicate the stamp that went with each unique postcard from one of four options, earning one point for each postcard stamped correctly. After stamping each unique postcard once, participants were asked to report its original frequency on a scale from 1 to 7. Finally, participants stamped all remaining postcards, such that they completed 48 additional memory test trials (i.e. they stamped each of the postcards in the 5-frequency condition four more times.) These trials were not included in any analyses, but their inclusion ensured that correctly encoding the stamps that belonged on the high-frequency postcards would be more valuable for participants despite each retrieval trial being worth one point. After completing the set of tasks, participants were told that they were going to play a second set of similar games. The second set of tasks was identical to the first, except that the stimuli were changed from postcards and stamps to landscape pictures and picture frames. The order of the stimulus sets was counterbalanced across participants, and data were combined across blocks for analyses.

Task structure.

Participants first learned the frequencies of each item (A) by viewing them in a continuous stream. They then were shown the information associated with each item (B). During retrieval, participants had to report the information associated with each item (C) as well as the item’s original frequency (D).

Across our behavioral analyses, we treated age as a continuous variable. To test for nonlinear effects of age, we first compared the fit of models with a linear age term and with both a linear and quadratic age term (Braams et al., 2015; Somerville et al., 2013). We dropped the quadratic age term when it did not significantly improve model fit. Because this study was cross-sectional, one concern was that the children, adolescents, and adults that we recruited may have come from different populations. Indeed, we observed a significant relation between age and age-normed Wechsler Abbreviated Scale of Intelligence (WASI; Wechsler, 2011) scores in our sample (β = −0.60, SE = 0.26, p = 0.0238), suggesting the children had slightly higher estimated IQs for their age relative to adults. To account for these age-related differences in reasoning ability, we included age-normed WASI scores as an interacting fixed effect in all analyses. Our aim in including WASI scores as a control variable was to partially account for confounding, population-level differences across our age groups, enabling us to more clearly examine the relation between age itself and our neurocognitive processes of interest.

Experiential learning of environmental statistics improved with age

During frequency learning, participants across our age range responded to new and repeated items with a high degree of accuracy (new items: mean = 0.90, SD = 0.30; repeated items: mean = 0.92, SD = 0.27; Appendix 2—figure 1A). Older participants demonstrated higher accuracy in correctly identifying both new and repeated items (generalized mixed-effects model results: new items: χ2(1) = 25.52, p < 0.001, repeated items: χ2(1) = 33.43, p < 0.001 Appendix 3—table 1 and 2). Participants were also more accurate in identifying items as ‘repeated’ as the number of times they saw each item increased, χ2 = 138.03, p < 0.001, indicating learning throughout the task. This effect varied as a function of age — younger participants demonstrated a larger effect of the number of item repetitions on response accuracy, as indicated by a significant interaction, χ2(1) = 17.41, p < 0.001.

Response times to both new and old items decreased with age (Appendix 2—figure 1B, Appendix 3—table 3 and 4; linear mixed-effects model results: new items: F(1, 85.99) = 32.51, p < 0.001; old items: F(1, 87.55) = 21.82, p < 0.001), such that reaction times decreased steeply throughout childhood before leveling off into late adolescence and early adulthood. Finally, response times for old items also decreased as the number of item repetitions increased, F(1, 69.94) = 282.21, p < 0.001.

Participants’ ability to distinguish old from new items was associated with a wide network of neural regions, some of which demonstrated greater activation in response to the last vs. first appearance of each item, and others of which demonstrated suppressed activation across repetitions. Specifically, whole-brain contrasts revealed greater recruitment of regions of the lateral occipital cortex, the frontal pole, precuneus, angular gyrus, and caudate (among other regions, see Figure 2A and Appendix 4—table 1), on the last vs. first appearance of each item. We observed widespread repetition suppression effects, reflected in decreases in neural responsivity in the lateral occipital cortex and temporal occipital cortex on the last vs. first appearance of each item. In line with our hypothesis, we also observed a robust decrease in activation in the parahippocampal cortex (Figure 2B, Appendix 4—table 2).

Neural activation during frequency learning.

(A) During frequency learning, participants demonstrated increased recruitment of regions in the frontal cortex, angular gyrus, and striatum on the last vs. first appearance of high-frequency items. (B) They demonstrated decreased activation in the lateral occipital cortex, temporal occipital cortex, and parahippocampal cortex. (C) Within a parahippocampal ROI (shown in green), the decrease in responses to each stimulus on its last vs. first appearance was greater in older participants.

We next examined whether repetition suppression in the parahippocampal cortex changed with age. We defined a parahippocampal region of interest (ROI) by drawing a 5 mm sphere around the peak voxel from the group-level first > last appearance contrast (x = 30, y = −39, z = −15), and mirrored it to encompass both right and left parahippocampal cortex (Figure 2C). For each participant, we modeled the neural response to each appearance of each high-frequency item. We then examined how neural activation changed as a function of repetition number and age. To account for nonlinear effects of repetition number, we included linear and quadratic repetition number terms. In line with our whole-brain analysis, we observed a main effect of repetition number, F(1, 5015.9) = 30.64, p<0.001, indicating that neural activation within the parahippocampal ROI decreased across repetitions (Appendix 3—table 5). Further, we observed a main effect of quadratic repetition number, F(1, 9881.0) = 7.47, p = 0.006, indicating that the reduction in neural activity was greatest across earlier repetitions (Figure 3A). Importantly, the influence of repetition number on neural activation varied with both linear age, F(1, 7267.5) = 7.2, p = 0.007, and quadratic age, F(1, 7260.8) = 6.9, p = 0.009. Finally, we also observed interactions between quadratic repetition number and both linear and quadratic age (ps < 0.026). These age-related differences suggest that repetition suppression was greatest in adulthood, with the steepest increases occurring from late adolescence to early adulthood (Figure 3).

Repetition suppression during frequency learning.

(A) Neural activation within a bilateral parahippocampal cortex ROI decreased across stimulus repetitions both linearly, F(1, 5015.9) = 30.64, p < 0.001, and quadratically, F(1, 9881.0) = 7.47, p = 0.006. Repetition suppression increased with linear age, F(1, 7267.5) = 7.2, p = 0.007, and quadratic age F(1, 7260.8) = 6.9, p = 0.009. The horizontal black lines indicate median neural activation values. The lower and upper edges of the boxes indicate the first and third quartiles of the grouped data, and the vertical lines extend to the smallest value no further than 1.5 times the interquartile range. Grey dots indicate data points outside those values. (B) The decrease in neural activation in the bilateral PHC ROI from the first to fifth repetition of each item also increased with both linear age, F(1, 78.32) = 3.97, p = 0.05, and quadratic age, F(1, 77.55) = 4.8, p = 0.031. The line on the scatter plot represents the best-fitting regression line from the model including both linear and quadratic age terms. The shaded region represents 95% confidence intervals.

For each participant for each item, we also computed a ‘repetition suppression index’ by taking the difference in mean beta values within our ROI on each item’s first and last appearance (Ward et al., 2013). These indices demonstrated a similar pattern of age-related variance — we found that the reduction of neural activity from the first to last appearance of the items varied positively with linear age, F(1, 78.32) = 3.97, p = 0.05, and negatively with quadratic age, F(1, 77.55) = 4.8, p = 0.031 (Figure 3B, Appendix 3—table 6). Taken together, our behavioral and neural results suggest that sensitivity to the repetition of items in the environment was prevalent from childhood to adulthood but increased with age.

Age-related differences in explicit knowledge of environmental structure

Could participants transform their sensitivity to environmental statistics into explicit reports of item frequency? To address this question, we computed participants’ frequency report error magnitudes by taking the absolute value of the difference between the item’s true frequency (i.e. 1 or 5) and each participant’s explicit report of its frequency (i.e. 1–7). We then examined how these report error magnitudes varied as a function of age and frequency condition (Appendix 3—table 7). We observed a main effect of age (F(1, 94.30) = 17.57, p < 0.001) such that error magnitudes decreased with increasing age (Children: Mean = 1.48, SD = 1.34; Adolescents: Mean = 1.10, SD = 1.12; Adults: Mean = 1.13, SD = 1.05). Error magnitudes were not related to frequency condition (p = 0.993), indicating that participants were not systematically better at representing the ‘true’ frequencies of items that appeared once or items that appeared five times.

To examine relations between online frequency learning and explicit knowledge, we tested whether repetition suppression indices for each item related to frequency reports (Appendix 3—table 8). We hypothesized that participants would report the items that elicited the greatest repetition suppression as most frequent. However, in line with other studies suggesting dissociations between repetition suppression and explicit memory (Ward et al., 2013), we did not observe any relation between repetition suppression indices and frequency reports, F(1, 1360.74) = 0.01, p = 0.903. Thus, while we observed parallel developmental improvements in online frequency learning and subsequent explicit reports, they may be driven by separable processes.

Age-related differences in value-guided memory

Participants’ frequency-learning performance and their explicit frequency reports indicate that older participants were better both at tracking repetitions of items within their environments and at explicitly representing item frequencies. Were participants able to use these representations of the structure of their environment to prioritize memory for high-value information? To address this question, we examined how frequency condition and age influenced memory accuracy (Appendix 3—table 9). Memory accuracy varied as a function of both linear (χ2(1) = 8.68, p = 0.003) and quadratic age (χ2(1) = 4.24, p = 0.039), such that older participants demonstrated higher memory accuracy, with the steepest improvements in memory accuracy occurring from childhood into early adolescence (Figure 4). In line with our hypothesis, we observed a main effect of frequency condition on memory, χ2(1) = 19.73, p < 0.001, indicating that individuals used naturalistic value signals to prioritize memory for high-value information. Critically, this effect interacted with both linear age (χ2(1) = 10.74, p = 0.001) and quadratic age (χ2(1) = 9.27, p = 0.002), such that the influence of frequency condition on memory increased to the greatest extent throughout childhood and early adolescence.

Memory accuracy by age and frequency condition.

Participants demonstrated prioritization of memory for high-value information, as indicated by higher memory accuracy for associations involving items in the five- relative to the one-frequency condition (χ2(1) = 19.73, p < 0.001). The effects of item frequency on associative memory increased throughout childhood and into adolescence (linear age x frequency condition: χ2(1) = 10.74, p = 0.001; quadratic age x frequency condition: χ2(1) = 9.27, p = 0.002). The thin grey lines connect each dots representing each participant's memory accuracy for items in the one- and five-frequency condition. The thicker colored lines represent the best-fitting regression lines from models including linear and quadratic age terms. The shaded regions represent 95% confidence intervals.

To determine whether the interaction between quadratic age and frequency condition on memory accuracy reflected an adolescent peak in value-guided memory prioritization, we re-ran our memory accuracy model without including any age terms and extracted each participant’s random slope across frequency conditions. We then submitted these random slopes to the ‘two-lines’ test (Simonsohn, 2018), which fits two regression lines with oppositely signed slopes to the data, algorithmically determining where the sign flip should occur. The results of this analysis revealed that the influence of frequency condition on memory significantly increased from age 8 to age 15.86 (b = 0.03, z = 2.71, p = 0.0068; Appendix 2—figure 2), but only marginally decreased from age 15.86 to age 25 (b = −0.02, z = 1.91, p = 0.0576). Thus, the interaction between frequency condition and quadratic age on memory performance suggests that the biggest age differences in value-guided memory occurred through childhood and early adolescence, with older adolescents and adults performing similarly.

Because we observed age-related differences in participants’ online learning of item frequencies and in their explicit frequency reports, we further examined whether these age differences in initial learning could account for the age differences we observed in associative memory. To do so, we ran an additional model in which we included each participant’s mean frequency learning accuracy, mean frequency learning accuracy on the last repetition of each item, and explicit frequency report error magnitude as covariates (Appendix 3—table 11). Here, explicit frequency report error magnitude predicted overall memory performance, χ2(1) = 13.05, p < 0.001, and we did not observe main effects of age or quadratic age on memory performance (ps > 0.20). However, we continued to observe a main effect of frequency condition, χ2(1) = 19.65 p < 0.001, as well as significant interactions between frequency condition and both linear age χ2(1) = 10.59, p = 0.001, and quadratic age χ2(1) = 9.15, p = 0.002. Thus, while age differences in initial learning related to overall memory performance, they did not account for age differences in the use of environmental regularities to strategically prioritize memory for valuable information.

Neural mechanisms of value-guided encoding

We next examined how neural activation during encoding supported the use of learned value to guide memory. Specifically, we examined whether participants demonstrated different patterns of neural activation during encoding of information associated with high- vs. low-frequency items. In line with our hypothesis, a whole-brain contrast revealed increased engagement of the left lateral PFC and bilateral caudate (1765 voxels at x = −51, y = 42, z = 9; 232 voxels at x = 18, y = 12, z = 6; and 54 voxels at x = 18, y = 18, and z = 12; Figure 5A, Appendix 4—table 4) during encoding of the pairs involving high-frequency items relative to pairs involving low-frequency items. To examine how this pattern of activation related to behavior, we computed a ‘memory difference score’ for each participant by subtracting their memory accuracy for associations involving low-frequency items from their accuracy for associations involving high-frequency items. We then included these memory difference scores as a covariate in our group-level GLM examining neural activation during encoding of pairs involving high- vs. low-frequency items. Participants who demonstrated the greatest difference in memory accuracy for pairs involving high-frequency vs. low-frequency items also demonstrated greater value-based modulation of left lateral PFC activation (232 voxels at x = −48, y = 21, z = 27; Figure 5B, Appendix 4—table 5).

Neural activation during encoding.

(A) During encoding of associations involving high- vs. low-frequency items, participants demonstrated greater engagement of the lateral PFC and caudate. (B) Participants who demonstrated the greatest value-based modulation of memory also demonstrated the greatest modulation of left prefrontal cortical activation during encoding of high- vs. low-value associations. (C) During encoding of both high- and low-value pairs, older participants demonstrated greater recruitment of the PFC relative to younger participants.

Because participants demonstrated effects of value on memory, neural signatures of encoding high- vs. low-value information may reflect successful vs. unsuccessful encoding. To de-confound the effects of value vs. subsequent memory accuracy on neural activation at encoding, we re-ran our high- vs. low-value contrast but restricted our analysis to associations that were subsequently retrieved correctly. We observed similar neural effects — increased recruitment of the left lateral prefrontal cortex and left caudate during encoding of high- vs. low- value pairs (1042 voxels at x = −42, y = 15, z = 30; 124 voxels at x = −12, y = −3, z = 9). Further, neural signatures of successful vs. unsuccessful encoding differed from those of high- vs. low-value encoding. Though we observed similar activation in left lateral PFC (1273 voxels at x = −48, y = 9, z = 27; Appendix 4—table 6), here, we did not observe differential recruitment of the caudate. In addition, consistent with previous observations of subsequent memory effects (Davachi, 2006), successful encoding was associated with increased activation in the right hippocampus (21 voxels at x = 24, y = −6, z = −21).

Age-related differences in neural activation during encoding

Next, we examined how neural activation during encoding vs. baseline (fixation) varied with age. Across trial types, during encoding, we observed widespread age-related increases in neural activation (Appendix 4—table 3), in regions including the left lateral PFC (120 voxels at x = −54, y = 12, z = 33) and the right lateral PFC (24 voxels at x = 48, y = 12, z = 30; Figure 5C).

To address our main question of interest — how age-related differences in differential neural activation during the encoding of high- vs. low-value information may support the development of adaptive memory — we conducted two ROI analyses. Given our a priori hypotheses about the role of the prefrontal cortex and striatum in value-guided encoding, and their exhibiting differential activation in the high- vs. low-value encoding group-level contrast, we examined neural activation within a prefrontal cortex and striatal ROI. Despite the absence of significant differential activation in the hippocampus and parahippocampal cortex, we also used the same ROI approach to test for age differences in activation in these a priori regions of interest but we did not observe any relations between age and hippocampal activation (see Appendix 2: Supplementary Results for details). The specific prefrontal and striatal ROIs were determined by taking the peak prefrontal voxel (x = −51, y = 42, z = 9) and the peak striatal voxel (x = −18, y = 12, z = 6) from the group-level high- vs. low-value associative encoding contrast and drawing 5 mm spheres around them. We then examined how the mean parameter estimate across voxels within each ROI for the high- vs. low-value encoding contrast related to both linear and quadratic age. Caudate activation did not vary significantly as a function of age (β = 0.16, SE = 0.11, p = 0.126) (Appendix 3—table 12), indicating that participants across our age range demonstrated similarly increased recruitment of the caudate while encoding high- vs. low-value associations. PFC activation, however, demonstrated a different pattern, varying as a function of both linear (β = 1.97, SE = 0.74, p = 0.01) and quadratic age (β = −1.73, SE = 0.73, p = 0.021), such that the difference in PFC engagement during encoding of high- vs. low-value associations increased to the greatest extent throughout childhood and early adolescence (Appendix 2—figure 3, Appendix 3—table 13).

The pattern of age-related differences that we observed in the PFC recruitment mirrored the age-related differences we observed in value-based memory. Given these parallel age effects across brain and behavior, we next asked whether age differences in PFC recruitment could account for our observed age differences in adaptive memory prioritization. First, we confirmed that in line with our whole-brain analysis, PFC modulation predicted memory difference scores, even when controlling for age (β = 0.34, SE = 0.10, p = 0.001). Next, we confirmed that these difference scores did in fact vary with age (β = 0.22, SE = 0.10, p = 0.041), with older participants demonstrating a larger difference in memory accuracy for high- vs. low- value associations. Critically, however, when controlling for PFC activation, age no longer related to memory difference scores (β = 0.15, SE = 0.10, p = 0.14). A formal mediation analysis revealed that PFC activation fully mediated the relation between linear age and memory difference scores (standardized indirect effect: .07, 95% confidence interval: [.01, .15], p = 0.017; standardized direct effect: .15, 95% confidence interval: [−0.03, .33], p = 0.108; Figure 6). This relation was directionally specific; age did not mediate the relation between PFC activation and memory difference scores (standardized indirect effect: .03, 95% confidence interval: [−0.007, .09], p = 0.13; standardized direct effect: .34, 95% confidence interval: [.14, .54], p < 0.001.) Further, when we included quadratic age, WASI scores, online frequency learning accuracy, online frequency learning accuracy on the final repetition of each item, and mean explicit frequency report error magnitudes as control variables in the mediation analysis, PFC activation continued to mediate the relation between linear age and memory difference scores (standardized indirect effect: .56, 95% confidence interval: [0.06, 1.35], p = 0.023; standardized direct effect: 1.75, 95% confidence interval: [0.12, 3.38], p = 0.034).

PFC activation mediates the relation between age and value-guided memory.

The increased engagement of left lateral PFC (ROI depicted in red) during encoding of high- vs. low-value information mediated the relation between age and memory difference scores (standardized indirect effect: .07, 95% confidence interval: [0.01, 0.15], p = 0.017; standardized direct effect: .15, 95% confidence interval: [−0.03, 0.33], p = 0.108). Path a shows the regression coefficient of the relation between age and PFC modulation. Path b shows the regression coefficient of the relation between PFC activation and memory difference scores, while controlling for age. Paths c and c’ show the regression coefficient of the relation between age and memory difference scores without and while controlling for PFC activation, respectively. † denotes p<0.06, * denotes p<0.05, ** denotes p<0.01.

Age- and value-based modulation of neural activation during retrieval

We next examined how the neural mechanisms of memory retrieval related to both age and learned value signals. As during encoding, a whole-brain contrast comparing retrieval trials to baseline revealed age-related differences in bilateral PFC recruitment (Figure 7A; 42 voxels at x = 51, y = 3, z = 21; 36 voxels at x = −60, y = 6, z = 21) as well as regions of occipital cortex (see Appendix 4—table 7) across trials during retrieval. We further tested whether participants demonstrated value-based modulation of neural activation at retrieval (Appendix 4—table 8). During retrieval of associations involving high- vs. low-frequency items, we continued to observe increased engagement of the left lateral PFC (Figure 7B; 116 voxels at x = −48, y = 21, z = 24) and the bilateral caudate (128 voxels at x = −12, y = −6, z = 15 and 77 voxels at x = 15, y = −3, z = 21). This activation was not related to age or memory difference scores.

Neural activation during retrieval.

(A) During retrieval, older participants demonstrated greater recruitment of the inferior frontal cortex relative to younger participants. (B) During retrieval of associations involving high- vs. low-frequency items, participants demonstrated greater engagement of the left lateral PFC and bilateral caudate.

Two distinct value representations influence memory

The overlap between the engagement of the neural systems we observed during encoding of high- vs. low-value information and those observed in prior studies of motivated memory that have used explicit value cues (Cohen et al., 2019b; Cohen et al., 2014) suggests that participants did indeed use learned regularities as value signals to guide memory. To what extent was memory supported by explicit representations of item frequency versus neural sensitivity to item repetitions during frequency learning? To examine the influence of these two types of value representations across age, we ran additional mixed-effects models. First, we examined how participants’ explicit representations of item frequency related to memory (Appendix 3—table 14). Participants demonstrated better associative memory for pairs involving items they reported were more frequent, χ2(1) = 31.20, p < 0.001 (Figure 8A). This effect was modulated by age (χ2(1) = 10.37, p = 0.001) and quadratic age (χ2(1) = 9.50, p = 0.002), indicating that participants’ beliefs about item frequency influenced memory to the greatest degree in adolescence and early adulthood. Further, replicating our previous behavioral findings (Nussenbaum et al., 2020), we found that the linear model including explicit frequency reports (BIC = 5438.37) fit the data better than the linear model including the true frequency condition (BIC = 5449.19, χ2 = 10.83, p < 0.001), indicating that participants’ representations of item frequency influenced memory to a greater extent than the true item frequencies.

Memory accuracy by reported frequency.

(A) Participants demonstrated increased associative memory accuracy for items that they reported as being more frequent (χ2(1) = 31.20, p < 0.001). This effect strengthened with increasing age (frequency report x linear age: χ2(1) = 10.37, p = 0.001; frequency report x quadratic age: χ2(1) = 9.50, p = 0.002). (B) Participants also demonstrated better memory for associations involving high-frequency items to which they demonstrated the greatest repetition suppression during frequency learning (χ2(1) = 11.21, p < 0.001). In both panels, the shading of the bars represents the number of trials included in each bin.

We further examined whether our neural measure of online frequency learning related to associative memory. Specifically, we asked whether greater sensitivity to item repetitions — as indexed by greater repetition suppression within the parahippocampal cortex — promoted better encoding of associative information (Appendix 3—table 15). Because we only had repetition suppression indices for items that appeared five times, our analysis was restricted to associations involving high-frequency items. We found that repetition suppression during frequency learning did indeed predict subsequent associative memory, χ2(1) = 11.21, p < 0.001 (Figure 8B). This effect did not interact with age, χ2(1) = 0.79, p = 0.374. Further, when we included both repetition suppression indices and explicit frequency reports in our model, both predictors continued to explain significant variance in memory accuracy (Frequency reports: χ2(1) = 21.16, p < 0.001, Repetition suppression: χ2(1) = 10.25, p = 0.001), suggesting that learned value signals that guide memory may be derived from multiple, distinct representations of prior experience.

Given the relations, we observed between memory and both repetition suppression and frequency reports, we examined whether they related to neural activation in both our caudate and PFC ROI during encoding. To do so, we computed each participant’s average repetition suppression index, and their ‘frequency distance’ — or the average difference in their explicit reports for items in the high- and low-frequency conditions. We expected that participants with greater average repetition suppression indices and greater frequency distances represented the high- and low-frequency items as more distinct from one another and therefore would show greater differences in neural activation at encoding across frequency conditions. In line with our prior analyses, both metrics varied with age (though repetition suppression only marginally (linear age: p = 0.067; quadratic age: p = 0.042); Appendix 3—tables 17 and 20), suggesting that older participants demonstrated better learning of the structure of the environment. We ran linear regressions examining the relations between each metric, age, and their interaction on neural activation in both the caudate and PFC. We observed no significant effects or interactions of average repetition suppression indices on neural activation (ps > 0.15; Appendix 3—tables 18 and 19). We did, however, observe a significant effect of frequency distance on PFC activation (β = 0.42, SE = 0.12, p = 0.0012), such that participants who believed that average frequencies of the high- and low-frequency items were further apart also demonstrated greater PFC activation during encoding of pairs with high- vs. low-frequency items (Appendix 3—tables 21 and 22). Here, we did not observe a significant effect of age on PFC activation (β = −0.03, SE = 0.13, p = 0.82), suggesting that age-related variance in PFC activation may be related to age differences in explicit frequency beliefs. Importantly, however, even when we accounted for both PFC activation and frequency distances, we continued to observe an effect of age on memory difference scores (β = 0.56, SE = 0.20, p = 0.006) (Appendix 3—table 23), which, together with our prior analyses (Appendix 3—table 16), suggest that developmental differences in value-guided memory are not driven solely by age differences in beliefs about the structure of the environment but also depend on the use of those beliefs to guide encoding.

Discussion

Prioritizing memory for useful information is essential throughout individuals’ lifetimes, but no prior work has investigated the development of the neural mechanisms that support value-guided memory prioritization from childhood to adulthood. The goal of the present study was to characterize how developmental differences in the neurocognitive processes that support both the learning and use of information value support improvements in adaptive memory from childhood to adulthood. In line with studies of motivated memory in adults (Cohen et al., 2014), we found that during encoding and retrieval, high- relative to low-value stimuli elicited increased activation in regions sensitive to value and motivational salience, including the caudate (Delgado et al., 2004), and those associated with strategic control processes, including the lateral PFC (Badre and Wagner, 2007; Cole and Schneider, 2007; Power and Petersen, 2013). Further, replicating previous work, we found that value-guided memory selectivity improved across childhood and adolescence (Castel et al., 2011; Hanten et al., 2007; Nussenbaum et al., 2020). Critically, here we demonstrate that increased engagement of the lateral PFC during encoding mediated the relation between age and memory selectivity. This relation was specific to the lateral PFC — although the caudate similarly demonstrated increased activation during encoding of high- vs. low-value information, value-based modulation of caudate activation did not vary as a function of age and did not relate to memory selectivity.

Two different signatures of value-learning predicted subsequent associative memory: Individuals demonstrated better memory for associations involving items that elicited stronger repetition suppression as well as for items that they reported as being more frequent. Moreover, the relation of these learning signals to memory performance varied with age. While all participants demonstrated a similar relation between repetition suppression and subsequent associative memory, the association between explicit frequency reports and memory was greater in older participants. These divergent developmental trajectories suggest that the influence of learned value on memory arises through distinct cognitive processes. One possibility is that while explicit beliefs about information value triggered the engagement of strategic control, stimulus familiarity (as indexed by repetition suppression [Gonsalves et al., 2005]) may have facilitated encoding of novel associations, even in the absence of controlled strategy use. Indeed, prior work suggests that stronger memory traces for constituent components enhance associative memory (Chalmers and Humphreys, 2003; Popov and Reder, 2020; Reder et al., 2016). However, in our previous behavioral work (Nussenbaum et al., 2020), we found that removing the relation between item frequency and reward value eliminated the memory benefit for associations involving high-frequency items, suggesting that stimulus familiarity itself did not account for the influence of item frequency on memory in our task. Still, when frequency does predict value, stimulus familiarity may serve as a proxy for information utility. This familiarity signal may exert age-invariant effects on subsequent memory, whereas explicit beliefs about item frequency may more strongly facilitate subsequent memory with increasing age.

Importantly, although we observed age-related differences in participants’ learning of the structure of their environments, the strengthening of the relation between frequency reports and associative memory with increasing age suggests that age differences in learning cannot fully account for age differences in value-guided memory. Even when accounting for individual differences in participants’ explicit knowledge of the structure of the environment, older participants demonstrated a stronger relation between their beliefs about item frequency and associative memory, suggesting that they used their beliefs to guide memory to a greater degree than younger participants. In addition, we continued to observe a robust interaction between age and frequency condition on associative memory, even when controlling for age-related differences in the accuracy of both online frequency learning and explicit frequency reports. Thus, although we observed age differences in the learning of environmental regularities and in their influence on subsequent associative memory encoding, our developmental memory effects cannot be fully explained by differences in initial learning.

Our neural results further suggest that developmental differences in memory were driven by both knowledge of the structure of the environment and use of that knowledge to guide encoding. Specifically, we observed age-related increases in both overall PFC engagement as well as its value-based modulation, which may reflect developmental differences in the engagement of strategic control. Our finding that lateral prefrontal cortex activation during encoding of high- vs. low-value information may underpin memory selectivity is also in line with prior studies of motivated memory in older adults (Cohen et al., 2016). Older adults have been shown to demonstrate decreased neural activation in response to value cues (Cohen et al., 2016; Geddes et al., 2018) but preserved memory selectivity (Castel, 2007; Castel et al., 2002; Cohen et al., 2016) supported by the strategic recruitment of the left lateral PFC during encoding of high- vs. low-value information (Cohen et al., 2016). The PFC may support enhanced attention (Uncapher et al., 2011; Uncapher and Wagner, 2009) and semantic elaboration (Kirchhoff and Buckner, 2006) during encoding, and more focused (Wais et al., 2012) and organized search and selection (Badre and Wagner, 2007; Yu et al., 2018) during retrieval. From childhood to young adulthood, individuals demonstrate improvements in the implementation of strategic memory processes (Bjorkland et al., 2009; Yu et al., 2018), which are paralleled by increases in PFC recruitment during both encoding and retrieval (Ghetti et al., 2010; Ghetti and Fandakova, 2020; Ofen et al., 2007; Shing et al., 2016; Tang et al., 2018). In line with this prior work, we similarly observed age-related improvements in overall memory performance and in prefrontal recruitment during encoding and retrieval of novel associations.

The development of adaptive memory requires not only the implementation of encoding and retrieval strategies, but also the flexibility to up- or down-regulate the engagement of control in response to momentary fluctuations in information value (Castel et al., 2007; Castel et al., 2013; Hennessee et al., 2017). Importantly, value-based modulation of lateral PFC engagement during encoding mediated the relation between age and memory selectivity, suggesting that developmental change in both the representation of learned value and value-guided cognitive control may underpin the emergence of adaptive memory prioritization. Prior work examining other neurocognitive processes, including response inhibition (Insel et al., 2017) and selective attention (Störmer et al., 2014), has similarly found that increases in the flexible upregulation of control in response to value cues enhance goal-directed behavior across development (Davidow et al., 2018), and may depend on the engagement of both striatal and prefrontal circuitry (Hallquist et al., 2018; Insel et al., 2017). Here, we extend these past findings to the domain of memory, demonstrating that value signals derived from the structure of the environment increasingly elicit prefrontal cortex engagement and strengthen goal-directed encoding across childhood and into adolescence.

Further, we also demonstrate that in the absence of explicit value cues, the engagement of prefrontal control processes may reflect beliefs about information value that are learned through experience. Here, we found that differential PFC activation during encoding of high- vs. low-value information reflected individual and age-related differences in beliefs about the structure of the environment; participants who represented the average frequencies of the low- and high-frequency items as further apart also demonstrated greater value-based modulation of lateral PFC activation. It is important to note, however, that we collected explicit frequency reports after associative encoding and retrieval. Thus, the relation between PFC activation and explicit frequency reports may be bidirectional — while participants may have increased the recruitment of cognitive control processes to better encode information they believed was more valuable, the engagement of more elaborative or deeper encoding strategies that led to stronger memory traces may have also increased participants’ subjective sense of an item’s frequency (Jonides and Naveh-Benjamin, 1987).

During retrieval, we continued to observe increased activation of the caudate and dlPFC for high- vs. low-value pairs. However, this activation did not significantly vary as a function of memory difference scores or age, suggesting that the developmental differences in value-guided memory that we observed were likely driven by age-related change in encoding processes.

We found that memory prioritization varied with quadratic age, and our follow-up tests probing the quadratic age effect did not reveal evidence for significant age-related change in memory prioritization between late adolescence and early adulthood. However, in our prior behavioral work using a very similar paradigm (Nussenbaum et al., 2020), we found that memory prioritization varied with linear age only. In line with theoretical proposals (Davidow et al., 2018), subtle differences in the control demands between the two tasks (e.g. reducing the number of ‘foils’ presented on each trial of the memory test here relative to our prior study), may have shifted the age range across which we observed differences in behavior, with the more demanding variant of our task showing more linear age-related improvements into early adulthood. In addition, the specific control demands of our task may have also influenced the age at which value-guided memory emerged. Future studies should test whether younger children can modulate encoding based on the value of information if the mnemonic demands of the task are simpler.

One important caveat is that our study was cross-sectional — it will be important to replicate our findings in a longitudinal sample to measure more directly how developmental changes in cognitive control within an individual contribute to changes in their ability to selectively encode useful information. Our mediation results, in particular, must be interpreted with caution, as simulations have demonstrated that in cross-sectional samples, variables can emerge as significant mediators of age-related change due largely to statistical artifact (Hofer et al., 2006; Lindenberger et al., 2011). Indeed, our finding that PFC activation mediates the relation between age and value-guided memory does not necessarily imply that within an individual, PFC development leads to improvements in memory selectivity. Longitudinal work in which individuals’ neural activity and memory performance is sampled densely within developmental windows of interest is needed to elucidate the complex relations between age, brain development, and behavior (Hofer et al., 2006; Lindenberger et al., 2011).

We did not find evidence to support two of our predictions. First, although we initially hypothesized that both the ventral and dorsal striatum may be involved in encoding of high-value information, the activation we observed was largely within the dorsal striatum, a region that may reflect the value of goal-directed actions (Liljeholm and O'Doherty, 2012). In our task, rather than each stimulus acquiring intrinsic value during frequency learning, participants may have represented the value of the ‘action’ of remembering each pair during encoding. Second, we did not observe differences in hippocampal engagement during encoding or retrieval of high- vs. low-value associates. This is somewhat surprising, as prior work has suggested that value cues bolster memory through their influence on medial temporal lobe activity (Adcock et al., 2006). However, because our task required the use of learned value signals, memory retrieval processes may have been required on every encoding trial to recall the frequency of each item. Thus, all items presented at encoding may have triggered the engagement of hippocampal-dependent retrieval processes (Squire, 1992) that underpin many forms of memory-guided behavior including attention (Stokes et al., 2012; Summerfield et al., 2006) and decision-making (Murty et al., 2016; Shadlen and Shohamy, 2016; Wang et al., 2020). Further, in line with prior work (Davachi, 2006), we did observe increased activation in the hippocampus during encoding of associations that were subsequently remembered vs. those that were subsequently forgotten. Medial temporal activation, including hippocampal activation, may thus more strongly reflect successful memory formation — whether or not it was facilitated by value — whereas the caudate and lateral prefrontal cortex may be more sensitive to fluctuations in the engagement of value-guided cognitive control.

There are multiple routes through which value signals influence memory (Cohen et al., 2019b), and in many contexts, reward-motivated memory may not require strategic control. Value anticipation and reward delivery lead to dopaminergic release in the VTA, which projects not only to corticostriatal circuits that implement goal-directed strategy selection (Liljeholm and O'Doherty, 2012), but also directly to the hippocampus and medial temporal lobes (Adcock et al., 2006; Lisman and Grace, 2005; Murty et al., 2017; Shohamy and Adcock, 2010; Stanek et al., 2019). Given the earlier development of subcortically restricted circuitry relative to the more protracted development of cortical-subcortical pathways (Somerville and Casey, 2010), it may be the case that the direct influence of reward on memory develops earlier than the controlled pathway we studied here. Rather than eliciting strategic control through incentivizing successful memory, this pathway can be engaged through direct delivery of rewards or reinforcement signals at the time of encoding (Ergo et al., 2020; Jang et al., 2019; Rosenbaum et al., 2020; Rouhani et al., 2018). In one such study, adolescents demonstrated greater reward-based modulation of hippocampal-striatal connectivity than adults, and the strength of this connectivity predicted reward-related memory (Davidow et al., 2016). However, other studies have not found evidence for developmental change in the influence of valenced outcomes on memory (Cohen et al., 2019a; Katzman and Hartley, 2020). The influence of different motivational and reward signals on memory across development may not be straightforward — individual and developmental differences in neurocognitive processes including sensitivity to valenced feedback (Ngo et al., 2019; Rosenbaum et al., 2020), curiosity (Fandakova and Gruber, 2021), and emotional processing (Adelman and Estes, 2013; Eich and Castel, 2016) may interact, leading to complex relations between age, motivation, and memory performance. Further work is needed to characterize both the influence of different types of reward signals on memory across development, as well as the development of the neural pathways that underlie age-related change in behavior.

The present study contributes to our understanding of the neurocognitive mechanisms that support memory across development. Specifically, we addressed the question of how motivated memory may operate in the absence of explicit value cues by examining the development of the neurocognitive mechanisms that support the learning and use of information value to guide encoding and retrieval. The present findings suggest that while development is marked by improvements in the ability to learn about the statistical structure of the environment, the emergence of adaptive memory also depends centrally on age-related differences in prefrontal control. Our findings demonstrate that prefrontal cortex development has implications not just for general memory processes but for the selective prioritization of useful information — a key component of adaptive memory throughout the lifespan.

Materials and methods

Participants

Ninety participants between the ages of 8.0 and 25.9 years took part in this experiment. Thirty participants were children between the ages of 8.0 and 12.7 years (n = 16 females), 30 participants were adolescents between the ages of 13.0 and 17.7 years (n = 16 females), and 30 participants were adults between the ages of 18.3 and 25.9 years (n = 15 females). Ten additional participants were tested but excluded from all analyses due to excessive motion during the fMRI scan (n = 8; see exclusion criteria below) or technical errors during data acquisition (n = 2). We based our sample size on other functional neuroimaging studies of the development of goal-directed behavior and memory across childhood and adolescence (Insel et al., 2017; Tang et al., 2018) as well as on our prior behavioral study that showed age-related change in the use of learned value to guide memory (Nussenbaum et al., 2020). According to self- or parental-report, participants were right-handed, had normal or corrected-to-normal vision, and no history of diagnosed psychiatric or learning disorders. Participants were recruited via flyers around New York University, and from science fairs and events throughout New York City. Based on self- or parent-report, 35.6% of participants were White, 26.7% were two or more races, 24.4% were Asian, 11.1% were Black and 2.2% were Native American. Additionally, 17.8% of the sample identified as Hispanic.

Research procedures were approved by New York University’s Institutional Review Board. Adult participants provided written consent prior to participating in the study. Children and adolescents provided written assent, and their parents or guardians provided written consent on their behalf, prior to their participation. All participants were compensated $60 for the experimental session, which involved a 1 hr MRI scan. Participants were told that they would receive an additional bonus payment based on their performance in the experiment; in reality, all participants received an additional $5 bonus payment.

Prior to participating in the scanning session, child and adolescent participants who had never participated in a MRI study in our lab completed a mock scanning session to acclimate to the scanning environment. Mock scan sessions took place during a separate lab visit, at least one day in advance of scheduled scans. In the mock scanner, participants practiced staying as still as possible. We attached a Wii-mote to their heads, and set it to ‘rumble’ whenever it sensed that the participant had moved. Participants completed a series of three challenges of increasing duration (10, 30, and 90 s) and decreasing angular tolerance (10, 5, and 2 degrees) in which they tried to prevent the Wii-mote from rumbling by lying very still (Casey et al., 2018).

Experimental tasks

Request a detailed protocol

Participants completed two blocks of three tasks (Figure 1), a variant of which we used in a previous behavioral study (Nussenbaum et al., 2020). Across tasks, participants made responses with two MRI-compatible button boxes, one for each hand. In between tasks, an experimenter reminded participants of the instructions for the next part, and participants viewed a diagram indicating which fingers and buttons they should use to make their responses. The tasks were presented using Psychtoolbox Version 3 (Brainard, 1997; Kleiner et al., 2007; Pelli, 1997) for Matlab 2017a (Mathworks Inc, 2017) and displayed on a screen behind the scanner, visible to participants via a mirror attached to the MRI head coil. FMRI BOLD activity was measured over eight functional runs, which ranged in duration from approximately 4 to 7.5 min.

The structure of each block of tasks was identical, but their narratives and stimuli differed. In one set of tasks, participants were told that they had a collection of postcards they needed to mail. Each type of postcard in their collection required a different type of stamp.

In the frequency-learning task, participants were told they had to sort through their postcards to learn how many of each type they had. They were told that they had more of some types of postcards relative to other types (e.g. they might have five postcards with the same, specific blue pattern but only one postcard with a specific red pattern [Figure 1]). Participants were instructed to try to keep track of how many of each kind of postcard they had, because it would be useful to them later on. Throughout the task, participants viewed 24 images of postcards. Twelve of these images were presented once and 12 of the images were presented five times, such that participants completed 72 trials total. On each trial, a postcard appeared in the center of the screen for 2 s. Across all tasks, stimulus presentation was followed by an inter-trial interval (ITI) of 2–6 s, which consisted of a black screen with a small, white fixation cross. Participants were instructed to press the button under their right index finger when they saw a new postcard they had not seen before and to press the button under their right middle finger when they saw a repeated postcard that they had already seen within the task. Participants were instructed to respond as quickly and as accurately as possible. The specific postcard assigned to each frequency condition (1 or 5) was counterbalanced across participants. The order of image presentation was randomized for each participant.

In the second task, the associative encoding task, participants were told that they would learn the correct stamp to put on each type of postcard. Participants were instructed that in the subsequent task, they would have to stamp all of their postcards, earning one point for each postcard stamped correctly. Critically, in the associative encoding task, regardless of the number of each type of postcard that they had (i.e. 1 or 5), participants saw each type of postcard with its corresponding stamp only once. Participants were instructed that they would earn more points if they focused on remembering the stamps that went on the types of postcards that they had the most of. Thus, participants had equal exposure to the to-be-encoded associations across frequency conditions. On each trial, participants viewed one of the types of postcards from the frequency task next to an image of a unique stamp (5 s). The stamp-postcard pairs, order of the trials, and side of the screen on which the stamp and postcard appeared were randomized for each participant.

Next, participants completed retrieval. In the first part of the retrieval task, participants viewed all 24 unique postcards, one at time. When each postcard appeared, participants also saw four stamps: the correct stamp, a foil stamp that had been presented with a high-frequency postcard in the previous paired-associates task, a foil stamp that had been presented with a low-frequency postcard, and a novel stamp. Participants used the four fingers on their right hands to select one of these four stamps. Participants had six seconds to make their selection. Regardless of when they made their selection, the card and all four stamps remained on the screen for 6 s. After participants selected a stamp, a faint, gray outline appeared around it. No feedback was given until the end of the set of tasks. The order of the postcard and the location of each stamp was randomized for each participant.

After stamping all 24 unique postcards once, participants’ memory for the postcards’ original frequencies was then probed. Participants again saw all 24 unique postcards, one at a time this time with the numbers 1–7 underneath them, and they were asked to provide frequency reports. Participants used three fingers on their left hand and all four fingers on their right hand to select the number that they believed matched the number of times they saw the card in the first task. As in the previous task, participants had six seconds to make their selection. Regardless of when they made their selection, the card and all seven numbers remained on the screen for 6 s. After participants selected a number, a faint, gray outline appeared around it. The order of the postcards was randomized for each participant.

Finally, participants stamped all remaining postcards, such that they completed 48 additional memory test trials (i.e. they stamped each of the postcards in the 5-frequency condition four more times.) These trials were not included in any analyses, but their inclusion ensured that correctly encoding the stamps that belonged on the high-frequency postcards would be more valuable for participants despite each retrieval trial being worth one point. Here, participants had 4 s to make each response. We did not measure neural activation during this run, so each trial was followed by a 500 ms black screen with a white fixation cross. At the end of the memory test, participants saw a screen that displayed how many postcards they stamped correctly.

After completing the three tasks, participants were told that they were going to play a second set of similar games. The second set of tasks was identical to the first, except that the stimuli were changed from postcards and stamps to landscape pictures and picture frames. The order of the stimulus sets was counterbalanced across participants.

Prior to entering the scanner, all participants completed a short task tutorial on a laptop to learn the overall task structure and the instructions for each part. The task tutorial comprised a full set of identical tasks but with only two stimuli within each frequency condition. Participants who first did the tasks with postcards completed a tutorial with four novel postcards and novel stamps; participants who first did the tasks with pictures completed a tutorial with four novel pictures and novel frames.

Child and adolescent participants were administered the Vocabulary and Matrix Reasoning subtests of the Wechsler Abbreviated Scale of Intelligence (WASI) (Wechsler, 2011) during the mock scanning session. Adults were administered the same two subtests immediately following their scan. We followed the standard procedure to compute age-normed IQ scores for each participant based on their performance on these two sub-tests.

Analysis of behavioral data

Request a detailed protocol

All behavioral data processing and statistical analyses were conducted in R version 3.5.1 (R Development Core Team, 2018). Data were combined across blocks (but we include analyss of block effects on memory performance in Appendix 2 (Appendix 2—tables 14), Trials in which participants failed to make a response were excluded from analyses. Mixed effects models were run using the ‘afex’ package version 0.21–2 (Singmann et al., 2020). Numeric variables were z-scored across the entire data set prior to their inclusion in each model. To determine the random effects structures of our mixed effects models, we began with the maximal model to minimize Type I errors (Barr et al., 2013). We included random participant intercepts and slopes across all fixed effects (except age and WASI scores) and their interactions. We also included random stimulus intercepts and slopes across all fixed effects and their interactions. Because stimuli were randomly paired during associative encoding and only repeated, on average, around four times across participants, our stimulus random effects accounted for individual items (e.g. postcard 1) rather than pairs of items (e.g. postcard 1 and stamp 5). We set the number of model iterations to one million and used the ‘bobyqa’ optimizer. When the maximal model gave convergence errors or failed to converge within a reasonable timeframe (~24 hr), we removed correlations between random slopes and random intercepts, followed by random slopes for interaction effects, followed by random slopes across stimuli. For full details about the fixed- and random-effects structure of all models, see ‘Appendix 3: Full Model Specification and Results.’ To test the significance of the fixed effects in our models, we used likelihood ratio tests for logistic models and F tests with Satterthwaite approximations for degrees of freedom for linear models. Mediation analyses were conducted with the ‘mediation’ R package (Tingley et al., 2014) and significance of the mediation effects was assessed via 10,000 bootstrapped samples.

For our memory analyses, trials were scored as ‘correct’ if the participant selected the correct association from the set of four possible options presented during the memory test, ‘incorrect’ if the participant selected an incorrect association, and ‘missed’ if the participant failed to respond within the 6 s response window. Missed trials were excluded from all analyses. Because participants had to select the correct association from four possible options, chance-level performance was 25%. Two child participants performed at or below chance-level on the memory test. They were included in all analyses reported in the manuscript; however, we report full details of the results of our memory analyses when we exclude these two participants in Appendix 3—table 10. Importantly, our main findings remain unchanged.

Image acquisition, preprocessing, and quality assessment

Request a detailed protocol

Participants were scanned at New York University’s Center for Brain Imaging using a Siemens Prisma 3T MRI scanner with a 64-channel head coil. Anatomical data were acquired with high-resolution, T1- weighted anatomical scans using a magnetization-prepared rapidly acquired gradient echo (MPRAGE) sequence (TR = 2.3 s, TE = 2.3 ms, TI = 0.9s; 8° flip angle;. 9 mm isotropic voxels, field of view = 192 x 256 x 256 voxels; acceleration: GRAPPA 2 in the phase-encoding direction, with 24 reference lines) and T2- weighted anatomical scans using a 3D turbo spin echo (TSE) sequence (T2: TR = 3.2 s, TE = 564 ms, Echo Train Length = 314; 120° flip angle, 9 mm isotropic voxels, field of view = 240 x 256 x 256 voxels; acceleration: GRAPPA 2x2 with 32 reference lines in both the phase- and slice-encoding directions). Functional data were acquired with a T2*-weighted, multi-echo EPI sequence with the following parameters: TR=2s, TEs=12.2, 29.48, 46.76, 64.04 ms; MB factor = 2; acceleration: GRAPPA 2, with 24 reference lines; effective echo spacing:. 245 ms; 44 axial slices; 75° flip angle, 3 mm isotropic voxels, from the University of Minnesota’s Center for Magnetic Resonance Research (Feinberg et al., 2010; Moeller et al., 2010; Xu et al., 2013).

All anatomical and functional MRI data were preprocessed using fMRIPrep v.1.5.1rc2 (Esteban et al., 2019), a robust preprocessing pipeline that adjusts to create the optimal workflow for the input dataset, and then visually inspected. FMRIPrep uses tedana (for implementation details, see Kundu et al., 2013; Kundu et al., 2012) to combine each four-echo time series based on the signal decay rate of each voxel, taking a weighted average of the four echoes that optimally balances signal strength and BOLD sensitivity. This approach enables the acquisition of BOLD data with a higher signal-to-noise ratio, giving us greater sensitivity to detect neural effects of interest (Kundu et al., 2013). This combined time series was then used in subsequent preprocessing steps (e.g. susceptibility distortion correction, confound estimation, registration). Runs in which more than 15% of TRs were censored for motion (relative motion > 0.9 mm framewise displacement) were excluded from neuroimaging analyses (see Appendix 1—table 1 for the number of participants included in each analysis). Participants who did not have at least one usable run of each task (frequency learning, associative encoding, retrieval), were excluded from all behavioral and neuroimaging analyses (n = 8), leaving N = 90 participants in our analyzed sample.

Analysis of fMRI data

Request a detailed protocol

Statistical analyses were completed in FSL v. 6.0.2. (Jenkinson et al., 2012; Smith et al., 2004). Preprocessed BOLD data, registered to fMRIPrep’s MNI152 template space and smoothed with a 5 mm Gaussian kernel, were combined across runs via fixed-effects analyses and then submitted to mixed-effects GLM analyses, implemented in FEAT 6.0.0 (Woolrich et al., 2001; Woolrich et al., 2004), to estimate relevant task effects. For all GLM analyses, nuisance regressors included six motion parameters and their derivatives, framewise displacement values, censored frames, the first six anatomical noise components (aCompCor) from fMRIPrep, and cosine regressors from fMRIprep to perform high-pass filtering of the data. All task-based temporal onset regressors were convolved with a double gamma hemodynamic response function and included temporal derivatives. Analyses were thresholded using a whole-brain correction of z > 3.1 and a cluster-defining threshold of p < 0.05 using FLAME 1.

Frequency-learning GLM

Request a detailed protocol

Our frequency-learning model included six task-based temporal onset regressors. Trials were divided based on appearance count and frequency condition to create the following regressors: (1) low-frequency items the first (and only) time they appeared, (2) high-frequency items the first time they appeared, (3) high-frequency items the second time they appeared, (4) High-frequency items the third time they appeared, (5) high-frequency items the fourth time they appeared, (6) high-frequency items the fifth time they appeared.

Repetition suppression analyses

Request a detailed protocol

For each stimulus in the high-frequency condition, we examined repetition suppression by measuring activation within a parahippocampal ROI during the presentation of each item during frequency learning. We defined our ROI by taking the peak voxel (x = 30, y = −39, z = −15) from the group-level first > last item appearance contrast for high-frequency items during frequency learning and drawing a 5 mm sphere around it. This voxel was located in the right parahippocampal cortex, though we observed widespread and largely symmetric activation in bilateral parahippocampal cortex. To encompass both left and right parahippocampal cortex within our ROI, we mirrored the peak voxel sphere. For each participant, we modeled the neural response to each appearance of each item using the Least Squares Single approach (Mumford et al., 2014). Each first-level model included a regressor for the trial of interest, as well as separate regressors for the onsets of all other items, grouped by repetition number (e.g. a regressor for item onsets on their first appearance, a regressor for item onsets on their second appearance, etc.). Values that fell outside five standard deviations from the mean level of neural activation across all subjects and repetitions were excluded from subsequent analyses (18 out of 10,320 values; .01% of observations). In addition to examining neural activation as a function of stimulus repetition, we also computed an index of repetition suppression for each high-frequency item by computing the difference in mean beta values within our ROI on its first and last appearance.

Associative encoding and retrieval GLMs

Request a detailed protocol

Our associative encoding and retrieval models included six task-based temporal onset regressors. Trials were divided based on frequency condition (high- vs. low-) and subsequent memory (remembered, forgotten, missed). Missed trials were included as nuisance regressors and not included in any contrasts.

Associative encoding regions of interest (ROIs)

Request a detailed protocol

Given our a priori hypotheses about the role of the prefrontal cortex and striatum in value-guided encoding, we examined neural activation within a prefrontal cortex and striatal ROI. The specific ROIs were determined by taking the peak prefrontal voxel (x = −51, y = 42, z = 9) and the peak striatal voxel (x = −18, y = 12, z = 6) from the group-level high > low value associative encoding contrast and drawing 5 mm spheres around them.

Appendix 1

Participant information

Appendix 1—figure 1
Participant age and sex distribution.
Appendix 1—table 1
Number of participants included in each analysis.
BlockData typeFrequency- learningAssociative encodingRetrievalFrequency reports
1Behavioral89NA9090
1Neural889090NA
2Behavioral86NA8585
2Neural848181NA

Appendix 2

Supplementary results

Frequency learning: Accuracy and reaction times

Appendix 2—figure 1
Frequency learning accuracy and reaction times.

(A) During frequency learning, older participants were more accurate in identifying items as new (χ2(1) = 25.54, p < 0.001) and as repeated (χ2(1) = 33.81, p < 0.001). All participants became more accurate in identifying items as repeated as the number of repetitions increased (χ2 = 138.20, p < 0.001), though younger participants demonstrated a greater increase in accuracy throughout learning (χ2(1) = 17.52, p < 0.001). (B) Older participants also responded to both new (F(1, 85.99) = 32.51, p < 0.001) and repeated (F(1, 87.55) = 21.82, p < 0.001) items more quickly than younger participants. Reaction times to old items became faster as the a function of item repetition number (F(1, 69.94) = 282.21, p < 0.001).

Relation between age and associative memory: Two-lines test

Appendix 2—figure 2
Relation between age and associative memory.

Results from the two-lines test (Simonsohn, 2018) revealed that the influence of frequency condition on memory accuracy increased throughout childhood and early adolescence, and did not significantly decrease from adolescence into early adulthood.

Prefrontal cortex activation during encoding

Appendix 2—figure 3
Prefrontal cortex activation during encoding.

Mean beta weights averaged over voxels within a prefrontal cortex ROI (see ‘methods’ in main text) during encoding of associations involving high- vs. low-frequency items increased with age. The increase was greatest in childhood before leveling out into late adolescence and early adulthood. The line represents the best-fitting regression line from the model including both linear and quadratic age. The shaded region represents 95% confidence intervals.

Hippocampal and parahippocampal cortex activation during encoding

A priori, we expected that regions in the medial temporal lobe that have been linked to successful memory formation, including the hippocampus and parahippocampal cortex (Davachi, 2006), may be differentially engaged during encoding of high- vs. low- value information. Further, we hypothesized that the differential engagement of these regions across age may contribute to age differences in value-guided memory. Though we did not see any significant clusters of activation in the hippocampus or parahippocampal cortex in our group level high value vs. low value encoding contrast, we conducted additional ROI analyses to test these hypotheses. As with our other ROI analyses, we first identified the peak voxel (based on its z-statistic; hippocampus: x = 24, y = 34, z = 23; parahippocampal cortex: x = 22, y = 41, z = 16) in each region from our group-level contrast, and then drew 5 mm spheres around them. We then examined how average parameter estimates within these spheres related to both age and memory difference scores.

First, we ran a linear regression modeling the effects of age, WASI scores, and their interaction on hippocampal activation. We did not observe a main effect of age on hippocampal activation, (β = 0.00, SE = 0.10, p > 0.99). We did, however, observe a significant age x WASI score interaction effect (β = 0.30, SE = 0.10, p = 0.003). Next, we conducted another linear regression to examine the effects of hippocampal activation, age, WASI scores, and their interaction on memory difference scores. In contrast to our prefrontal cortex activation results, activation in the hippocampus did not relate to memory difference scores, (β = −0.02, SE = 0.03, p = 0.50).

We repeated these analyses with our parahippocampal cortex sphere. Here, we did not observe any significant effects of age on parahippocampal activation (β = −0.07, SE = 0.11, p = 0.50), nor did we observe any effects of parahippocampal activation on memory difference scores (β = 0.01, SE = 0.03, p = 0.25).

Effects of block order and type on associative memory

Block order

To examine whether participants’ memory varied across blocks, we re-ran our associative memory accuracy model with block order (e.g. 1 or 2) as an additional interacting fixed effect. Our full model included frequency condition, WASI scores, linear and quadratic age, and block order as interacting fixed effects. We included random intercepts and random slopes across frequency condition and block order for each participant, and random intercepts and random slopes across frequency condition, IQ, linear and quadratic age, and block order for each stimulus. We did not observe a significant effect of block order on associative memory (p = 0.676; Appendix 2—table 1), nor did block order interact with any other predictors (ps > 0.18). Thus, we did not observe any evidence that participants performed the task differently across blocks.

Block type

To examine whether participants’ memory varied across blocks depending on their content, we re-ran our associative memory accuracy model with block type (e.g. pictures/frames or postcards/stamps) as an additional interacting fixed effect. Our full model included frequency condition, WASI scores, linear and quadratic age, and block type as interacting fixed effects. We included random intercepts and random slopes across frequency condition and block type for each participant, and random intercepts and random slopes across frequency condition, IQ, linear and quadratic age, and block type for each stimulus. We did not observe a main effect of block type on associative memory (p = 0.061; Appendix 2—table 2; Appendix 2—figure 4). We did, however, observe a significant block type x frequency condition interaction, such that participants better remembered low-value pairs in the block with the pictures. Importantly, when we included block type as a covariate, we continued to observe a robust influence of frequency condition on associative memory, as well as significant interactions between frequency condition and both age terms.

Appendix 2—figure 4
Associative memory across blocks.

Participants demonstrated a greater influence of frequency condition on associative memory in the task block involving postcards and stamps relative to the task block involving pictures and frames (χ2(1) = 0.4.40, p = 0.036).

Appendix 2—table 1
Associative memory accuracy by frequency condition with block order.
Estimate95% CIΧ2p
Intercept0.260.12–0.40
Age1.420.52–2.338.970.003
Age2−0.97−1.87 - −0.084.390.036
WASI0.280.14–0.4115.09<0.001
Frequency Condition−0.22−0.31 - −0.1319.47<0.001
Block Order−0.02−0.11–0.070.180.676
Age x WASI0.15−0.80–1.100.100.756
Age2 x WASI−0.10−1.00–0.800.040.833
Age x Block Order−0.23−0.85–0.390.530.468
Age2 x Block Order0.13−0.48–0.750.180.675
Age x Frequency Condition−1.05−1.68 - −0.4310.310.001
Age2 x Frequency Condition0.97−0.35–1.598.940.003
WASI x Frequency Condition−0.04−0.13–0.050.650.419
Block Order x Frequency Condition−0.03−0.10–0.050.760.384
Block Order x WASI−0.05−0.15–0.041.130.288
Age x WASI x Frequency Condition−0.53−1.19–0.132.430.119
Age2 x WASI x Frequency Condition0.54−0.08–1.172.830.092
Age x Block Order x Frequency Condition−0.35−0.86–0.161.800.180
Age2 x Block Order x Frequency Condition0.27−0.23–0.771.100.294
WASI x Block Order x Frequency Condition−0.05−0.12–0.031.350.246
WASI x Age x Block Order0.05−0.61–0.700.020.887
WASI x Age2 x Block Order−0.01−0.63–0.620.000.985
Age x WASI x Frequency Condition x Block Order0.35−0.19–0.891.610.205
Age2 x WASI x Frequency Condition x Block Order−0.33−0.85–0.181.620.203
Appendix 2—table 2
Associative memory accuracy by frequency condition with block type.
Estimate95% CIΧ2p
Intercept0.260.12–0.40
Age1.400.49–2.318.610.003
Age2−0.95−1.85 - −0.054.160.041
WASI0.270.14–0.4014.46<0.001
Frequency Condition−0.22−0.31 - −0.1320.07<0.001
Block Type−0.10−0.20–0.003.520.061
Age x WASI0.17−0.79–1.120.120.734
Age2 x WASI−0.10−1.01–0.800.050.822
Age x Block Type0.23−0.39–0.850.530.465
Age2 x Block Type−0.28−0.89–0.330.790.375
Age x Frequency Condition−1.05−1.68 - −0.4210.060.002
Age2 x Frequency Condition0.960.34–1.598.640.003
WASI x Frequency Condition−0.04−0.13–0.050.650.419
Block Type x Frequency Condition−0.08−0.15 - −0.014.400.036
Block Type x WASI0.01−0.08–0.100.030.866
Age x WASI x Frequency Condition−0.51−1.18–0.152.240.135
Age2 x WASI x Frequency Condition0.52−0.11–1.152.560.109
Age x Block Type x Frequency Condition0.31−0.19–0.821.440.230
Age2 x Block Type x Frequency Condition−0.29−0.79–0.211.280.258
WASI x Block Type x Frequency Condition−0.03−0.10–0.050.440.505
WASI x Age x Block Type0.62−0.03–1.273.440.064
WASI x Age2 x Block Type−0.60−1.22–0.023.570.059
Age x WASI x Frequency Condition x Block Type−0.31−0.84–0.231.280.258
Age2 x WASI x Frequency Condition x Block Type0.34−0.17–0.851.680.195

Given that we observed a significant block type x frequency condition interaction on memory, we next examined whether the relation between age and lateral PFC activation during encoding varied across block type. To do so, we used fslmeants to extract the mean parameter estimate within our PFC ROI for the 5 vs. 1 encoding contrast for each participant, for each block. We then examined how these parameter estimates varied with age. To do so, we ran a linear mixed-effects model with linear age, quadratic age, WASI scores, and block type as interacting fixed effects, and included random participant intercepts. We did not observe a significant effect of block type on PFC activation, nor did it interact with any other predictors (ps > 0.22). In line with the analyses reported in the main text, we observed a significant relation between linear age and PFC activation (p = 0.007; Appendix 2—table 3).

Appendix 2—table 3
High- vs. low-value encoding PFC activation by age with block type.
Estimate95% CIdfFp
Intercept−0.01−0.17–0.15
Age1.600.47–2.731, 79.467.760.007
Age2−1.39−2.50 - −0.281, 79.246.040.016
WASI0.15−0.02–0.321, 86.823.060.084
Block Type−0.04−0.19–0.111, 80.300.300.585
Age x WASI0.45−0.79–1.691, 87.350.500.479
Age2 x WASI−0.53−1.69–0.631, 86.030.790.376
Age x Block Type0.11−0.94–1.161, 78.810.040.843
Age2 x Block Type−0.12−1.16–0.911, 78.590.050.818
WASI x Block Type−0.06−0.22–0.101, 86.370.540.462
Age x WASI x Block Type−0.73−1.89–0.441, 86.581.490.226
Age2 x WASI x Block Type0.68−0.41–1.771, 85.261.480.227

Finally, we examined how PFC activation across blocks influenced memory difference scores via a linear mixed-effects model with PFC activation, age, WASI scores, and block type as interacting fixed effects. Our model also included random participant intercepts. Here, including quadratic age did not improve model fit (Χ2(8) = 0.00, p = 1). We found that PFC activation related to memory difference scores (p = 0.002; Appendix 2—table 4). No other main effects or interactions were significant (ps > 0.063).

Appendix 2—table 4
Memory difference scores by PFC activation and age with block type.
Estimate95% CIdfFp
Intercept0.090.05–0.13
PFC Activation0.060.02–0.101, 153.789.850.002
Age0.02−0.01–0.061, 84.351.520.221
WASI−0.00−0.04–0.041, 86.860.040.836
Block Type0.03−0.00–0.071, 82.482.920.091
PFC Activation x Age0.01−0.03–0.061, 153.530.280.598
PFC Activation x WASI0.03−0.01–0.081, 154.292.130.147
Age x WASI−0.01−0.05–0.031, 86.370.220.642
PFC Activation x Block Type−0.01−0.04–0.031, 151.410.090.770
Age x Block Type0.01−0.03–0.041, 82.910.130.718
WASI x Block Type0.02−0.02–0.051, 85.400.760.385
PFC Activation x Age x WASI−0.01−0.06–0.041, 152.590.140.711
PFC Activation x Age x Block Type0.03−0.02–0.081, 152.381.560.214
PFC Activation x WASI x Block Type−0.04−0.09–0.001, 151.643.490.064
Age x WASI x Block Type−0.02−0.05–0.011, 85.201.230.270
PFC Activation x Age x WASI x Block Type0.05−0.01–0.101, 154.253.000.085

Appendix 3

Full model specification and results

For each model described in the manuscript, we report here its full random-effects structure (when relevant) and effect estimates.

Model 1: Frequency-learning accuracy: new items

We examined how participants’ accuracy in identifying new items during frequency-learning varied as a function of age, WASI scores, and their interaction via a mixed-effects logistic regression (Appendix 3—table 1). We included random intercepts for each participant and each stimulus. Including quadratic age in the model did not improve model fit (Χ2(2) = 3.29, p = 0.19).

Appendix 3—table 1
Frequency-learning accuracy: new items.
Estimate95% CIΧ2p
Intercept3.082.73–3.44
Age0.920.58–1.2525.52<0.001
WASI0.430.09–0.776.180.013
Age x WASI0.25−0.08–0.572.250.134

Model 2: Frequency-learning accuracy: repeated items

We examined how participants’ accuracy in identifying repeated items during frequency learning varied as a function of the number of times the item had appeared, age, WASI scores, and their interactions via a mixed-effects logistic regression (Appendix 3—table 2). We included random intercepts and random slopes across item appearances for each participant and random intercepts for each stimulus. Including quadratic age did not improve model fit (Χ2(4) = 1.80, p = 0.77).

Appendix 3—table 2
Frequency-learning accuracy: repeated item appearances.
Estimate95% CIΧ2p
Intercept3.833.46–4.20
Appearance1.531.28–1.78138.03<0.001
Age0.970.64–1.2933.43<0.001
WASI0.460.13–0.797.580.006
Appearance x Age0.450.24–0.6717.41<0.001
Appearance x WASI0.05−0.17–0.260.180.672
Age x WASI0.12−0.20–0.450.570.449
Appearance x Age x WASI0.04−0.17–0.250.140.707

Model 3: Frequency-learning reaction times: new items

We examined how participants’ reaction times when they correctly identified new items during frequency learning varied as a function of age, WASI scores, and their interaction via a mixed-effects linear regression (Appendix 3—table 3). We included random intercepts for each participant and each stimulus, and random slopes across age and WASI scores for each stimulus. We also estimated the correlation between random stimulus intercepts and slopes. Including quadratic age did not improve model fit (Χ2(13) = 21.26, p = 0.068).

Appendix 3—table 3
Frequency-learning reaction times: new items.
Estimate95% CIdfFp
Intercept1.121.09–1.15
Age−0.08−0.11 - −0.051, 85.9932.51<.001
WASI−0.01−0.04–0.021, 82.340.56.457
Age x WASI−0.02−0.05 - −0.011, 83.142.12.149

Model 4: Frequency-learning reaction times: repeated items

We examined how participants’ reaction times when they correctly identified repeated items during frequency learning varied as a function of the number of times the item had appeared, age, WASI scores, and their interactions via a mixed-effects linear regression (Appendix 3—table 4). We included random intercepts and random slopes across item appearances for each participant, random intercepts and slopes across age, WASI scores, and item appearances for each stimulus, and estimated the correlation between random stimulus intercepts and slopes. Including quadratic age did not improve model fit (Χ2(9) = 3.18, p = 0.96).

Appendix 3—table 4
Frequency-learning reaction times: repeated items.
Estimate95% CIdfFp
Intercept1.031.00–1.06
Age−0.07−0.10 – −0.041, 87.5521.82<0.001
WASI−0.03−0.06 – −0.011, 86.272.650.108
Appearance−0.08−0.09 – −0.071, 69.94282.21<0.001
Age x WASI−0.01−0.03–0.021, 84.970.220.641
Age x Appearance−0.01−0.01–0.001, 77.061.260.265
WASI x Appearance0.00−0.01–0.011, 75.790.000.992
Age x WASI x Appearance0.00−0.01–0.011, 74.960.680.413

Model 5: Parahippocampal cortex neural activation by stimulus repetition and age

For items in the high-frequency condition, we examined how neural activation in a parahippocampal cortex ROI varied as a function of age, quadratic age, stimulus repetition number, quadratic stimulus repetition number, WASI scores, and their interactions (Appendix 3—table 5). We included random intercepts for each participant and stimulus, random slopes across linear and quadratic repetition number for each participant, and random slopes across linear and quadratic repetition number, linear and quadratic age, and their interactions for each stimulus stimuli.

Appendix 3—table 5
Parahippocampal cortex neural activation by stimulus repetition and age.
Estimate95% CIdfFp
Intercept65.7052.23–79.17
Age−78.41−164.25–7.431, 82.783.210.077
Age282.54−2.27–167.351, 82.773.640.060
Repetition−30.20−40.89 – −19.501, 5015.9430.64<0.001
Repetition214.524.10–24.931, 9881.007.470.006
WASI−1.11−13.49–11.271, 83.310.030.861
Age x Repetition101.6527.40–175.901, 7267.467.200.007
Age x Repetition2−88.18161.49 - −14.871, 9857.855.560.018
Age2 x Repetition97.99−171.28 - −24.711, 7260.706.870.009
Age2 x Repetition282.7610.40–155.111, 9854.925.030.025
WASI x Age28.87−61.72–119.471, 83.510.390.534
WASI x Age2−20.73−106.34–64.881, 83.470.230.636
WASI x Repetition−7.56−18.46–3.351, 7402.991.840.175
WASI x Repetition27.40−3.38–18.181, 7857.101.810.178
WASI x Age x Repetition−52.32−130.26–25.611, 7243.451.730.188
WASI x Age x Repetition242.15−34.79–119.081, 9868.651.150.283
WASI x Age2 x Repetition45.97−27.58–119.531, 7235.301.500.221
WASI x Age2 x Repetition2−38.01−110.62–34.591, 9867.591.050.305

Model 6: Repetition suppression indices and age

For items in the high-frequency condition, we examined how repetition suppression varied as a function of age, quadratic age, WASI scores, and their interactions (Appendix 3—table 6). We included random intercepts for each participant and stimulus, and random slopes across age, quadratic age, and WASI scores for each stimulus.

Appendix 3—table 6
Repetition suppression indices.
Estimate95% CIdfFp
Intercept45.5435.63–55.45
Age−61.34−8.56–10.771, 78.323.970.050
Age266.52−121.70 - −0.981, 77.554.800.031
WASI1.117.01–126.031, 58.060.050.823
Age x WASI60.90−2.53–124.341, 77.383.540.064
Age2 x WASI−51.17−111.03–8.701, 77.162.810.098

Model 7: Frequency report error magnitudes

We examined how the magnitude of participants’ errors in their frequency reports varied as a function of age, WASI scores, frequency condition, and their interactions via a mixed-effects linear regression (Appendix 3—table 7). We included random intercepts and random slopes across frequency conditions for each participant and random intercepts and random slopes across age, WASI scores, and frequency conditions for each stimulus. Including quadratic age did not improve model fit (Χ2(8) = 7.96, p = 0.437).

Appendix 3—table 7
Frequency report error magnitudes.
Estimate95% CIdfFp
Intercept1.211.12–1.30
Age−0.18−0.27 - - 0.101, 94.3017.57<0.001
WASI−0.11−0.19 - −0.021, 83.806.470.014
Frequency Condition−0.00−0.11–0.111, 93.810.000.993
Age x WASI−0.07−0.15–0.011, 86.353.240.075
Age x Frequency Condition−0.05−0.17–0.061, 85.480.950.332
WASI x Frequency Condition−0.03−0.14–0.091, 86.550.200.652
Age x WASI x Frequency Condition−0.07−0.17–0.031, 85.941.770.187

Model 8: Frequency reports by repetition suppression indices

For items in the high-frequency condition, we examined how frequency reports varied as a function of age, WASI scores, repetition suppression, and their interactions via a mixed-effects linear regression (Appendix 3—table 8). We included random intercepts and random slopes across repetition suppression for each participant, and random intercepts and random slopes across repetition suppression, WASI scores, age, and their interactions for each stimulus. Including a quadratic age term did not improve model fit (Χ2(8) = 6.45, p = 0.60).

Appendix 3—table 8
Frequency reports by repetition suppression.
Estimate95% CIdfFp
Intercept4.444.26–4.63
Age0.260.09–0.421, 82.878.930.004
WASI0.190.02–0.361, 85.194.770.032
Repetition Suppression0.00−0.06–0.071, 1360.740.010.903
Age x WASI0.09−0.07–0.251, 84.131.340.251
Age x Repetition Suppression0.06−0.00–0.121, 938.873.610.058
WASI x Repetition Suppression0.04−0.02–0.101, 58.811.520.222
Age x WASI x Repetition Suppression−0.03−0.08–0.031, 313.721.100.296

Model 9: Associative memory accuracy

We examined how memory accuracy varied as a function of age, quadratic age, WASI scores, frequency condition, and their interactions via a mixed-effects logistic regression (Appendix 3—table 9). We included random intercepts and random slopes across frequency conditions for each participant, and random intercepts and random slopes across frequency condition, WASI scores, age, and quadratic age for each stimulus.

Appendix 3—table 9
Associative memory accuracy by frequency condition.
Estimate95% CIΧ2p
Intercept0.260.12–0.40
Age1.380.49–2.288.680.003
Age2−0.95−1.83 – −0.064.240.039
WASI0.260.13–0.3914.18<0.001
Frequency Condition−0.21−0.30 – −0.1319.73<0.001
Age x WASI0.18−0.76–1.120.140.704
Age2 x WASI−0.12−1.01–0.770.070.789
Age x Frequency Condition−1.06−1.68 – −0.4510.740.001
Age2 x Frequency Condition0.980.37–1.599.270.002
WASI x Frequency Condition−0.04−0.13–0.050.860.355
Age x WASI x Frequency Condition−0.50−1.15–0.152.260.133
Age2 x WASI x Frequency Condition0.52−0.10–1.132.650.104

Model 10: Associative memory accuracy excluding participants who performed below chance

Two participants (both children) responded correctly to 25% or fewer memory test trials. We re-ran our memory accuracy model, excluding these two participants (Appendix 3—table 10).

Appendix 3—table 10
Associative memory accuracy by frequency condition (below-chance subjects excluded).
Estimate95% CIΧ2p
Intercept0.300.16–0.44
Age1.190.29–2.096.480.011
Age2−0.79−1.68–0.102.980.084
WASI0.240.11–0.3712.44<0.001
Frequency Condition−0.22−0.31 – −0.1320.04<0.001
Age x WASI0.31−0.62–1.240.430.513
Age2 x WASI−0.23−1.11–0.660.250.615
Age x Frequency Condition−1.07−1.70 - −0.4310.250.001
Age2 x Frequency Condition0.990.36–1.618.970.003
WASI x Frequency Condition−0.04−0.14–0.050.810.368
Age x WASI x Frequency Condition−0.50−1.16–0.162.170.141
Age2 x WASI x Frequency Condition0.52−0.11–1.152.540.111

Model 11: Associative memory accuracy controlling for individual differences in frequency learning

We examined how memory accuracy varied as a function of age, quadratic age, WASI scores, frequency condition, and their interactions via a mixed-effects logistic regression (Appendix 3—table 11). We also included mean frequency report error magnitudes, overall mean accuracy during frequency learning, and mean accuracy on the last appearance of each item during frequency learning as non-interacting fixed effects. We included random intercepts and random slopes across frequency conditions for each participant, and random intercepts and random slopes across frequency condition, WASI scores, age, and quadratic age for each stimulus.

Appendix 3—table 11
Associative memory accuracy by frequency condition (with frequency-learning covariates).
Estimate95% CIΧ2p
Intercept0.250.12–0.38
Age0.59−0.31–1.491.620.203
Age2−0.30−1.17–0.560.470.491
WASI0.160.03–0.296.020.014
Frequency Condition−0.21−0.30 – −0.1319.65<0.001
Mean Frequency Report Error Magnitude−0.26−0.39 – −0.1213.05<0.001
Frequency-learning Accuracy0.17−0.04–0.382.360.125
Frequency-learning Accuracy (last item appearance)−0.10−0.30–0.091.060.304
Age x WASI0.11−0.75–0.960.060.807
Age2 x WASI−0.08−0.89–0.720.040.843
Age x Frequency Condition−1.06−1.67 - −0.4410.590.001
Age2 x Frequency Condition0.970.36–1.589.150.002
WASI x Frequency Condition−0.04−0.13–0.050.830.362
Age x WASI x Frequency Condition−0.50−1.15–0.152.220.136
Age2 x WASI x Frequency Condition0.51−0.10–1.132.610.106

Model 12: High- vs. low-value encoding caudate activation and age

We ran a linear regression to examine the effects of age, WASI scores, and their interaction on differential caudate activation during encoding of high- vs. low-value information (Appendix 3—table 12). Including quadratic age did not improve model fit (Χ2(2) = 2.26, pp = 0.11).

Appendix 3—table 12
High vs. low-value encoding caudate activation by age.
EstimateSEtp
Intercept−0.07.104
Age0.16.1071.550.126
WASI0.18.1101.640.105
Age x WASI−0.29.101−2.860.005

Model 13: High- vs. low-value encoding PFC activation and age

We ran a linear regression to examine the effects of linear and quadratic age, WASI scores, and their interactions on differential PFC activation during encoding of high- vs. low-value information (Appendix 3—table 13).

Appendix 3—table 13
High vs. low-value encoding PFC activation by age.
EstimateSEtp
Intercept0.00.105
Age1.97.7432.650.009
Age2−1.73.734−2.350.021
WASI0.26.1092.340.022
Age x WASI0.93.7891.180.240
Age2 x WASI−1.02.745−1.370.174

Model 14: Relation between frequency reports and associative memory accuracy

We examined how memory accuracy varied as a function of age, quadratic age, WASI scores, frequency reports, and their interactions via a mixed-effects logistic regression (Appendix 3—table 14). We included random intercepts and random slopes across frequency reports for each participant, and random intercepts and random slopes across frequency reports, age, quadratic age, and WASI scores for each stimulus.

Appendix 3—table 14
Associative memory accuracy by frequency report.
Estimate95% CIΧ2p
Intercept0.260.12–0.40
Age1.460.55–2.389.250.002
Age2−1.01−1.92 – −0.114.640.031
WASI0.270.13–0.4013.94<0.001
Frequency Report0.280.19–0.3731.20<0.001
Age x WASI0.15−0.81–1.120.090.759
Age2 x WASI−0.10−1.01–0.820.040.838
Age x Frequency Report1.130.46–1.7910.370.001
Age2 x Frequency Report−1.07−1.73 – −0.419.500.002
WASI x Frequency Report0.02−0.08–0.110.100.754
Age x WASI x Frequency Report0.37−0.35–1.091.000.316
Age2 x WASI x Frequency Report−0.33−1.01–0.350.890.345

Model 15: Influence of repetition suppression on associative memory accuracy

For associations involving items in the high-frequency condition, we examined how memory accuracy varied as a function of age, quadratic age, WASI scores, repetition suppression, and their interactions via a mixed-effects logistic regression (Appendix 3—table 15). We included random intercepts and random slopes across repetition suppression for each participant, and random intercepts and random slopes across repetition suppression, WASI scores, age, quadratic age, and their interactions for each stimulus.

Appendix 3—table 15
Associative memory accuracy by repetition suppression.
Estimate95% CIΧ2p
Intercept0.510.32–0.69
Age2.631.43–3.8316.87<0.001
Age2−2.13−3.31 – −0.9411.55<0.001
WASI0.310.14–0.4911.47<0.001
Repetition Suppression0.230.10–0.3711.21<0.001
Age x WASI0.78−0.49–2.051.440.230
Age2 x WASI−0.75−1.95–0.451.470.225
Age x Repetition Suppression−0.42−1.32–0.490.790.374
Age2 x Repetition Suppression0.64−0.28–1.571.790.181
WASI x Repetition Suppression0.00−0.13–0.140.000.954
Age x WASI x Repetition Suppression−0.17−1.02–0.680.150.700
Age2 x WASI x Repetition Suppression0.27−0.57–1.100.370.541

Model 16: Effects of frequency reports and repetition suppression on associative memory accuracy

For associations involving items in the high-frequency condition, we examined how memory accuracy varied as a function of age, quadratic age, WASI scores, repetition suppression, frequency reports, and their interactions via a mixed-effects logistic regression (Appendix 3—table 16). We included random intercepts and random slopes across repetition suppression, frequency reports, and their interaction for each participant. We also included random intercepts and random slopes across repetition suppression, frequency reports, age, quadratic age, and WASI scores for each stimulus.

Appendix 3—table 16
Associative memory accuracy by repetition suppression and frequency reports.
Estimate95% CIΧ2p
Intercept0.510.32–0.69
Repetition Suppression0.230.09–0.3610.250.001
WASI0.260.08–0.447.590.006
Frequency Report0.30.17–0.4221.16<0.001
Age2.471.25–3.6914.4<0.001
Age2−2.02−3.23 – −0.819.980.002
Repetition Suppression x WASI0.00−0.13–0.140.000.968
Repetition Suppression x Frequency Report−0.11−0.25–0.042.180.140
WASI x Frequency Report−0.05−0.18–0.080.480.488
Repetition Suppression x Age−0.12−1.05–0.810.060.804
Repetition Suppression x Age20.33−0.62–1.270.450.503
WASI x Age0.92−0.36–2.211.960.161
WASI x Age2−0.89−2.12–0.332.040.153
Frequency Report x Age0.12−0.74–0.990.080.783
Frequency Report x Age2−0.09−0.97–0.790.040.843
RS x WASI x Frequency Report−0.04−0.19–0.110.270.603
RS x WASI x Age−0.28−1.16–0.610.360.550
RS x WASI x Age20.39−0.49–1.260.710.400
RS x Frequency Report x Age0.13−0.78–1.030.070.786
RS x Frequency Report x Age2−0.12−1.09–0.850.060.809
WASI x Frequency Report x Age−0.33−1.31–0.650.440.509
WASI x Frequency Report x Age20.44−0.52–1.400.790.374
RS x WASI x Frequency Report x Age−0.73−1.68–0.222.290.130
RS x WASI x Frequency Report x Age20.83−0.14–1.802.860.091

Model 17: Relation between age and mean repetition suppression indices

We ran a linear regression to examine how average repetition suppression indices (across items, for each participant) varied as a function of age, quadratic age, WASI scores, and their interactions.

Appendix 3—table 17
Mean repetition suppression indices by age.
EstimateSEtp
Intercept44.764.53
Age−59.0031.82−1.850.067
Age264.9331.372.070.042
WASI1.864.720.390.695
Age x WASI57.0633.891.680.096
Age2 x WASI−48.0231.93−1.500.136

Model 18: Relation between mean repetition suppression indices and neural activation in caudate

We ran a linear regression to examine how differential caudate activation in response to high- vs. low-value information during encoding related to average repetition suppression indices age, WASI scores, and their interactions (Appendix 3—table 18). Including quadratic age did not improve model fit (Χ2(4) = 1.37, p = 0.25).

Appendix 3—table 18
Caudate activation by repetition suppression indices.
EstimateSEtp
Intercept8.231.54
Repetition Suppression2.311.601.440.153
Age1.481.640.910.367
WASI1.931.641.170.244
Repetition Suppression x Age1.291.420.910.366
Repetition Suppression x WASI0.261.540.170.864
Age x WASI−4.401.50−2.940.004
Repetition Suppression x Age x WASI−0.091.22−0.070.945

Model 19: Relation between mean repetition suppression indices and PFC neural activation

We ran a linear regression to examine how differential PFC activation in response to high- vs. low-value information during encoding related to average repetition suppression indices, age, quadratic age, WASI scores, and their interactions (Appendix 3—table 19).

Appendix 3—table 19
PFC activation by repetition suppression indices.
EstimateSEtp
Intercept28.664.51
Repetition Suppression−6.784.83−1.400.165
Age103.3635.332.930.004
Age2−94.2535.13−2.680.009
WASI13.594.742.870.005
Repetition Suppression x Age−54.1534.13−1.590.112
Repetition Suppression x Age253.8532.791.640.105
Repetition Suppression x WASI−3.284.49−0.730.467
Age x WASI30.4334.440.880.380
Age2 x WASI−38.5932.41−1.910.237
Repetition Suppression x Age x WASI−41.3329.80−1.390.169
Repetition Suppression x Age2 x WASI41.4627.581.500.137

Model 20: Relation between age and frequency distance

We ran a linear regression to examine how mean frequency distances related to age, quadratic age, WASI scores, and their interactions (Appendix 3—table 20).

Appendix 3—table 20
Frequency distance by age.
EstimateSETp
Intercept2.23.09
Age2.75.674.10<0.001
Age2−2.28.66−3.44<0.001
WASI0.38.103.89<0.001
Age x WASI0.33.710.460.646
Age2 x WASI−0.12.67−0.180.857

Model 21: Relation between frequency distance and neural activation in caudate

We ran a linear regression to examine how differential caudate activation in response to high- vs. low-value information during encoding related to average repetition suppression indices age, WASI scores, and their interactions (Appendix 3—table 21). Including quadratic age did not improve model fit (Χ2(4) = 1.36, p = 0.25).

Appendix 3—table 21
Caudate activation by frequency distance.
EstimateSEtp
Intercept7.001.78
Frequency Distance2.551.861.370.175
Age1.391.890.740.463
WASI1.131.770.640.526
Frequency Distance x Age2.271.771.290.202
Frequency Distance x WASI−1.501.77−0.850.399
Age x WASI−5.491.64−3.430.001
Frequency Distance x Age x WASI1.711.401.220.227

Model 22: Relation between frequency distance and PFC neural activation

We ran a linear regression to examine how differential PFC activation in response to high- vs. low-value information during encoding related to frequency distance, age, WASI scores, and their interactions (Appendix 3—table 22). Including quadratic age did not improve model fit (Χ2(4) = 1.49, p = 0.21).

Appendix 3—table 22
PFC activation by frequency distance.
EstimateSEtp
Intercept−0.020.12
Frequency Distance0.420.123.360.001
Age−0.030.13−0.220.824
WASI0.080.120.640.522
Frequency Distance x Age−0.180.12−1.510.136
Frequency Distance x WASI0.150.121.230.223
Age x WASI−0.170.11−1.520.132
Frequency Distance x Age x WASI0.050.090.520.607

Model 23: Effects of frequency distance and PFC neural activation on memory difference scores

We ran a linear regression to examine how memory difference scores were related to differential PFC activation in response to high- vs. low-value information during encoding, frequency distance, age, quadratic age, WASI scores, and their interactions (Appendix 3—table 23).

Appendix 3—table 23
Memory difference scores by PFC activation and frequency distance.
EstimateSEtp
Intercept0.07
Frequency Distance−0.020.03−0.550.582
Age0.560.22.830.006
Age2−0.510.2−2.60.012
WASI0.020.030.850.398
PFC Activation0.080.041.950.055
Frequency Distance x Age0.290.21.50.139
Frequency Distance x Age2−0.270.2−1.360.178
Frequency Distance x WASI0.010.030.380.709
Age x WASI0.030.190.170.866
Age2 x WASI−0.050.18−0.250.800
Frequency Distance x PFC Activation0.010.040.250.800
Age x PFC Activation−0.150.34−0.430.672
Age2 x PFC Activation0.210.340.620.538
WASI x PFC Activation−0.040.04−0.840.406
Frequency Distance x Age x WASI−0.360.22−1.660.102
Frequency Distance x Age2 x WASI0.330.21.620.111
Frequency Distance x Age x PFC Activation−0.070.27−0.240.809
Frequency Distance x Age2 x PFC Activation0.080.280.280.778
Frequency Distance x WASI x PFC Activation−0.050.03−1.490.142
Age2 x WASI x PFC Activation0.460.361.290.202
Age2 x WASI x PFC Activation−0.460.36−1.280.204
Frequency Distance x Age2 x WASI x PFC Activation0.270.270.990.324
Frequency Distance x Age2 x WASI x PFC Activation−0.280.27−1.070.288

Appendix 4

Supplemental neural results

Appendix 4—table 1
Frequency-learning: Last vs. first item appearance cluster table.
RegionxyzCluster sizez-max
Frontal pole−4251024196.33
Precuneus0−663313228.99
Left lateral occipital cortex / angular gyrus−57−663013197.2
Right lateral occipital cortex / angular gyrus51−63336376.03
Right middle temporal gyrus66−33−123045.48
Right cerebellum15−87−271645.33
Precentral gyrus3−18751244.53
Left cerebellum−42−75−42924.72
Left middle temporal gyrus−57-3−27734.44
Left caudate-6129624.91
Right caudate9246605.04
Occipital pole-3−9612464.81
Appendix 4—table 2
Frequency-learning: First vs. last item appearance cluster table.
RegionxyzCluster sizez-max
Right temporal fusiform cortex / lateral occipital cortex / parahippocampal gyrus30−39−1521378.62
Left temporal fusiform cortex / lateral occipital cortex / parahippocampal gyrus−33−63−1518587.53
Cingulate gyrus99421006
Right precuneus18−519806.05
Left postcentral gyrus−51−1557744.73
Right precentral gyrus42333735.54
Juxtapositional lobule cortex6657244.11
Left amygdala−24-6−15224.23
Cingulate gyrus3-333224.68
Central opercular cortex39-315214.8
Appendix 4—table 3
Encoding: Encoding vs. baseline by linear age cluster table.
RegionxyzCluster sizez-max
Right lateral occipital cortex45−81211325
Left precentral gyrus / middle frontal gyrus−5412331205.9
Left lateral occipital cortex−45−8421544.69
Left lateral occipital cortex−27−7242504.32
Right lateral occipital cortex30−6945444.17
Left cerebellum−18−45−48394.55
Superior frontal gyrus-31260394.35
Left supramarginal gyrus−36−4536374.57
Left middle frontal gyrus−33363334.6
Right superior parietal lobule33−4542313.76
Right inferior frontal gyrus481230243.97
Appendix 4—table 4
Encoding: High- vs. low-value cluster table.
RegionxyzCluster sizez-max
Superior parietal lobule / lateral occipital cortex / temporal occipital fusiform cortex / cerebellum−33−545142626.23
Left frontal pole / inferior frontal gyrus / middle frontal gyrus−5142917656.92
Left caudate / thalamus−181262325.67
Right caudate181812544.17
Right precentral gyrus51933504.26
Right precentral gyrus24-954474.79
Left cerebellum-3−510384.73
Right frontal pole513912354.53
Right postcentral gyrus42−3651284.32
Right putamen2715-3264.52
Left thalamus-3−24-3234.65
Left putamen−30−15-6205.7
Appendix 4—table 5
Encoding: High- vs. low-value by memory difference scores cluster table.
RegionxyzCluster sizez-max
Left lateral occipital cortex−45−69-63774.78
Left middle frontal gyrus / inferior frontal gyrus−4821272324.95
Right lateral occipital cortex (inferior)39−9031894.89
Right temporal occipital fusiform cortex42−57-9874.45
Right lateral occipital cortex (superior)27−6633614.83
Appendix 4—table 6
Encoding: Remembered vs. not remembered cluster table.
RegionxyzCluster sizez-max
Right lateral occipital cortex / temporal occipital fusiform gyrus48−75-912965.75
Left lateral occipital cortex / temporal occipital fusiform gyrus / inferior temporal gyrus−48−72-612735.97
Left inferior frontal gyrus−489271034.22
Left inferior frontal gyrus−5721-3403.81
Right hippocampus / amygdala24-6−21214.06
Appendix 4—table 7
Retrieval: Retrieval vs. baseline by linear age cluster table.
RegionxyzCluster sizez-max
Right lateral occipital cortex45−81211475.28
Left lateral occipital cortex−39−8721594.25
Right precentral gyrus / inferior frontal gyrus51321424.21
Right precentral gyrus39−1239414.4
Left postcentral gyrus / supramarginal gyrus−63−2448384.45
Left precentral gyrus−60621364.39
Right lateral occipital cortex27−6057343.76
Cingulate gyrus / left thalamus3−330304.74
Left supramarginal gyrus−69−2424284.33
Appendix 4—table 8
Retrieval: High- vs. low-value cluster table.
RegionxyzCluster sizez-max
Precuneus cortex0−72392994.81
Left lateral occipital cortex−33−69512365.28
Left caudate / thalamus−12-6151285.15
Right cerebellum3−81−301256.12
Left inferior frontal gyrus / middle frontal gyrus−4821241164.57
Left frontal orbital cortex−3030-3884.82
Right cerebellum39−69−36854.65
Right caudate15-321774.96
Left middle frontal gyrus−451248744.31
Cerebellum-3−60−36604.57
Left inferior temporal gyrus−57−60−15604.33
Cingulate gyrus0−333254.09
Right frontal orbital cortex33333214.69
Left frontal pole−36573203.79
Right lateral occipital cortex27−6642204.24

Data availability

Task code, behavioral data, and analysis code are available on the Open Science Framework: https://osf.io/2fkbj/ Unthresholded z-statistic maps from the neuroimaging analyses are available on Neurovault: https://neurovault.org/collections/BUMNZQXA/ BIDS-formatted neuroimaging data are available on Open Neuro: https://openneuro.org/datasets/ds003499.

The following data sets were generated
    1. Nussenbaum K
    2. Hartley CA
    (2021) Open Science Framework
    ID 2fkb. Developmental change in prefrontal cortex recruitment supports the emergence of value-guided memory.
    1. Nussenbaum K
    (2021) OpenNeuro
    Developmental change in prefrontal cortex recruitment supports the emergence of value-guided memory.
    https://doi.org/10.18112/openneuro.ds003499.v1.0.1

References

  1. Book
    1. Bjorkland DF
    2. Dukes C
    3. Brown DR
    (2009) The development of memory strategies
    In: Courage M, Cowan N, editors. The Development of Memory in Infancy and Childhood. Psychology Press. pp. 145–176.
    https://doi.org/10.4324/9780203934654
    1. Jonides J
    2. Naveh-Benjamin M
    (1987) Estimating frequency of occurrence
    Journal of Experimental Psychology: Learning, Memory, and Cognition 13:230–240.
    https://doi.org/10.1037/0278-7393.13.2.230
  2. Software
    1. R Development Core Team
    (2018) R: A language and environment for statistical computing
    R Foundation for Statistical Computing, Vienna, Austria.
    1. Rouhani N
    2. Norman KA
    3. Niv Y
    (2018) Dissociable effects of surprising rewards on learning and memory
    Journal of Experimental Psychology: Learning, Memory, and Cognition 44:1430–1443.
    https://doi.org/10.1037/xlm0000518
  3. Book
    1. Wechsler D
    (2011)
    Wechsler Abbreviated Scale of Intelligence (WASI-II
    NCS Pearson.

Decision letter

  1. Thorsten Kahnt
    Reviewing Editor; Northwestern University, United States
  2. Christian Büchel
    Senior Editor; University Medical Center Hamburg-Eppendorf, Germany
  3. Vishnu Murty
    Reviewer; Temple University, United States

Our editorial process produces two outputs: (i) public reviews designed to be posted alongside the preprint for the benefit of readers; (ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Acceptance summary:

This study will be relevant to those interested in the neurodevelopment of reward learning, episodic memory, and memory-guided decision-making. The combination of a clever task and thorough data analysis make this an impactful paper.

Decision letter after peer review:

Thank you for submitting your article "Developmental change in prefrontal cortex recruitment supports the emergence of value-guided memory" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, one of whom is a member of our Board of Reviewing Editors, and the evaluation has been overseen by Christian Büchel as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Vishnu Murty (Reviewer #2).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

All reviewers agreed that your manuscript presents important and interesting results suitable for publication in eLife. They generally liked the study and felt that the experimental design and data analysis approaches were clever and thorough. However, reviewers also raised several issues that we would like you to address in a revised version of the manuscript:

1. Do younger subjects not prioritize memory or do they not learn what information is valuable (e.g., R1, comment 3; R2, comment 2; R3, comment 2)?

2. Please address the modest relationship between the repetition suppression index in the parahippocampal cortex and age (e.g., R1, comment 1and2).

3. It would be important to be more specific regarding anatomical subregions in the introduction and discussion (R2, comment 1).

4. It would be interesting to connect the neural measures during initial frequency learning and memory-guided adaptive encoding (e.g., R2, comment 5).

5. Please discuss problems with cross-sectional designs, especially as they relate to mediation analyses (e.g., R3, comment 1).

6. Please discuss the implications of the quadratic effects (e.g., R2, comment 3).

We also encourage you to consider the other suggestions made by individual reviewers.

Reviewer #1 (Recommendations for the authors):

1. Given the modest relationship between the repetition suppression index in the parahippocampal cortex and age, it would be important to make sure this relationship is robust. One way would be to use more data to improve the estimation of the repetition suppression index. For instance, data from all presentations could be used to estimate linear and quadratic effects of repetition number. This would not only improve the reliability of the estimates, but also offer the opportunity to examine whether the dynamics of frequency learning across multiple presentations changes with age.

2. At the very least, it would be important to include a scatter plot of repetition index versus age in the main text, and to show the parameter estimates across all repetitions in the parahippocampal cortex (e.g., per age group).

3. Given the effects of age on both initial frequency learning (new vs. old response accuracy and frequency estimation error) and the effects of item frequency on memory, it would be important to control for frequency learning when testing the effects of age on memory prioritization, and whether lateral PFC mediates this relationship.

4. Please include more information on how the four echo time series were combined. What is considered optimal in this regard?

Reviewer #2 (Recommendations for the authors):

Before I start critiquing individual components of the study, I want to say I am quite excited about these results, and believe this a very clever design and interesting neuroimaging findings.

1. I found the introduction quite nice, and it really provided a strong foundation for a very timely paper. However, I think more specificity is needed for the predictions based on anatomical targets. For example, the ventral and dorsal lateral PFC are thought to sub-serve quite discrete processes, both of which could be relevant for this study. I think discussing prior work, and the authors own work, with a greater deal of anatomical specificity would help readers better interpret that findings. A similar weakness was in the lack of discussion of where in the prior and work value-related differences emerged. For example there is a lot of discussion of prior work by Davidow and Shohamy, but those all occurred in ventral striatum which does not overlap with the current findings.

2. For the frequency learning task, it would be helpful to report on accuracy and RT for the 5th trial only. It would be extremely important to know if children and adults were entering the second phase of the experiment with the same acquired knowledge, or different amounts of acquired knowledge. If this 5th trial accuracy is different across age groups, the authors need to include this measure as a co-variate in all neural analyses, as age differences in biases towards high frequency information may result from not having learned the information rather than not being able to effectively use the information to guide adaptive encoding.

3. I think more discussion is warranted on why some neurobehavioral targets show quadratic effects while others show linear. Quadratic effects are often predicted in theoretical models, but from my read on the literature rarely show up in developmental analyses. These data could be leveraged to better understand those models by explaining why some processes are quadratic and others are linear.

4. The authors show directly compare the model fits of the mediation models that manipulate directionality.

5. I found the predictors from the frequency analysis predicting behavior during memory encoding/retrieval to perhaps be the most interesting finding in the paper, especially given that both implicit and explicit measures were predicting memory independently. However, then I was left wanting to know (somewhat desperately!), how much these signals related to the lateral PFC and caudate signals seen during memory encoding. I think this type of analysis would really help make the paper a complete package.

6. Regarding the discussion, I think it would be helpful for the authors to discuss a few features of their data and how they relate to development. The first would be the WASI findings which were quite prominent in most analyses, and in a few showed interactions with age. If this measure is a proxy of executive function, discussing the role of executive function for adaptive memory could help provide a more concrete mechanisms of adaptive memory formation across development. Additionally, it would be helpful for the authors to discuss the negative findings during retrieval. The fact that these age-related differences in adaptive memory processes are most likely seeming from encoding versus retrieval is not only highly interesting, but are a major prediction that stems from an animal literature on dopaminergic influences on hippocampal-dependent memory.

7. It would be helpful to report all behavioral results in the youngest sample only, as the interpretation of the data is quite different if children can or cannot perform these processes.

8. Could the authors provide data on differences in behavioral performance for both task content (which could drive motivational differences) or task order (which might lead to practice effects). If notable differences emerge across these factors, I strongly believe they needed to be included as co-variates in ally analyses.

Reviewer #3 (Recommendations for the authors):

1) Empirical findings directly comparing cross-sectional and longitudinal effects have demonstrated that cross-sectional analyses of age differences do not readily generalize to longitudinal research (e.g., Raz et al., 2005; Raz and Lindenberger, 2012). Formal analyses have demonstrated that proportion of explained age-related variance in cross-sectional mediation models may stem from various factors, including similar mean age trends, within-time correlations between a mediator and an outcome, or both (Lindenberger et al., 2011; see also Hofer, Flaherty, and Hoffman, 2006; Maxwell and Cole, 2007). Thus, the results of the mediation analysis showing that PFC activation explains age-related variance in memory difference scores, cannot be taken to imply that changes in PFC activation are correlated with changes in value-guided memory. While the general limitations of a cross-sectional study are noted in the Discussion of the manuscript, it would be important to discuss the critical limitations of the mediation analysis. While the main conclusions of the paper do not critically depend on this analysis, it would be important to alert the reader to the limited information value in performing cross-sectional mediation analyses of age variance.

2) It would be helpful to provide more information on how chance memory performance was handled during data analysis, especially as it is more likely to occur in younger participants. Related to this, please connect the points that belong to the same individual in Figure 3 to facilitate evaluation of individual differences in the memory difference scores.

3) I would like to see some consideration of how the different signatures of value learning, repetition suppression and reported item frequency, are related to the observed PFC and caudate effects during memory encoding. Such a discussion would help the reader connect the findings on learning and using information value across development.

4) A point worthy of discussion are the implications of the finding that younger participants demonstrated greater deviations in their frequency reports for the development of value learning, given that frequency reports were found to predict associative memory accuracy.

5) It would be helpful to include (supplementary) figures accompanying the behavioral results from the frequency learning phase.

6) Supplementary Figure S2 – providing an (additional) plot of the estimated age effects would help match the displayed results better to the description in the main text.

https://doi.org/10.7554/eLife.69796.sa1

Author response

Essential revisions:

All reviewers agreed that your manuscript presents important and interesting results suitable for publication in eLife. They generally liked the study and felt that the experimental design and data analysis approaches were clever and thorough. However, reviewers also raised several issues that we would like you to address in a revised version of the manuscript:

1. Do younger subjects not prioritize memory or do they not learn what information is valuable (e.g., R1, comment 3; R2, comment 2; R3, comment 2)?

Thank you for raising this important point. Indeed, one of our main findings is that older participants are better both at learning the structure of their environments and also at using structured knowledge to strategically prioritize memory. In our original manuscript, we described results of a model that included participants’ explicit frequency reports as a predictor of memory. Model comparison revealed that participants’ frequency reports — which we interpret as reflecting their beliefs about the structure of the environment — predicted memory more strongly than the item’s true frequency. In other words, participants’ beliefs about the structure of the environment (even if incorrect) more strongly influenced their memory encoding than the true structure of the environment. Critically, however, frequency reports interacted with age to predict memory (Figure 8). Even when we accounted for age-related differences in knowledge of the structure of the environment, older participants demonstrated a stronger influence of frequency on memory, suggesting they were better able to use their beliefs to control subsequent associative encoding. We have now clarified our interpretation of this model in our discussion on p. 23:

“Importantly, though we observed age-related differences in participants’ learning of the structure of their environment, the strengthening of the relation between frequency reports and associative memory with increasing age suggests that age differences in learning cannot fully account for age differences in value-guided memory. Even when accounting for individual differences in participants’ explicit knowledge of the structure of the environment, older participants demonstrated a stronger relation between their beliefs about item frequency and associative memory, suggesting that they used their beliefs to guide memory to a greater degree than younger participants.”

As noted by the reviewer, however, our initial memory analysis did not account for age-related differences in participants’ initial, online learning of item frequency, and our neural analyses further did not account for age differences in explicit frequency reports. We have now run additional control analyses to account for the potential influence of individual differences in frequency learning on associative memory. Specifically, for each participant, we computed three metrics: (1) their overall accuracy during frequency learning, (2) their overall accuracy for the last presentation of each item during frequency learning (as suggested by Reviewer 2), and (3) the mean magnitude of the error in their frequency reports. We then included these metrics as covariates in our memory analyses.

When we include these control variables in our model, we continue to observe a robust effect of frequency condition (p < 0.001) as well as robust interactions between frequency condition and linear and quadratic age (ps < 0.003) on associative memory accuracy. We also observed a main effect of frequency error magnitude on memory accuracy (p < 0.001). Here, however, we no longer observe main effects of age or quadratic age on overall memory accuracy. Given the relation we observed between frequency error magnitudes and age, the results from this model suggests that there may be age-related improvements in overall memory that influence both memory for associations as well as learning of and memory for item frequencies. The fact that age no longer relates to overall memory when controlling for frequency error magnitudes suggest that age-related variance in memory for item frequencies and memory for associations are strongly related within individuals. Importantly, however, age-related variance in memory for item frequencies did not explain age-related variance in the influence of frequency condition on associative memory, suggesting that there are developmental differences in the use of knowledge of environmental structure to prioritize valuable information in memory that persist even when controlling for age-related differences in initial learning of environmental regularities. Given the importance of this analysis in elucidating the relation between the learning of environmental structure and value-guided memory, we have now updated the results in the main text of our manuscript to include them. Specifically, on p. 13, we now write:

“Because we observed age-related differences in participants’ online learning of item frequencies and in their explicit frequency reports, we further examined whether these age differences in initial learning could account for the age differences we observed in associative memory. To do so, we ran an additional model in which we included each participant’s mean frequency learning accuracy, mean frequency learning accuracy on the last repetition of each item, and explicit report error magnitude as covariates. Here, explicit report error magnitude predicted overall memory performance, χ2(1) = 13.05, p < 0.001, and we did not observe main effects of age or quadratic age on memory performance (ps > 0.20). However, we continued to observe a main effect of frequency condition, χ2(1) = 19.65 p < 0.001, as well as significant interactions between frequency condition and both linear age χ2(1) = 10.59, p = 0.001, and quadratic age χ2(1) = 9.15, p = 0.002. Thus, while age differences in initial learning related to overall memory performance, they did not account for age differences in the use of environmental regularities to strategically prioritize memory for valuable information.”

In addition, as suggested by the reviewer, we also included the three covariates as control variables in our mediation analysis. When controlling for online frequency learning and explicit frequency report errors, PFC activity continued to mediate the relation between age and memory difference scores. We have now included these results on p. 16 – 17 of the main text:

“Further, when we included quadratic age, WASI scores, online frequency learning accuracy, online frequency learning accuracy on the final repetition of each item, and mean explicit frequency report error magnitudes as control variables in the mediation analysis, PFC activation continued to mediate the relation between linear age and memory difference scores (standardized indirect effect:.56, 95% confidence interval: [.06, 1.35], p = 0.023; standardized direct effect; 1.75, 95% confidence interval: [.12,.3.38], p = 0.034).”

We also refer to these analyses when we interpret our findings in our discussion. On p. 23, we write:

“In addition, we continued to observe a robust interaction between age and frequency condition on associative memory, even when controlling for age-related change in the accuracy of both online frequency learning and explicit frequency reports. Thus, though we observed age differences in the learning of environmental regularities and in their influence on subsequent associative memory encoding, our developmental memory effects cannot be fully explained by differences in initial learning.”

We thank the reviewer for this constructive suggestion, as we believe these control analyses strengthen our interpretation of age differences in both the learning and use of environmental regularities to prioritize memory.

2. Please address the modest relationship between the repetition suppression index in the parahippocampal cortex and age (e.g., R1, comment 1 and 2).

As recommended, we have now computed neural activation within our parahippocampal region of interest not just for the first and last appearance of each item during frequency learning, but for all appearances. Specifically we extended our repetition suppression analysis described in the manuscript to include all image repetitions (p. 36 – 37). Our new methods description reads:

“For each stimulus in the high-frequency condition, we examined repetition suppression by measuring activation within a parahippocampal ROI during the presentation of each item during frequency learning. We defined our ROI by taking the peak voxel (x = 30, y = -39, z = -15) from the group-level first > last item appearance contrast for high-frequency items during frequency learning and drawing a 5 mm sphere around it. This voxel was located in the right parahippocampal cortex, though we observed widespread and largely symmetric activation in bilateral parahippocampal cortex. To encompass both left and right parahippocampal cortex within our ROI, we mirrored the peak voxel sphere. For each participant, we modeled the neural response to each appearance of each item using the Least Squares-Separate approach (Mumford et al., 2014). Each first-level model included a regressor for the trial of interest, as well as separate regressors for the onsets of all other items, grouped by repetition number (e.g., a regressor for item onsets on their first appearance, a regressor for item onsets on their second appearance, etc.). Values that fell outside five standard deviations from the mean level of neural activation across all subjects and repetitions were excluded from subsequent analyses (18 out of 10,320 values; .01% of observations). In addition to examining neural activation as a function of stimulus repetition, we also computed an index of repetition suppression for each high-frequency item by computing the difference in mean β values within our ROI on its first and last appearance.”

As suggested, we ran a mixed effects model examining the influence of linear and quadratic age and linear and quadratic repetition number on neural activation. In line with our whole-brain analysis, we observed a robust effect of linear and quadratic repetition number, suggesting that neural activation decreased non-linearly across stimulus repetitions. In addition, we observed significant interactions between our age and repetition number terms, suggesting that repetition suppression increased into early adulthood. Thus, although the relation we observed between age and repetition suppression is modest, the results from our new analyses suggest it is robust. Because these results largely aligned with the pattern of age-related change we observed in our analysis of repetition suppression indices, we continued to use that compressed metric in subsequent analyses looking at relations with behavior. However, we have updated our Results section to include the full analysis taking into account all item repetitions, as suggested. Our updated manuscript now reads (p. 9):

“We next examined whether repetition suppression in the parahippocampal cortex changed with age. We defined a parahippocampal region of interest (ROI) by drawing a 5mm sphere around the peak voxel from the group-level first > last appearance contrast (x = 30, y = -39, z = -15), and mirrored it to encompass both right and left parahippocampal cortex (Figure 2C). For each participant, we modeled the neural response to each appearance of each high-frequency item. We then examined how neural activation changed as a function of repetition number and age. To account for non-linear effects of repetition number, we included linear and quadratic repetition number terms. In line with our whole-brain analysis, we observed a main effect of repetition number, F(1, 5016.0) = 30.64, p < 0.001, indicating that neural activation within the parahippocampal ROI decreased across repetitions. Further, we observed a main effect of quadratic repetition number, F(1, 9881.0) = 7.47, p = 0.006, indicating that the reduction in neural activity was greatest across earlier repetitions (Figure 3A). Importantly, the influence of repetition number on neural activation varied with both linear age, F(1, 7267.5) = 7.2, p = 0.007 and quadratic age , F(1, 7260.8) = 6.9, p = 0.009. Finally, we also observed interactions between quadratic repetition number and both linear and quadratic age (ps < 0.026). These age-related differences suggest that repetition suppression was greatest in adulthood, with the steepest increases occurring from late adolescence to early adulthood (Figure 3).”

For each participant for each item, we also computed a “repetition suppression index” by taking the difference in mean β values within our ROI on each item’s first and last appearance (Ward et al., 2013). These indices demonstrated a similar pattern of age-related variance — we found that the reduction of neural activity from the first to last appearance of the items varied positively with linear age, F(1, 78.32) = 3.97, p = 0.05, and negatively with quadratic age, F(1, 77.55) = 4.8, p = 0.031 (Figure 3B). Taken together, our behavioral and neural results suggest that sensitivity to the repetition of items in the environment was prevalent from childhood to adulthood but increased with age.”

In addition, in the main text on p. 10, we have now included the suggested scatter plot (see new Figure 3B) as well as a modified version of our previous figure S2 to show neural activation across all repetitions in the parahippocampal cortex (see new Figure 3A). We thank the reviewer for this helpful suggestion, as we believe these new figures much more clearly illustrate the repetition suppression effects we observed during frequency learning.

3. It would be important to be more specific regarding anatomical subregions in the introduction and discussion (R2, comment 1).

We agree with the reviewer that our introduction and discussion would benefit from more anatomical granularity, and we did indeed have a priori predictions about more specific neural regions that might be involved in our task.

First, we expected that both the ventral and dorsal striatum might be responsive to stimulus value across our age range. Prior work has suggested that activity in the ventral striatum often correlates with the intrinsic value of a stimulus, whereas activity in the dorsal striatum may reflect goal-directed action values (Liljeholm and O’Doherty, 2012). In our task, we expected that high-frequency items may acquire intrinsic value during frequency learning that is then reflected in the striatal response to these items during encoding. However, because participants were not rewarded when they encountered these images, but rather incentivized to encode associations involving them, we hypothesized that the dorsal striatum may represent the value of the ‘action’ of remembering each pair. In line with this prediction, the dorsal striatum, and the caudate in particular, have also been shown to be engaged during value-guided cognitive control (Hikosaka et al., 2014; Insel et al., 2017).

We have now revised our introduction to include greater specificity in our anatomical predictions on p. 3:

“When individuals need to remember information associated with previously encountered stimuli (e.g., the grocery store aisle where an ingredient is located), frequency knowledge may be instantiated as value signals, engaging regions along the mesolimbic dopamine pathway that have been implicated in reward anticipation and the encoding of stimulus and action values. These areas include the ventral tegmental area (VTA) and the ventral and dorsal striatum (Adcock et al., 2006; Liljeholm and O’Doherty, 2012; Shigemune et al., 2014).”

Though we initially predicted that encoding of high-value information would be associated with increased activation in both the ventral and dorsal striatum, the activation we observed was largely within the dorsal striatum, and specifically, the caudate. We have now revised our discussion accordingly on p. 26:

“Though we initially hypothesized that both the ventral and dorsal striatum may be involved in encoding of high-value information, the activation we observed was largely within the dorsal striatum, a region that may reflect the value of goal-directed actions (Liljeholm and O’Doherty, 2012). In our task, rather than each stimulus acquiring intrinsic value during frequency learning, participants may have represented the value of the ‘action’ of remembering each pair during encoding.”

Second, while the ventromedial PFC often reflects value, given the control demands of our task, we expected to see greater activity in the dorsolateral PFC, which is often engaged in tasks that require the implementation of cognitive control (Botvinick and Braver, 2015). Thus, we hypothesized that individuals would show increased activation in the dlPFC during encoding of high- vs. low-value information, and that this activation would vary as a function of age. We have now clarified this hypothesis on p. 3:

“Value responses in the striatum may signal the need for increased engagement of the dorsolateral prefrontal cortex (dlPFC) (Botvinick and Braver, 2015), which supports the implementation of strategic control.”

In our discussion, we review disparate findings in the developmental literature and discuss factors that may contribute to these differences across studies. For example, in our discussion of Davidow et al., (2016), we highlight differences between their task design and the present study, focusing on how their task involved immediate receipt of reward at the time of encoding, while our task incentivized memory accuracy. We further note that studies that involve reward delivery at the time of encoding may engage different neural pathways than those that promote goal-directed encoding. Beyond Davidow et al., (2016), there are no other neuroimaging studies that examine the influence of reward on memory across development. Thus, we cannot relate our present neural findings to prior work on the development of value-guided memory. As we note in our discussion (p. 28), “Further work is needed to characterize both the influence of different types of reward signals on memory across development, as well as the development of the neural pathways that underlie age-related change in behavior.”

4. It would be interesting to connect the neural measures during initial frequency learning and memory-guided adaptive encoding (e.g., R2, comment 5).

Thank you for this valuable suggestion. We agree that it would be interesting to link frequency-learning behavior to neural activity at encoding. As such, we have now conducted additional analyses to explore these relations.

In the original version of our manuscript, we examined behavior at the item level through mixed-effects models, and neural activation during encoding at the participant level. Thus, to examine the relation between frequency-learning metrics and neural activation at encoding, we created two additional participant-level metrics. For each participant we computed their average repetition suppression index, and a measure of frequency distance. The average repetition suppression index reflects the overall extent to which the participant demonstrated repetition suppression in response to the fifth presentation of the high-frequency items, and is computed by averaging each participant’s repetition suppression indices across items. We hypothesized that participants who demonstrated the greatest degree of repetition suppression might be the most sensitive to the difference between the 1- and 5-frequency items, and therefore, show the greatest differences in striatal and PFC activation during encoding of high- vs. low-value information. The frequency distance metric reflects the average distance between participants’ explicit frequency reports for items that appeared once and items that appeared five times, and is computed by averaging their explicit frequency reports for items in each frequency condition, and then subtracting the average reports in the low-frequency condition from those in the high-frequency condition. We hypothesized that participants with the largest frequency distances might similarly be the most sensitive to the difference between the 1- and 5-frequency items, and therefore, show the greatest differences in striatal and PFC activation during encoding of high- vs. low-value information.

We first wanted to confirm that the relations we observed between repetition suppression, frequency reports, and age, could also be observed at the participant level. In line with our prior, behavioral analyses, we found that age related to both mean repetition suppression indices (marginally; linear age: p = 0.067; quadratic age: p = 0.042); and frequency distances (linear and quadratic age: ps < 0.001).

In addition, we further tested whether these two metrics related to memory performance. In contrast to our item-level findings, we did not observe a significant relation between repetition suppression indices and memory (p = 0.83). We did observe an effect of frequency distance on memory performance. Specifically, we observed significant interactions between frequency distance and age (p = 0.014) and frequency distance and quadratic age (p = 0.021) on memory difference scores, such that the influence of frequency distance on memory difference scores increased with increasing age from childhood to adolescence.

We next examined how mean repetition suppression indices and frequency distances related to differential neural activation during encoding of high- and low-value pairs. In line with our memory findings, we did not observe any significant relations between mean repetition suppression indices and neural activation in the caudate or prefrontal cortex during encoding (ps > 0.15).

Frequency distance did not relate to caudate activation during encoding nor did we observe a frequency distance x age interaction effect (ps > 0.16). Frequency distance did, however, relate to differential PFC activation during encoding of high- vs. low-value pairs. Specifically, we observed a main effect of frequency distance on PFC activation (p = 0.0012), such that participants whose explicit reports of item frequency, were on average, more distinct across frequency conditions, demonstrated increased PFC activation during encoding of pairs involving high- vs. low-frequency items. Interestingly, when we included frequency distance in our model, we no longer observed a significant effect of age on differential PFC activation, nor did we observe a significant frequency distance x age interaction (ps > 0.13). These findings suggest that PFC activation during encoding may have, in part, reflected participants’ beliefs about the structure of the environment, with participants demonstrating stronger differential engagement of control processes across conditions when their representations of the conditions themselves were more distinct.

Finally, we examined how age, frequency distance, and PFC activation related to memory difference scores. Here, even when controlling for both frequency distance and PFC activation, we continued to observe main effects of age and quadratic age on memory difference scores (linear age: p = 0.006; quadratic age: p = 0.001). In line with our analysis of the relation between frequency reports and memory, these results suggest that age-related variance in value-guided memory may depend on both knowledge of the structure of the environment and use of that knowledge to effectively control encoding.

We have now added these results to our manuscript on p. 13 – 14. We write:

“Given the relations we observed between memory and both repetition suppression and frequency reports, we examined whether they related to neural activation in both our caudate and PFC ROI during encoding. […] Importantly, however, even when we accounted for both PFC activation and frequency distances, we continued to observe an effect of age on memory difference scores (β = 0.56, SE = 0.20, p = 0.006), which, together with our prior analyses, suggest that developmental differences in value-guided memory are not driven solely by age differences in beliefs about the structure of the environment but also depend on the use of those beliefs to guide encoding.”

We have added the full model results to Appendix 3.

Given these results, we have now revised our interpretation of our neural data. Our memory analyses demonstrate that across our age range, we observed age-related differences in both the acquisition of knowledge of the structure of the environment and in its use. Originally, we interpreted the PFC activation as reflecting the use of learned value to guide memory. However, the strong relation we found between frequency distance and PFC activation suggests that the age differences in PFC activation that we observed may also be related to age differences in knowledge of the structure of the environment that governs when control processes should be engaged most strongly. However, these results must be interpreted cautiously. Participants provided explicit frequency reports after they completed the encoding and retrieval tasks, and so explicit frequency reports may have been influenced not only by participants’ memories of online frequency learning, but also by the strength with which they encoded the item and its paired associate, and the experience of successfully retrieving it.

We have now revised our discussion to consider these results. On p. 23, we now write,

“Our neural results further suggest that developmental differences in memory were driven by both knowledge of the structure of the environment and use of that knowledge to guide encoding.”

On p. 24, we write,

“The development of adaptive memory requires not only the implementation of encoding and retrieval strategies, but also the flexibility to up- or down-regulate the engagement of control in response to momentary fluctuations in information value (Castel et al., 2007, 2013; Hennessee et al., 2017). Importantly, value-based modulation of lateral PFC engagement during encoding mediated the relation between age and memory selectivity, suggesting that developmental change in both the representation of learned value and value-guided cognitive control may underpin the emergence of adaptive memory prioritization. Prior work examining other neurocognitive processes, including response inhibition (Insel et al., 2017) and selective attention (Störmer et al., 2014), has similarly found that increases in the flexible upregulation of control in response to value cues enhance goal-directed behavior across development (Davidow et al., 2018), and may depend on the engagement of both striatal and prefrontal circuitry (Hallquist et al., 2018; Insel et al., 2017). Here, we extend these past findings to the domain of memory, demonstrating that value signals derived from the structure of the environment increasingly elicit prefrontal cortex engagement and strengthen goal-directed encoding across childhood and into adolescence.”

And on p. 25, we have added an additional paragraph:

“Further, we also demonstrate that in the absence of explicit value cues, the engagement of prefrontal control processes may reflect beliefs about information value that are learned through experience. Here, we found that differential PFC activation during encoding of high- vs. low-value information reflected individual and age-related differences in beliefs about the structure of the environment; participants who represented the average frequencies of the low- and high-frequency items as further apart also demonstrated greater value-based modulation of lateral PFC activation. It is important to note, however, that we collected explicit frequency reports after associative encoding and retrieval. Thus the relation between PFC activation and explicit frequency reports may be bidirectional — while participants may have increased the recruitment of cognitive control processes to better encode information they believed was more valuable, the engagement of more elaborative or deeper encoding strategies that led to stronger memory traces may have also increased participants’ subjective sense of an item’s frequency (Jonides and Naveh-Benjamin, 1987).”

5. Please discuss problems with cross-sectional designs, especially as they relate to mediation analyses (e.g., R3, comment 1).

Thank you for raising this critical point. We have expanded our discussion to specifically note the limitations of our mediation analysis and to more strongly emphasize the need for future longitudinal studies to reveal how changes in neural circuitry may support the emergence of motivated memory across development. Specifically, on p. 26, we now write:

“One important caveat is that our study was cross-sectional — it will be important to replicate our findings in a longitudinal sample to more directly measure how developmental changes in cognitive control within an individual contribute to changes in their ability to selectively encode useful information. Our mediation results, in particular, must be interpreted with caution as simulations have demonstrated that in cross-sectional samples, variables can emerge as significant mediators of age-related change due largely to statistical artifact (Hofer, Flaherty, and Hoffman, 2006; Lindenberger et al., 2011). Indeed, our finding that PFC activation mediates the relation between age and value-guided memory does not necessarily imply that within an individual, PFC development leads to improvements in memory selectivity. Longitudinal work in which individuals’ neural activity and memory performance is sampled densely within developmental windows of interest is needed to elucidate the complex relations between age, brain development, and behavior (Hofer, Flaherty, and Hoffman, 2006; Lindenberger et al., 2011).”

6. Please discuss the implications of the quadratic effects (e.g., R2, comment 3).

We agree with the reviewer that more discussion is warranted here. While many cognitive processes tend to improve with increasing age, the significant interaction between quadratic age and frequency condition on memory accuracy could reflect a number of different patterns of developmental variance. Because quadratic curves are U-shaped, the significant interaction between quadratic age and frequency condition could reflect a peak in value-guided memory in adolescence. However, the combination of linear and quadratic effects can also capture “plateauing” effects, where the influence of age on a particular cognitive process decreases at a particular developmental timepoint. To determine how to interpret the quadratic effect of age on value-guided memory — and specifically, to test for the presence of an adolescent peak — we ran an additional analysis.

To test for an adolescent peak in value-guided memory, we first fit our memory accuracy model without any age terms, and then extracted the random slope across frequency conditions for each subject. We then conducted a ‘two lines test’ (Simonsohn, 2018) to examine the relation between age and these random slopes. In brief, the two-lines test fits the data with two linear models — one with a positive slope and one with a negative slope, algorithmically determining the breakpoint in the estimates where the signs of the slopes change. When we analyzed our memory data in this way, we found a robust, positive relation between age and value-guided memory (see newly added Appendix 2 – Figure 3) from childhood to mid-adolescence, that peaked around age 16 (age 15.86). From age ~16 to early adulthood, however, we observed only a marginal negative relation between age and value-guided memory (p = 0.0567). Thus, our findings do not offer strong evidence in support of an adolescent peak in value-guided memory — instead, they suggest that improvements in value-guided memory are strongest from childhood to adolescence.

To more clearly demonstrate the relation between age and value-guided memory, we have now included the results of the two-lines test in the Results section of our main text. On p. 12 – 13, we write:

“In line with our hypothesis, we observed a main effect of frequency condition on memory, χ2(1) = 21.51, p < 0.001, indicating that individuals used naturalistic value signals to prioritize memory for high-value information. Critically, this effect interacted with both linear age (χ2(1) = 11.03, p < 0.001) and quadratic age (χ2(1) = 9.51, p = 0.002), such that the influence of frequency condition on memory increased to the greatest extent throughout childhood and early adolescence.

To determine whether the interaction between quadratic age and frequency condition on memory accuracy reflected an adolescent peak in value-guided memory prioritization, we re-ran our memory accuracy model without including any age terms, and extracted each participant’s random slope across frequency conditions. We then submitted these random slopes to the “two-lines” test (Simonsohn, 2018), which fits two regression lines with oppositely signed slopes to the data, algorithmically determining where the sign flip should occur. The results of this analysis revealed that the influence of frequency condition on memory significantly increased from age 8 to age 15.86 (b = 0.03, z = 2.71, p = 0.0068; Appendix 2 – Figure 3), but only marginally decreased from age 15.86 to age 25 (b = -.02, z = 1.91, p = 0.0576). Thus, the interaction between frequency condition and quadratic age on memory performance suggests that the biggest age differences in value-guided memory occurred through childhood and early adolescence, with older adolescents and adults performing similarly.”

That said, this developmental trajectory is likely specific to the particular demands of our task. In our previous behavioral study that used a very similar paradigm (Nussenbaum, Prentis, and Hartley, 2020), we observed only a linear relation between age and value-guided memory. Although the task used in our behavioral study was largely similar to the task we employed here, there were subtle differences in the design that may have extended the age range through which we observed improvements in memory prioritization. In particular, in our previous behavioral study, the memory test required participants to select the correct associate from a grid of 20 options (i.e., 1 correct and 19 incorrect options), whereas here, participants had to select the correct associate from a grid of 4 options (1 correct and 3 incorrect options). In our prior work, the need to differentiate the ‘correct’ option from many more foils may have increased the demands on either (or both) memory encoding or memory retrieval, requiring participants to encode and retrieve more specific representations that would be less confusable with other memory representations. By decreasing the task demands in the present study, we may have shifted the developmental curve we observed toward earlier developmental timepoints.

We originally did not emphasize our quadratic findings in the discussion of our manuscript because, given the marginal decrease in memory selectivity we observed from age 16 to age 25 and the different age-related findings across our two studies, we did not want to make strong claims about the specific shape of developmental change. However, we agree with the reviewer that these points are worthy of discussion within the manuscript. We have now amended our discussion on p. 25 accordingly:

“We found that memory prioritization varied with quadratic age, and our follow-up tests probing the quadratic age effect did not reveal evidence for significant age-related change in memory prioritization between late adolescence and early adulthood. However, in our prior behavioral work using a very similar paradigm (Nussenbaum et al., 2020), we found that memory prioritization varied with linear age only. In line with theoretical proposals (Davidow et al., 2018), subtle differences in the control demands between the two tasks (e.g., reducing the number of ‘foils’ presented on each trial of the memory test here relative to our prior study), may have shifted the age range across which we observed differences in behavior, with the more demanding variant of our task showing more linear age-related improvements into early adulthood. In addition, the specific control demands of our task may have also influenced the age at which value-guided memory emerged. Future studies should test whether younger children can modulate encoding based on the value of information if the mnemonic demands of the task are simpler.”

We thank the reviewer for this helpful suggestion, and believe our additions that expand on the quadratic age effects help clarify our developmental findings.

We also encourage you to consider the other suggestions made by individual reviewers.

We appreciate the thoughtful comments and suggestions from the reviewers and have addressed all of them, in turn, below. Given the overlap across many of the reviewer comments, some of our responses are repeated where relevant.

Reviewer #1 (Recommendations for the authors):

1. Given the modest relationship between the repetition suppression index in the parahippocampal cortex and age, it would be important to make sure this relationship is robust. One way would be to use more data to improve the estimation of the repetition suppression index. For instance, data from all presentations could be used to estimate linear and quadratic effects of repetition number. This would not only improve the reliability of the estimates, but also offer the opportunity to examine whether the dynamics of frequency learning across multiple presentations changes with age.

Thank you for this helpful suggestion. As recommended, we have now computed neural activation within our parahippocampal region of interest not just for the first and last appearance of each item during frequency learning, but for all appearances. Specifically we extended our repetition suppression analysis described in the manuscript to include all image repetitions (p. 36 – 37). Our new methods description reads:

“For each stimulus in the high-frequency condition, we examined repetition suppression by measuring activation within a parahippocampal ROI during the presentation of each item during frequency learning. […] In addition to examining neural activation as a function of stimulus repetition, we also computed an index of repetition suppression for each high-frequency item by computing the difference in mean β values within our ROI on its first and last appearance.”

As suggested, we ran a mixed effects model examining the influence of linear and quadratic age and linear and quadratic repetition number on neural activation. In line with our whole-brain analysis, we observed a robust effect of linear and quadratic repetition number, suggesting that neural activation decreased non-linearly across stimulus repetitions. In addition, we observed significant interactions between our age and repetition number terms, suggesting that repetition suppression increased into early adulthood. Thus, although the relation we observed between age and repetition suppression is modest, the results from our new analyses suggest it is robust. Because these results largely aligned with the pattern of age-related change we observed in our analysis of repetition suppression indices, we continued to use that compressed metric in subsequent analyses looking at relations with behavior. However, we have updated our Results section to include the full analysis taking into account all item repetitions, as suggested. Our updated manuscript now reads (p. 9):

“We next examined whether repetition suppression in the parahippocampal cortex changed with age. […] These age-related differences suggest that repetition suppression was greatest in adulthood, with the steepest increases occurring from late adolescence to early adulthood (Figure 3).”

For each participant for each item, we also computed a “repetition suppression index” by taking the difference in mean β values within our ROI on each item’s first and last appearance (Ward et al., 2013). These indices demonstrated a similar pattern of age-related variance — we found that the reduction of neural activity from the first to last appearance of the items varied positively with linear age, F(1, 78.32) = 3.97, p = 0.05, and negatively with quadratic age, F(1, 77.55) = 4.8, p = 0.031 (Figure 3B). Taken together, our behavioral and neural results suggest that sensitivity to the repetition of items in the environment was prevalent from childhood to adulthood but increased with age.”

2. At the very least, it would be important to include a scatter plot of repetition index versus age in the main text, and to show the parameter estimates across all repetitions in the parahippocampal cortex (e.g., per age group).

In the main text on p. 10, we have now included the suggested scatter plot (see new Figure 3B) as well as a modified version of our previous figure S2 to show neural activation across all repetitions in the parahippocampal cortex (see new Figure 3A). We thank the reviewer for this helpful suggestion, as we believe these new figures much more clearly illustrate the repetition suppression effects we observed during frequency learning.

3. Given the effects of age on both initial frequency learning (new vs. old response accuracy and frequency estimation error) and the effects of item frequency on memory, it would be important to control for frequency learning when testing the effects of age on memory prioritization, and whether lateral PFC mediates this relationship.

Thank you for raising this important point. Indeed, one of our main findings is that older participants are better both at learning the structure of their environments and also at using structured knowledge to strategically prioritize memory. In our original manuscript, we described results of a model that included participants’ explicit frequency reports as a predictor of memory. Model comparison revealed that participants’ frequency reports — which we interpret as reflecting their beliefs about the structure of the environment — predicted memory more strongly than the item’s true frequency. In other words, participants’ beliefs about the structure of the environment (even if incorrect) more strongly influenced their memory encoding than the true structure of the environment. Critically, however, frequency reports interacted with age to predict memory (Figure 8). Even when we accounted for age-related differences in knowledge of the structure of the environment, older participants demonstrated a stronger influence of frequency on memory, suggesting they were better able to use their beliefs to control subsequent associative encoding. We have now clarified our interpretation of this model in our discussion on p. 23:

“Importantly, though we observed age-related differences in participants’ learning of the structure of their environment, the strengthening of the relation between frequency reports and associative memory with increasing age suggests that age differences in learning cannot fully account for age differences in value-guided memory. Even when accounting for individual differences in participants’ explicit knowledge of the structure of the environment, older participants demonstrated a stronger relation between their beliefs about item frequency and associative memory, suggesting that they used their beliefs to guide memory to a greater degree than younger participants.”

As noted by the reviewer, however, our initial memory analysis did not account for age-related differences in participants’ initial, online learning of item frequency, and our neural analyses further did not account for age differences in explicit frequency reports. We have now run additional control analyses to account for the potential influence of individual differences in frequency learning on associative memory. Specifically, for each participant, we computed three metrics: (1) their overall accuracy during frequency learning, (2) their overall accuracy for the last presentation of each item during frequency learning (as suggested by Reviewer 2), and (3) the mean magnitude of the error in their frequency reports. We then included these metrics as covariates in our memory analyses.

When we include these control variables in our model, we continue to observe a robust effect of frequency condition (p < 0.001) as well as robust interactions between frequency condition and linear and quadratic age (ps < 0.003) on associative memory accuracy. We also observed a main effect of frequency error magnitude on memory accuracy (p < 0.001). Here, however, we no longer observe main effects of age or quadratic age on overall memory accuracy. Given the relation we observed between frequency error magnitudes and age, the results from this model suggests that there may be age-related improvements in overall memory that influence both memory for associations as well as learning of and memory for item frequencies. The fact that age no longer relates to overall memory when controlling for frequency error magnitudes suggest that age-related variance in memory for item frequencies and memory for associations are strongly related within individuals. Importantly, however, age-related variance in memory for item frequencies did not explain age-related variance in the influence of frequency condition on associative memory, suggesting that there are developmental differences in the use of knowledge of environmental structure to prioritize valuable information in memory that persist even when controlling for age-related differences in initial learning of environmental regularities. Given the importance of this analysis in elucidating the relation between the learning of environmental structure and value-guided memory, we have now updated the results in the main text of our manuscript to include them. Specifically, on p. 13, we now write:

“Because we observed age-related differences in participants’ online learning of item frequencies and in their explicit frequency reports, we further examined whether these age differences in initial learning could account for the age differences we observed in associative memory. To do so, we ran an additional model in which we included each participant’s mean frequency learning accuracy, mean frequency learning accuracy on the last repetition of each item, and explicit report error magnitude as covariates. Here, explicit report error magnitude predicted overall memory performance, χ2(1) = 13.05, p < 0.001, and we did not observe main effects of age or quadratic age on memory performance (ps > 0.20). However, we continued to observe a main effect of frequency condition, χ2(1) = 19.65 p < 0.001, as well as significant interactions between frequency condition and both linear age χ2(1) = 10.59, p = 0.001, and quadratic age χ2(1) = 9.15, p = 0.002. Thus, while age differences in initial learning related to overall memory performance, they did not account for age differences in the use of environmental regularities to strategically prioritize memory for valuable information.”

In addition, as suggested by the reviewer, we also included the three covariates as control variables in our mediation analysis. When controlling for online frequency learning and explicit frequency report errors, PFC activity continued to mediate the relation between age and memory difference scores. We have now included these results on p. 16 – 17 of the main text:

“Further, when we included quadratic age, WASI scores, online frequency learning accuracy, online frequency learning accuracy on the final repetition of each item, and mean explicit frequency report error magnitudes as control variables in the mediation analysis, PFC activation continued to mediate the relation between linear age and memory difference scores (standardized indirect effect:.56, 95% confidence interval: [.06, 1.35], p = 0.023; standardized direct effect: 1.75, 95% confidence interval: [.12,.3.38], p = 0.034).”

We also refer to these analyses when we interpret our findings in our discussion. On p. 23, we write:

“In addition, we continued to observe a robust interaction between age and frequency condition on associative memory, even when controlling for age-related change in the accuracy of both online frequency learning and explicit frequency reports. Thus, though we observed age differences in the learning of environmental regularities and in their influence on subsequent associative memory encoding, our developmental memory effects cannot be fully explained by differences in initial learning.”

We thank the reviewer for this constructive suggestion, as we believe these control analyses strengthen our interpretation of age differences in both the learning and use of environmental regularities to prioritize memory.

4. Please include more information on how the four echo time series were combined. What is considered optimal in this regard?

As noted in the manuscript, we preprocessed our data with fMRIprep, which uses the tedana T2* pipeline (Kundu et al., 2011; Kundu et al., 2013; Kundu et al., 2017) to combine the four echoes. Images acquired at longer delays after the excitation pulse (longer echo times) have higher signal dropout but greater BOLD sensitivity. Thus the ‘optimal’ combination of echoes takes a weighted average of the four echoes that balances signal strength and sensitivity for each voxel. More specifically, tedana first fits a model to estimate both the total signal in each voxel before decay as well as the rate at which the signal in each voxel decays over time (Kundu et al., 2017). Then, using the estimate for the rate of signal decay, tedana combines the signals across the four echoes using a weighted average, where the “weight” of each echo is determined by:

TE*e(TE/T2*)

where TE is the echo time and T2* is the rate of signal decay.

By weighting the echoes in this way, the combined data (which we use for all analyses) takes advantage of the signal strength of the earlier echoes and the sensitivity of the later echoes. Because signal decay is modeled separately for each voxel, this method of combining echoes enables differential weighting of images acquired at shorter and longer TEs for different regions of the brain. Our description of the combination of the echoes as “optimal” is meant to reflect the fact that the echoes are weighted differently for different voxels, depending on the rate at which the signal within them decays. We have now clarified this in the manuscript on p. 35, and emphasized to the reader where detailed information on the implementation of the multi-echo combination procedure can be obtained. We have also added text detailing the benefits of the use of multi-echo sequences. Specifically, we now write:

“FMRIPrep uses tedana (for implementation details, see Kundu et al., 2013, 2012) to combine each four-echo time series based on the signal decay rate of each voxel, taking a weighted average of the four echoes that optimally balances signal strength and BOLD sensitivity. This approach enables the acquisition of BOLD data with a higher signal-to-noise ratio, giving us greater sensitivity to detect neural effects of interest (Kundu et al., 2013).”

Reviewer #2 (Recommendations for the authors):

Before I start critiquing individual components of the study, I want to say I am quite excited about these results, and believe this a very clever design and interesting neuroimaging findings.

1. I found the introduction quite nice, and it really provided a strong foundation for a very timely paper. However, I think more specificity is needed for the predictions based on anatomical targets. For example, the ventral and dorsal lateral PFC are thought to sub-serve quite discrete processes, both of which could be relevant for this study. I think discussing prior work, and the authors own work, with a greater deal of anatomical specificity would help readers better interpret that findings. A similar weakness was in the lack of discussion of where in the prior and work value-related differences emerged. For example there is a lot of discussion of prior work by Davidow and Shohamy, but those all occurred in ventral striatum which does not overlap with the current findings.

We agree with the reviewer that our introduction and discussion would benefit from more anatomical granularity, and we did indeed have a priori predictions about more specific neural regions that might be involved in our task.

First, we expected that both the ventral and dorsal striatum might be responsive to stimulus value across our age range. Prior work has suggested that activity in the ventral striatum often correlates with the intrinsic value of a stimulus, whereas activity in the dorsal striatum may reflect goal-directed action values (Liljeholm and O’Doherty, 2012). In our task, we expected that high-frequency items may acquire intrinsic value during frequency learning that is then reflected in the striatal response to these items during encoding. However, because participants were not rewarded when they encountered these images, but rather incentivized to encode associations involving them, we hypothesized that the dorsal striatum may represent the value of the ‘action’ of remembering each pair. In line with this prediction, the dorsal striatum, and the caudate in particular, have also been shown to be engaged during value-guided cognitive control (Hikosaka et al., 2014; Insel et al., 2017).

We have now revised our introduction to include greater specificity in our anatomical predictions on p. 3:

“When individuals need to remember information associated with previously encountered stimuli (e.g., the grocery store aisle where an ingredient is located), frequency knowledge may be instantiated as value signals, engaging regions along the mesolimbic dopamine pathway that have been implicated in reward anticipation and the encoding of stimulus and action values. These areas include the ventral tegmental area (VTA) and the ventral and dorsal striatum (Adcock et al., 2006; Liljeholm and O’Doherty, 2012; Shigemune et al., 2014).”

Though we initially predicted that encoding of high-value information would be associated with increased activation in both the ventral and dorsal striatum, the activation we observed was largely within the dorsal striatum, and specifically, the caudate. We have now revised our discussion accordingly on p. 26:

“Though we initially hypothesized that both the ventral and dorsal striatum may be involved in encoding of high-value information, the activation we observed was largely within the dorsal striatum, a region that may reflect the value of goal-directed actions (Liljeholm and O’Doherty, 2012). In our task, rather than each stimulus acquiring intrinsic value during frequency learning, participants may have represented the value of the ‘action’ of remembering each pair during encoding.”

Second, while the ventromedial PFC often reflects value, given the control demands of our task, we expected to see greater activity in the dorsolateral PFC, which is often engaged in tasks that require the implementation of cognitive control (Botvinick and Braver, 2015). Thus, we hypothesized that individuals would show increased activation in the dlPFC during encoding of high- vs. low-value information, and that this activation would vary as a function of age. We have now clarified this hypothesis on p. 3:

“Value responses in the striatum may signal the need for increased engagement of the dorsolateral prefrontal cortex (dlPFC) (Botvinick and Braver, 2015), which supports the implementation of strategic control.”

In our discussion, we review disparate findings in the developmental literature and discuss factors that may contribute to these differences across studies. For example, in our discussion of Davidow et al. (2016), we highlight differences between their task design and the present study, focusing on how their task involved immediate receipt of reward at the time of encoding, while our task incentivized memory accuracy. We further note that studies that involve reward delivery at the time of encoding may engage different neural pathways than those that promote goal-directed encoding. Beyond Davidow et al., (2016), there are no other neuroimaging studies that examine the influence of reward on memory across development. Thus, we cannot relate our present neural findings to prior work on the development of value-guided memory. As we note in our discussion (p. 28), “Further work is needed to characterize both the influence of different types of reward signals on memory across development, as well as the development of the neural pathways that underlie age-related change in behavior.“

2. For the frequency learning task, it would be helpful to report on accuracy and RT for the 5th trial only. It would be extremely important to know if children and adults were entering the second phase of the experiment with the same acquired knowledge, or different amounts of acquired knowledge. If this 5th trial accuracy is different across age groups, the authors need to include this measure as a co-variate in all neural analyses, as age differences in biases towards high frequency information may result from not having learned the information rather than not being able to effectively use the information to guide adaptive encoding.

We continued to observe age differences in frequency-learning accuracy and reaction times (RTs) on the fifth and final presentation of each image (see newly added Figure S2). While age differences in RTs may be reflective of processing or motor response speed, we agree with the reviewer that age differences in accuracy likely indicate that children and adults did not necessarily acquire equivalent knowledge of environmental structure prior to encoding associated information.

Indeed, in our original manuscript, we included a model with participants’ frequency reports as a predictor of memory performance rather than each item’s true frequency. Not only did participant frequency reports predict associative memory, model comparison revealed that they were actually a better predictor of associative memory than each item’s true frequency. Importantly, we observed a frequency report x age interaction effect, indicating that older participants’ demonstrated greater modulation of memory by their beliefs about the structure of the environment. In other words, even when we control for age differences in beliefs about the structure of the environment, older participants continued to demonstrate greater use of beliefs about environmental structure to guide memory. Thus, taken together with our frequency-learning results, we believe results from this model demonstrate that there are age differences both in learning the structure of the environment, and in using learned regularities to guide encoding. Critically, age differences in using these learned regularities emerge even when controlling for age differences in learning. We have now clarified and expanded our interpretation of this model in our manuscript discussion on p. 23:

“Importantly, though we observed age-related differences in participants’ learning of the structure of their environment, the strengthening of the relation between frequency report and associative memory with increasing age suggests that age differences in learning cannot fully account for age differences in value guided memory. Even when accounting for individual differences in participants’ explicit knowledge of the structure of the environment, older participants demonstrated a stronger relation between their beliefs about item frequency and associative memory, suggesting that they used their beliefs to guide memory to a greater degree than younger participants.”

Of course, this model only controls for participants explicit beliefs about the structure of the environment. As the reviewer notes, there may be age differences in online frequency learning that also influence encoding. Further, our neural analyses did not account for age differences in explicit frequency reports. We have now run additional control analyses to account for the potential influence of individual differences in frequency learning on associative memory. Specifically, for each participant, we computed three metrics: (1) their overall accuracy during frequency learning, (2) their overall accuracy for the last presentation of each item during frequency learning (as suggested), and (3) the mean magnitude of the error in their frequency reports. We then included these metrics as covariates in our memory analyses.

When we include these control variables in our model, we continue to observe a robust effect of frequency condition (p < 0.001) as well as robust interactions between frequency condition and linear and quadratic age (ps < 0.003). We also observed a main effect of frequency error magnitude on memory accuracy (p < 0.001). Here, however, we no longer observe main effects of age or quadratic age on overall memory accuracy. Given the relation we observed between frequency error magnitudes and age, the results from this model suggests that there may be age-related improvements in overall memory that influence both memory for associations as well as learning of and memory for item frequencies. The fact that age no longer relates to overall memory when controlling for frequency error magnitudes suggest that age-related variance in memory for item frequencies and memory for associations are strongly related within individuals. Importantly, however, age-related variance in memory for item frequencies did not explain age-related variance in the influence of frequency condition on associative memory, suggesting that there are developmental differences in the use of knowledge of environmental structure to prioritize valuable information in memory that persist even when controlling for age-related differences in initial learning of environmental regularities. Given the importance of this analysis in elucidating the relation between the learning of environmental structure and value-guided memory, we have now updated the results in the main text of our manuscript to include them. On p. 13, we now write:

“Because we observed age-related differences in participants’ online learning of item frequencies and in their explicit frequency reports, we further examined whether these age-differences in initial learning could account for the age differences we observed in associative memory. To do so, we ran an additional model in which we included each participant’s mean frequency learning accuracy, mean frequency learning accuracy on the last repetition of each item, and explicit report error magnitude as covariates. Here, explicit report error magnitude strongly predicted overall memory performance, χ2(1) = 13.05, p < 0.001, and we did not observe main effects of age or quadratic age on memory performance (ps > 0.20). However, we continued to observe a main effect of frequency condition, χ2(1) = 19.65 p < 0.001, as well as significant interactions between frequency condition and both linear age χ2(1) = 10.59, p = 0.001, and quadratic age χ2(1) = 9.15, p = 0.002. Thus, while age differences in initial learning related to overall memory performance, they did not account for age differences in the use of environmental regularities to strategically prioritize memory for valuable information.”

In addition, as suggested by the reviewer, we also included the three covariates as control variables in our mediation. When controlling for online frequency learning and explicit frequency report errors, PFC activity continues to mediate the relation between age and memory difference scores. We have now included these results on p. 16 – 17 of the main text:

“Further, when we included quadratic age, WASI scores, online frequency learning accuracy, online frequency learning accuracy on the final repetition of each item, and mean explicit frequency report error magnitudes as control variables in the mediation analysis, PFC activation continued to mediate the relation between linear age and memory difference scores (standardized indirect effect:.56, 95% confidence interval: [.06, 1.31], p = 0.032; standardized direct effect; 1.75, 95% confidence interval: [.11,.3.52], p = 0.030).”

We also refer to these analyses when we interpret our findings in our discussion. Specifically, on p. 23, we write:

“In addition, we continued to observe a robust interaction between age and frequency condition on associative memory, even when controlling for age-related change in the accuracy of both online frequency learning and explicit frequency reports. Thus, though we observed age differences in the learning of environmental regularities and in their influence on subsequent associative memory encoding, our developmental memory effects cannot be fully explained by differences in initial learning.”

We thank the reviewer for this suggestion, as we believe these control analyses strengthen our interpretation of age differences in both the learning and use of environmental regularities to prioritize memory.

3. I think more discussion is warranted on why some neurobehavioral targets show quadratic effects while others show linear. Quadratic effects are often predicted in theoretical models, but from my read on the literature rarely show up in developmental analyses. These data could be leveraged to better understand those models by explaining why some processes are quadratic and others are linear.

We agree with the reviewer that more discussion is warranted here. While many cognitive processes tend to improve with increasing age, the significant interaction between quadratic age and frequency condition on memory accuracy could reflect a number of different patterns of developmental variance. Because quadratic curves are U-shaped, the significant interaction between quadratic age and frequency condition could reflect a peak in value-guided memory in adolescence. However, the combination of linear and quadratic effects can also capture “plateauing” effects, where the influence of age on a particular cognitive process decreases at a particular developmental timepoint. To determine how to interpret the quadratic effect of age on value-guided memory — and specifically, to test for the presence of an adolescent peak — we ran an additional analysis.

To test for an adolescent peak in value-guided memory, we first fit our memory accuracy model without any age terms, and then extracted the random slope across frequency conditions for each subject. We then conducted a ‘two lines test’ (Simonsohn, 2018) to examine the relation between age and these random slopes. In brief, the two-lines test fits the data with two linear models — one with a positive slope and one with a negative slope, algorithmically determining the breakpoint in the estimates where the signs of the slopes change. When we analyzed our memory data in this way, we found a robust, positive relation between age and value-guided memory (see newly added Appendix 2 – Figure 3) from childhood to mid-adolescence, that peaked around age 16 (age 15.86). From age ~16 to early adulthood, however, we observed only a marginal negative relation between age and value-guided memory (p = 0.0567). Thus, our findings do not offer strong evidence in support of an adolescent peak in value-guided memory — instead, they suggest that improvements in value-guided memory are strongest from childhood to adolescence.

To more clearly demonstrate the relation between age and value-guided memory, we have now included the results of the two-lines test in the Results section of our main text. On p. 12 – 13, we write:

“In line with our hypothesis, we observed a main effect of frequency condition on memory, χ2(1) = 21.51, p < 0.001, indicating that individuals used naturalistic value signals to prioritize memory for high-value information. […] Thus, the interaction between frequency condition and quadratic age on memory performance suggests that the biggest age differences in value-guided memory occurred through childhood and early adolescence, with older adolescents and adults performing similarly.”

That said, this developmental trajectory is likely specific to the particular demands of our task. In our previous behavioral study that used a very similar paradigm (Nussenbaum, Prentis, and Hartley, 2020), we observed only a linear relation between age and value-guided memory. Although the task used in our behavioral study was largely similar to the task we employed here, there were subtle differences in the design that may have extended the age range through which we observed improvements in memory prioritization. In particular, in our previous behavioral study, the memory test required participants to select the correct associate from a grid of 20 options (i.e., 1 correct and 19 incorrect options), whereas here, participants had to select the correct associate from a grid of 4 options (1 correct and 3 incorrect options). In our prior work, the need to differentiate the ‘correct’ option from many more foils may have increased the demands on either (or both) memory encoding or memory retrieval, requiring participants to encode and retrieve more specific representations that would be less confusable with other memory representations. By decreasing the task demands in the present study, we may have shifted the developmental curve we observed toward earlier developmental timepoints.

We originally did not emphasize our quadratic findings in the discussion of our manuscript because, given the marginal decrease in memory selectivity we observed from age 16 to age 25 and the different age-related findings across our two studies, we did not want to make strong claims about the specific shape of developmental change. However, we agree with the reviewer that these points are worthy of discussion within the manuscript. We have now amended our discussion on p. 25 accordingly:

“We found that memory prioritization varied with quadratic age, and our follow-up tests probing the quadratic age effect did not reveal evidence for significant age-related change in memory prioritization between late adolescence and early adulthood. However, in our prior behavioral work using a very similar paradigm (Nussenbaum et al., 2020), we found that memory prioritization varied with linear age only. In line with theoretical proposals (Davidow et al., 2018), subtle differences in the control demands between the two tasks (e.g., reducing the number of ‘foils’ presented on each trial of the memory test here relative to our prior study), may have shifted the age range across which we observed differences in behavior, with the more demanding variant of our task showing more linear age-related improvements into early adulthood. In addition, the specific control demands of our task may have also influenced the age at which value-guided memory emerged. Future studies should test whether younger children can modulate encoding based on the value of information if the mnemonic demands of the task are simpler. ”

We thank the reviewer for this helpful suggestion, and believe our additions that expand on the quadratic age effects help clarify our developmental findings.

4. The authors show directly compare the model fits of the mediation models that manipulate directionality.

As noted in the manuscript (p. 16), we found that PFC activity during encoding of pairs involving high- vs. low-frequency items mediated the relation between age and memory differences scores (standardized indirect effect:.07, 95% confidence interval: [.01,.15], p = 0.017; standardized direct effect:.15, 95% confidence interval: [-.03,.33], p = 0.108), but age did not mediate the relation between PFC activity and memory difference scores (standardized indirect effect:.03, 95% confidence interval: [-.007,.09], p = 0.13; standardized direct effect;.34, 95% confidence interval: [.14,.54], p < 0.001.). Directly comparing the fits of the mediation models does not seem possible — In both cases, the full models include the same dependent variable (memory difference scores) and the same two predictor variables (PFC activation and age). Thus, both models will fit the data equivalently well.

However, to further elucidate the directionality of the effects we observed, we examined the AIC difference between the linear regression testing the main effect of the predictor and the linear regression that included the hypothesized mediator to determine the extent to which the mediator improved model fit. Specifically, to examine the extent to which including PFC activation in the model improved the fit of our regression examining the relation between age and memory difference scores, we computed the model AICs with and without PFC activation included as a predictor.

In this case, adding each subject’s PFC activation reduced the AIC by 9.07 and significantly improved model fit (F(1) = 11.39, p = 0.001). Including age as a predictor in the regression examining the relation between PFC activation and memory difference scores only reduced the model AIC by.22 and did not significantly improve model fit (F(1) = 2.17, p = 0.14). Thus, we believe these additional quantitative metrics provide further evidence for the direction of the mediation. These results capture information that is redundant with the mediation analyses included in the manuscript. As such we have chosen not to add them.

5. I found the predictors from the frequency analysis predicting behavior during memory encoding/retrieval to perhaps be the most interesting finding in the paper, especially given that both implicit and explicit measures were predicting memory independently. However, then I was left wanting to know (somewhat desperately!), how much these signals related to the lateral PFC and caudate signals seen during memory encoding. I think this type of analysis would really help make the paper a complete package.

Thank you for this valuable suggestion. We agree that it would be interesting to link frequency-learning behavior to neural activity at encoding. As such, we have now conducted additional analyses to explore these relations.

In the original version of our manuscript, we examined behavior at the item level through mixed-effects models, and neural activation during encoding at the participant level. Thus, to examine the relation between frequency-learning metrics and neural activation at encoding, we created two additional participant-level metrics. For each participant we computed their average repetition suppression index, and a measure of frequency distance. The average repetition suppression index reflects the overall extent to which the participant demonstrated repetition suppression in response to the fifth presentation of the high-frequency items, and is computed by averaging each participant’s repetition suppression indices across items. We hypothesized that participants who demonstrated the greatest degree of repetition suppression might be the most sensitive to the difference between the 1- and 5-frequency items, and therefore, show the greatest differences in striatal and PFC activation during encoding of high- vs. low-value information. The frequency distance metric reflects the average distance between participants’ explicit frequency reports for items that appeared once and items that appeared five times, and is computed by averaging their explicit frequency reports for items in each frequency condition, and then subtracting the average reports in the low-frequency condition from those in the high-frequency condition. We hypothesized that participants with the largest frequency distances might similarly be the most sensitive to the difference between the 1- and 5-frequency items, and therefore, show the greatest differences in striatal and PFC activation during encoding of high- vs. low-value information.

We first wanted to confirm that the relations we observed between repetition suppression, frequency reports, and age, could also be observed at the participant level. In line with our prior, behavioral analyses, we found that age related to both mean repetition suppression indices (marginally; linear age: p = 0.067; quadratic age: p = 0.042); and frequency distances (linear and quadratic age: ps < 0.001).

In addition, we further tested whether these two metrics related to memory performance. In contrast to our item-level findings, we did not observe a significant relation between repetition suppression indices and memory (p = 0.83). We did observe an effect of frequency distance on memory performance. Specifically, we observed significant interactions between frequency distance and age (p = 0.014) and frequency distance and quadratic age (p = 0.021) on memory difference scores, such that the influence of frequency distance on memory difference scores increased with increasing age from childhood to adolescence.

We next examined how mean repetition suppression indices and frequency distances related to differential neural activation during encoding of high- and low-value pairs. In line with our memory findings, we did not observe any significant relations between mean repetition suppression indices and neural activation in the caudate or prefrontal cortex during encoding (ps > 0.15).

Frequency distance did not relate to caudate activation during encoding nor did we observe a frequency distance x age interaction effect (ps > 0.16). Frequency distance did, however, relate to differential PFC activation during encoding of high- vs. low-value pairs. Specifically, we observed a main effect of frequency distance on PFC activation (p = 0.0012), such that participants whose explicit reports of item frequency, were on average, more distinct across frequency conditions, demonstrated increased PFC activation during encoding of pairs involving high- vs. low-frequency items. Interestingly, when we included frequency distance in our model, we no longer observed a significant effect of age on differential PFC activation, nor did we observe a significant frequency distance x age interaction (ps > 0.13). These findings suggest that PFC activation during encoding may have, in part, reflected participants’ beliefs about the structure of the environment, with participants demonstrating stronger differential engagement of control processes across conditions when their representations of the conditions themselves were more distinct.

Finally, we examined how age, frequency distance, and PFC activation related to memory difference scores. Here, even when controlling for both frequency distance and PFC activation, we continued to observe main effects of age and quadratic age on memory difference scores (linear age: p = 0.006; quadratic age: p = 0.001). In line with our analysis of the relation between frequency reports and memory, these results suggest that age-related variance in value-guided memory may depend on both knowledge of the structure of the environment and use of that knowledge to effectively control encoding.

We have now added these results to our manuscript on p. 13 – 14. We write:

“Given the relations we observed between memory and both repetition suppression and frequency reports, we examined whether they related to neural activation in both our caudate and PFC ROI during encoding. […] Here, we did not observe a significant effect of age on PFC activation (β = -.03, SE = 0.13, p = 0.82), suggesting that age-related variance in PFC activation may be related to age differences in explicit frequency beliefs. Importantly, however, even when we accounted for both PFC activation and frequency distances, we continued to observe an effect of age on memory difference scores (β = 0.56, SE = 0.20, p = 0.006), which, together with our prior analyses, suggest that developmental differences in value-guided memory are not driven solely by age differences in beliefs about the structure of the environment but also depend on the use of those beliefs to guide encoding.”

We have added the full model results to Appendix 3.

Given these results, we have now revised our interpretation of our neural data. Our memory analyses demonstrate that across our age range, we observed age-related differences in both the acquisition of knowledge of the structure of the environment and in its use. Originally, we interpreted the PFC activation as reflecting the use of learned value to guide memory. However, the strong relation we found between frequency distance and PFC activation suggests that the age differences in PFC activation that we observed may also be related to age differences in knowledge of the structure of the environment that governs when control processes should be engaged most strongly. However, these results must be interpreted cautiously. Participants provided explicit frequency reports after they completed the encoding and retrieval tasks, and so explicit frequency reports may have been influenced not only by participants’ memories of online frequency learning, but also by the strength with which they encoded the item and its paired associate, and the experience of successfully retrieving it.

We have now revised our discussion to consider these results. On p. 23, we now write,

“Our neural results further suggest that developmental differences in memory were driven by both knowledge of the structure of the environment and use of that knowledge to guide encoding.”

On p. 24, we write,

“The development of adaptive memory requires not only the implementation of encoding and retrieval strategies, but also the flexibility to up- or down-regulate the engagement of control in response to momentary fluctuations in information value (Castel et al., 2007, 2013; Hennessee et al., 2017). Importantly, value-based modulation of lateral PFC engagement during encoding mediated the relation between age and memory selectivity, suggesting that developmental change in both the representation of learned value and value-guided cognitive control may underpin the emergence of adaptive memory prioritization. Prior work examining other neurocognitive processes, including response inhibition (Insel et al., 2017) and selective attention (Störmer et al., 2014), has similarly found that increases in the flexible upregulation of control in response to value cues enhance goal-directed behavior across development (Davidow et al., 2018), and may depend on the engagement of both striatal and prefrontal circuitry (Hallquist et al., 2018; Insel et al., 2017). Here, we extend these past findings to the domain of memory, demonstrating that value signals derived from the structure of the environment increasingly elicit prefrontal cortex engagement and strengthen goal-directed encoding across childhood and into adolescence.”

And on p. 25, we have added an additional paragraph:

“Further, we also demonstrate that in the absence of explicit value cues, the engagement of prefrontal control processes may reflect beliefs about information value that are learned through experience. Here, we found that differential PFC activation during encoding of high- vs. low-value information reflected individual and age-related differences in beliefs about the structure of the environment; participants who represented the average frequencies of the low- and high-frequency items as further apart also demonstrated greater value-based modulation of lateral PFC activation. It is important to note, however, that we collected explicit frequency reports after associative encoding and retrieval. Thus the relation between PFC activation and explicit frequency reports may be bidirectional — while participants may have increased the recruitment of cognitive control processes to better encode information they believed was more valuable, the engagement of more elaborative or deeper encoding strategies that led to stronger memory traces may have also increased participants’ subjective sense of an item’s frequency (Jonides and Naveh-Benjamin, 1987).”

6. Regarding the discussion, I think it would be helpful for the authors to discuss a few features of their data and how they relate to development. The first would be the WASI findings which were quite prominent in most analyses, and in a few showed interactions with age. If this measure is a proxy of executive function, discussing the role of executive function for adaptive memory could help provide a more concrete mechanisms of adaptive memory formation across development.

Thank you for this suggestion. The main focus of our study was to examine age-related change in the influence of learned value signals on memory. Because this study was cross-sectional, one concern was that the children, adolescents, and adults that we recruited may have come from different populations. For example, since we conduct recruitment events at local science fairs and tell participants that they will receive a picture of their brain if they come in, one concern was that the children who signed up to participate may have been particularly excited about science and research, and therefore fundamentally different from the adults we recruited, who may have been more motivated by the monetary compensation. Indeed, we did observe a negative relation between age and age-normed IQ in our sample, suggesting the children had slightly higher IQs for their age relative to adults. Thus, we included WASI scores in all of our analyses to account for these age-related differences in IQ. Our aim in including IQ as a control variable was to partially account for confounding, population-level differences across our age groups, enabling us to more clearly examine the relation between age itself and our neurocognitive processes of interest. We did not intend to examine the role of IQ in motivated memory processes.

We have now clarified why we collected WASI data and used the scores in our analyses in our methods section (p. 7):

“Because this study was cross-sectional, one concern was that the children, adolescents, and adults that we recruited may have come from different populations. Indeed, we observed a significant relation between age and age-normed Wechsler Abbreviated Scale of Intelligence (WASI; Wechsler, 2011) scores in our sample (β = -.60, SE = 0.26, p = 0.0238), suggesting the children had slightly higher estimated IQs for their age relative to adults. To account for these age-related differences in reasoning ability, we included age-normed WASI scores as an interacting fixed effect in all analyses. Our aim in including WASI scores as a control variable was to partially account for confounding, population-level differences across our age groups, enabling us to more clearly examine the relation between age itself and our neurocognitive processes of interest.”

That said, as the reviewer notes, we did observe IQ effects across our analyses. However, in our main memory analysis, we observed a main effect of age-normed WASI IQ score on memory accuracy, but no interactions between WASI scores and frequency condition. This pattern of results suggests that individuals with higher IQs were better at forming associations in memory, but not necessarily better at selectively encoding high-value associations. Similarly, we found that individuals with higher WASI scores were also better at learning the structure of the environment, as demonstrated by a main effect of WASI score on frequency report error magnitudes, but the influence of WASI score on frequency report error magnitudes did not vary as a function of age. Taken together, our results are highly consistent with prior studies that have suggested that IQ may relate to learning and memory across development (Deary et al., 2010; Rose et al., 2012).

The reason we have not included a discussion of WASI scores in the manuscript is because our primary goal was to elucidate the relation between age and motivated memory processes. Across our analyses, we observed few interactions between WASI scores and age. We did observe significant interactions between age and WASI scores on differential neural activation within both the caudate and hippocampus during encoding of high- vs. low-value information. We do not have a strong interpretation for these results, so we report them but do not discuss them.

We agree with the reviewer that it would be valuable to identify more concrete mechanisms that drive developmental change in the component processes of adaptive memory. However, as we now clarify in our methods, the purpose of including the WASI was to account for age-related differences in reasoning ability that may confound our interpretation of our age effects. Thus, we do not want to overinterpret our WASI findings.

Additionally, it would be helpful for the authors to discuss the negative findings during retrieval. The fact that these age-related differences in adaptive memory processes are most likely seeming from encoding versus retrieval is not only highly interesting, but are a major prediction that stems from an animal literature on dopaminergic influences on hippocampal-dependent memory.

Most of our hypotheses about age-related change in value-guided memory were about changes in encoding processes. In our whole-brain analyses, we found that neural activation at encoding related to memory difference scores. Neural activation at retrieval, however, did not vary by memory difference scores or by age. Given these findings, we focused our subsequent analyses on encoding.

As the reviewer notes, the animal literature makes interesting theoretical predictions about the role of dopamine in value-guided memory encoding processes. Because we don’t have data that can tie our findings to dopaminergic mechanisms, we refrain from too much speculation. However, we have now revised our discussion to emphasize the specificity of our findings to the encoding phase of the task. We now discuss our retrieval findings in more depth on p. 25:

“During retrieval, we continued to observe increased activation of the caudate and dlPFC for high- vs. low-value pairs. However, this activation did not significantly vary as a function of memory difference scores or age, suggesting that the developmental differences in value-guided memory that we observed were likely driven by age-related change in encoding processes.”

7. It would be helpful to report all behavioral results in the youngest sample only, as the interpretation of the data is quite different if children can or cannot perform these processes.

Thank you for this suggestion. Determining whether the “youngest” participants can or cannot use learned value to influence memory is complicated for several reasons. First, drawing a line to determine who to include in our analysis of the ‘youngest’ participants feels fairly arbitrary — our youngest children included were age 8, but we only have 5 8-year-olds in our sample, and so we would not expect to observe any significant influence of frequency condition within such a small sample. We can include more participants, but by doing so, we then need to add older participants (e.g., 9- and 10-year-olds) to this analysis, potentially muddling what conclusions can be drawn. Thus, we do not believe this question is best answered by statistics — instead, we have chosen to display individual points for each participant that depict their memory performance in Figure 4 on p. 12. Per Reviewer 3’s suggestion, we have also now connected the points belonging to the same participants to more clearly display individual participants’ performance. We believe this plot can give readers the best sense of how participants across our age range performed on the task.

Perhaps more importantly, in line with theoretical accounts of the development of value-guided cognitive control (Davidow, Insel, and Somerville, 2018), we also believe that the specific age at which the ability to use value to guide memory encoding emerges will be highly dependent on the nature of the control demands elicited by the specific task used. For example, younger children may have demonstrated stronger effects of value on memory if they only had to encode two associations (instead of 24), if the stimuli themselves were easier to represent in memory (e.g., child-friendly, nameable objects rather than more abstract images and scenes), or if the value differences were greater (e.g., 1 vs. 10 repetitions instead of 1 vs. 5). Thus, rather than making a claim about the specific age at which motivated memory emerges, we have instead expanded our discussion of this issue on p. 25 – 26 of the manuscript:

“We found that memory prioritization varied with quadratic age, and our follow-up tests probing the quadratic age effect did not reveal evidence for significant age-related change in memory prioritization between late adolescence and early adulthood. However, in our prior behavioral work using a very similar paradigm (Nussenbaum et al., 2020), we found that memory prioritization varied with linear age only. In line with theoretical proposals (Davidow et al., 2018), subtle differences in the control demands between the two tasks (e.g., reducing the number of ‘foils’ presented on each trial of the memory test here relative to our prior study), may have shifted the age range across which we observed differences in behavior, with the more demanding variant of our task showing more linear age-related improvements into early adulthood. In addition, the specific control demands of our task may have also influenced the age at which value-guided memory emerged. Future studies should test whether younger children can modulate encoding based on the value of information if the mnemonic demands of the task are simpler.”

8. Could the authors provide data on differences in behavioral performance for both task content (which could drive motivational differences) or task order (which might lead to practice effects). If notable differences emerge across these factors, I strongly believe they needed to be included as co-variates in ally analyses.

Thank you for raising this important point. During the review process, we also received feedback on our preprint from a colleague who suggested that in addition to including participant random effects in our models, we should also include stimulus random effects (Barr et al., 2013). As noted by the reviewer, task content may influence motivation or memorability — the stimulus-level random effects should account for variance in memory performance that may be driven by differences in how participants encode individual stimuli, beyond any effects of our frequency manipulation. Thus, in our revised manuscript, we have updated our modeling approach accordingly. On p. 33 of our methods section, we now write:

“To determine the random effects structures of our mixed effects models, we began with the maximal model to minimize Type I errors (Barr et al., 2013). We included random participant intercepts and slopes across all fixed effects (except age and IQ) and their interactions. We also included random stimulus intercepts and slopes across all fixed effects and their interactions. Because stimuli were randomly paired during associative encoding and only repeated, on average, around 4 times across participants, our stimulus random effects accounted for individual items (e.g., postcard 1) rather than pairs of items (e.g., postcard 1 and stamp 5). We set the number of model iterations to one million and use the “bobyqa” optimizer. When the maximal model gave convergence errors or failed to converge within a reasonable timeframe (~24 hours), we removed correlations between random slopes and random intercepts, followed by random slopes for interaction effects, followed by random slopes across stimuli. For full details about the fixed- and random-effects structure of all models, see Appendix 3: Full Model Specification and Results.”

Importantly, all of the effects that we reported in the original version of our manuscript hold when accounting for stimulus-level random effects, indicating that our results are robust to individual differences across stimuli.

As the reviewer suggested, we also ran two additional models to test the influence of block order and block type (postcards and stamps vs. pictures and picture frames) on associative memory. We did not observe any significant main effects of or interactions with block order (main effect: χ2(1) = 0.18, p = 0.675, all interaction p values > 0.20), but we did observe a significant block type x frequency condition interaction effect (p = 0.036). This seems to be driven by better memory performance for associations involving low-frequency pictures relative to low-frequency postcards (see newly added Appendix 2 – Figure 5). Importantly, all the other effects in our model (e.g., frequency condition, age x frequency condition, etc.) hold when we account for block type.

A priori, we planned to collapse data across our two blocks for all analyses; we split our experiment into two blocks to make frequency learning and associative memory encoding easier for participants while still ensuring we had enough trials for our neural analyses to be adequately powered — our experiment was not designed to examine the effects of block type or order. However, given that we observed a frequency condition x block type interaction on associative memory, we also examined each participant’s average β weight (parameter estimate) within our PFC ROI for each block separately. We used fslmeants to extract these parameter estimates, and then examined their relations with both age and memory difference scores, which we also computed separately for each block.

Briefly, when we included block type as a covariate in our analyses, we continued to observe significant effects of age and quadratic age on differential PFC activation across frequency conditions during encoding. Further, we continued to observe a relation between PFC activation and memory difference scores, suggesting that differential engagement of the PFC during encoding of high- vs. low-value associations may, in part, account for individual and developmental differences in value-guided memory.

Thus, while we did observe differences in behavior that related to task content, our main conclusions hold when controlling for them. We have now included these additional analyses of task content in Appendix 2: Supplementary Analyses (Because they span four pages, we have not included them here).

We refer to these analyses in our methods, on p. 33:

“Data were combined across blocks (but we include an analysis of block effects on memory performance in Appendix 2: Supplementary Analyses).”

We thank the reviewer for prompting us to more fully examine different influences on memory performance in our task, and believe that these supplementary results provide greater evidence for our central argument.

Reviewer #3 (Recommendations for the authors):

1) Empirical findings directly comparing cross-sectional and longitudinal effects have demonstrated that cross-sectional analyses of age differences do not readily generalize to longitudinal research (e.g., Raz et al., 2005; Raz and Lindenberger, 2012). Formal analyses have demonstrated that proportion of explained age-related variance in cross-sectional mediation models may stem from various factors, including similar mean age trends, within-time correlations between a mediator and an outcome, or both (Lindenberger et al., 2011; see also Hofer, Flaherty, and Hoffman, 2006; Maxwell and Cole, 2007). Thus, the results of the mediation analysis showing that PFC activation explains age-related variance in memory difference scores, cannot be taken to imply that changes in PFC activation are correlated with changes in value-guided memory. While the general limitations of a cross-sectional study are noted in the Discussion of the manuscript, it would be important to discuss the critical limitations of the mediation analysis. While the main conclusions of the paper do not critically depend on this analysis, it would be important to alert the reader to the limited information value in performing cross-sectional mediation analyses of age variance.

Thank you for raising this critical point. We have expanded our discussion to specifically note the limitations of our mediation analysis and to more strongly emphasize the need for future longitudinal studies to reveal how changes in neural circuitry may support the emergence of motivated memory across development. Specifically, on p. 26, we now write:

“One important caveat is that our study was cross-sectional — it will be important to replicate our findings in a longitudinal sample to more directly measure how developmental changes in cognitive control within an individual contribute to changes in their ability to selectively encode useful information. Our mediation results, in particular, must be interpreted with caution as simulations have demonstrated that in cross-sectional samples, variables can emerge as significant mediators of age-related change due largely to statistical artifact (Hofer, Flaherty, and Hoffman, 2006; Lindenberger et al., 2011). Indeed, our finding that PFC activation mediates the relation between age and value-guided memory does not necessarily imply that within an individual, PFC development leads to improvements in memory selectivity. Longitudinal work in which individuals’ neural activity and memory performance is sampled densely within developmental windows of interest is needed to elucidate the complex relations between age, brain development, and behavior (Hofer, Flaherty, and Hoffman, 2006; Lindenberger et al., 2011).”

2) It would be helpful to provide more information on how chance memory performance was handled during data analysis, especially as it is more likely to occur in younger participants. Related to this, please connect the points that belong to the same individual in Figure 3 to facilitate evaluation of individual differences in the memory difference scores.

Thank you for raising this important point. On each memory test trial, participants viewed the item (either a postcard or picture) above images of four possible paired associates (see Figure 1 on p. 6). On each memory test trial, participants had 6 seconds to select one of these items. If participants did not make a response within 6 seconds, that trial was considered ‘missed.’ Missed trials were excluded from behavioral analyses and regressed out in neural analyses. If participants selected the correct associate, memory accuracy was coded as ‘1;’ if they selected an incorrect associate, accuracy was coded as ‘0.’ On each trial, there was 1 correct option and 3 incorrect options. As such, chance-level memory performance was 25%. We have now clarified this on p. 34 and included a dashed line indicating chance-level performance within Figure 4 (formerly Figure 3) on p. 12. In addition, we have also updated Figure 4 to connect the points belonging to the same participants, as suggested by the reviewer.

Out of 90 participants, 2 children performed at or below chance (<= 25% memory accuracy). Interpreting the behavior of the participants who responded to fewer than 12 out of 48 trials correctly is challenging. On the one hand, they might not have remembered anything and responded correctly on these trials due to randomly guessing. On the other hand, they may have implemented an encoding strategy of focusing only on a small number of pairs. Thus, a priori, based on the analysis approach we implemented in our prior, behavioral study (Nussenbaum et al., 2019), we decided to include all participants in our memory analyses, regardless of their overall accuracy. However, when we exclude these two participants from our memory analyses, our main findings still hold. Specifically, we continue to observe main effects of frequency condition and age, and interactions between frequency condition and both linear and quadratic age on associative memory accuracy (ps < 0.012).

We have now clarified these details about chance-level performance in the methods section of our manuscript on p. 34.

“For our memory analyses, trials were scored as ‘correct’ if the participant selected the correct association from the set of four possible options presented during the memory test, ‘incorrect’ if the participant selected an incorrect association, and ‘missed’ if the participant failed to respond within the 6-second response window. Missed trials were excluded from all analyses. Because participants had to select the correct association from four possible options, chance-level performance was 25%. Two child participants performed at or below chance-level on the memory test. They were included in all analyses reported in the manuscript; however, we report full details of the results of our memory analyses when we exclude these two participants in Appendix 3 (Table 15). Importantly, our main findings remain unchanged.”

In Appendix 3, we include a table with the full results from our memory model without these two participants.

3) I would like to see some consideration of how the different signatures of value learning, repetition suppression and reported item frequency, are related to the observed PFC and caudate effects during memory encoding. Such a discussion would help the reader connect the findings on learning and using information value across development.

Thank you for this valuable suggestion. We agree that it would be interesting to link frequency-learning behavior to neural activity at encoding. As such, we have now conducted additional analyses to explore these relations.

In the original version of our manuscript, we examined behavior at the item level through mixed-effects models, and neural activation during encoding at the participant level. Thus, to examine the relation between frequency-learning metrics and neural activation at encoding, we created two additional participant-level metrics. For each participant we computed their average repetition suppression index, and a measure of frequency distance. The average repetition suppression index reflects the overall extent to which the participant demonstrated repetition suppression in response to the fifth presentation of the high-frequency items, and is computed by averaging each participant’s repetition suppression indices across items. We hypothesized that participants who demonstrated the greatest degree of repetition suppression might be the most sensitive to the difference between the 1- and 5-frequency items, and therefore, show the greatest differences in striatal and PFC activation during encoding of high- vs. low-value information. The frequency distance metric reflects the average distance between participants’ explicit frequency reports for items that appeared once and items that appeared five times, and is computed by averaging their explicit frequency reports for items in each frequency condition, and then subtracting the average reports in the low-frequency condition from those in the high-frequency condition. We hypothesized that participants with the largest frequency distances might similarly be the most sensitive to the difference between the 1- and 5-frequency items, and therefore, show the greatest differences in striatal and PFC activation during encoding of high- vs. low-value information.

We first wanted to confirm that the relations we observed between repetition suppression, frequency reports, and age, could also be observed at the participant level. In line with our prior, behavioral analyses, we found that age related to both mean repetition suppression indices (marginally; linear age: p = 0.067; quadratic age: p = 0.042); and frequency distances (linear and quadratic age: ps < 0.001).

In addition, we further tested whether these two metrics related to memory performance. In contrast to our item-level findings, we did not observe a significant relation between repetition suppression indices and memory (p = 0.83). We did observe an effect of frequency distance on memory performance. Specifically, we observed significant interactions between frequency distance and age (p = 0.014) and frequency distance and quadratic age (p = 0.021) on memory difference scores, such that the influence of frequency distance on memory difference scores increased with increasing age from childhood to adolescence.

We next examined how mean repetition suppression indices and frequency distances related to differential neural activation during encoding of high- and low-value pairs. In line with our memory findings, we did not observe any significant relations between mean repetition suppression indices and neural activation in the caudate or prefrontal cortex during encoding (ps > 0.15).

Frequency distance did not relate to caudate activation during encoding nor did we observe a frequency distance x age interaction effect (ps > 0.16). Frequency distance did, however, relate to differential PFC activation during encoding of high- vs. low-value pairs. Specifically, we observed a main effect of frequency distance on PFC activation (p = 0.0012), such that participants whose explicit reports of item frequency, were on average, more distinct across frequency conditions, demonstrated increased PFC activation during encoding of pairs involving high- vs. low-frequency items. Interestingly, when we included frequency distance in our model, we no longer observed a significant effect of age on differential PFC activation, nor did we observe a significant frequency distance x age interaction (ps > 0.13). These findings suggest that PFC activation during encoding may have, in part, reflected participants’ beliefs about the structure of the environment, with participants demonstrating stronger differential engagement of control processes across conditions when their representations of the conditions themselves were more distinct.

Finally, we examined how age, frequency distance, and PFC activation related to memory difference scores. Here, even when controlling for both frequency distance and PFC activation, we continued to observe main effects of age and quadratic age on memory difference scores (linear age: p = 0.006; quadratic age: p = 0.001). In line with our analysis of the relation between frequency reports and memory, these results suggest that age-related variance in value-guided memory may depend on both knowledge of the structure of the environment and use of that knowledge to effectively control encoding.

We have now added these results to our manuscript on p. 13 – 14. We write:

“Given the relations we observed between memory and both repetition suppression and frequency reports, we examined whether they related to neural activation in both our caudate and PFC ROI during encoding. […] Importantly, however, even when we accounted for both PFC activation and frequency distances, we continued to observe an effect of age on memory difference scores (β = 0.56, SE = 0.20, p = 0.006), which, together with our prior analyses, suggest that developmental differences in value-guided memory are not driven solely by age differences in beliefs about the structure of the environment but also depend on the use of those beliefs to guide encoding.”

We have added the full model results to Appendix 3.

Given these results, we have now revised our interpretation of our neural data. Our memory analyses demonstrate that across our age range, we observed age-related differences in both the acquisition of knowledge of the structure of the environment and in its use. Originally, we interpreted the PFC activation as reflecting the use of learned value to guide memory. However, the strong relation we found between frequency distance and PFC activation suggests that the age differences in PFC activation that we observed may also be related to age differences in knowledge of the structure of the environment that governs when control processes should be engaged most strongly. However, these results must be interpreted cautiously. Participants provided explicit frequency reports after they completed the encoding and retrieval tasks, and so explicit frequency reports may have been influenced not only by participants’ memories of online frequency learning, but also by the strength with which they encoded the item and its paired associate, and the experience of successfully retrieving it.

We have now revised our discussion to consider these results. On p. 23, we now write,

“Our neural results further suggest that developmental differences in memory were driven by both knowledge of the structure of the environment and use of that knowledge to guide encoding.”

On p. 24, we write,

“The development of adaptive memory requires not only the implementation of encoding and retrieval strategies, but also the flexibility to up- or down-regulate the engagement of control in response to momentary fluctuations in information value (Castel et al., 2007, 2013; Hennessee et al., 2017). Importantly, value-based modulation of lateral PFC engagement during encoding mediated the relation between age and memory selectivity, suggesting that developmental change in both the representation of learned value and value-guided cognitive control may underpin the emergence of adaptive memory prioritization. Prior work examining other neurocognitive processes, including response inhibition (Insel et al., 2017) and selective attention (Störmer et al., 2014), has similarly found that increases in the flexible upregulation of control in response to value cues enhance goal-directed behavior across development (Davidow et al., 2018), and may depend on the engagement of both striatal and prefrontal circuitry (Hallquist et al., 2018; Insel et al., 2017). Here, we extend these past findings to the domain of memory, demonstrating that value signals derived from the structure of the environment increasingly elicit prefrontal cortex engagement and strengthen goal-directed encoding across childhood and into adolescence.”

And on p. 25, we have added an additional paragraph:

“Further, we also demonstrate that in the absence of explicit value cues, the engagement of prefrontal control processes may reflect beliefs about information value that are learned through experience. Here, we found that differential PFC activation during encoding of high- vs. low-value information reflected individual and age-related differences in beliefs about the structure of the environment; participants who represented the average frequencies of the low- and high-frequency items as further apart also demonstrated greater value-based modulation of lateral PFC activation. It is important to note, however, that we collected explicit frequency reports after associative encoding and retrieval. Thus the relation between PFC activation and explicit frequency reports may be bidirectional — while participants may have increased the recruitment of cognitive control processes to better encode information they believed was more valuable, the engagement of more elaborative or deeper encoding strategies that led to stronger memory traces may have also increased participants’ subjective sense of an item’s frequency (Jonides and Naveh-Benjamin, 1987).”

4) A point worthy of discussion are the implications of the finding that younger participants demonstrated greater deviations in their frequency reports for the development of value learning, given that frequency reports were found to predict associative memory accuracy.

Thank you for raising this important point. Indeed, one of our main findings is that older participants are better both at learning the structure of their environments and also at using structured knowledge to strategically prioritize memory. In our original manuscript, we described results of a model that included participants’ explicit frequency reports as a predictor of memory. Model comparison revealed that participants’ frequency reports — which we interpret as reflecting their beliefs about the structure of the environment — predicted memory more strongly than the item’s true frequency. In other words, participants’ beliefs about the structure of the environment (even if incorrect) more strongly influenced their memory encoding than the true structure of the environment. Critically, however, frequency reports interacted with age to predict memory (Figure 8). Even when we accounted for age-related differences in knowledge of the structure of the environment, older participants demonstrated a stronger influence of frequency on memory, suggesting they were better able to use their beliefs to control subsequent associative encoding. We have now clarified our interpretation of this model in our discussion on p. 23:

“Importantly, though we observed age-related differences in participants’ learning of the structure of their environment, the strengthening of the relation between frequency reports and associative memory with increasing age suggests that age differences in learning cannot fully account for age differences in value-guided memory. Even when accounting for individual differences in participants’ explicit knowledge of the structure of the environment, older participants demonstrated a stronger relation between their beliefs about item frequency and associative memory, suggesting that they used their beliefs to guide memory to a greater degree than younger participants.”

As noted by the reviewer, however, our initial memory analysis did not account for age-related differences in participants’ initial, online learning of item frequency, and our neural analyses further did not account for age differences in explicit frequency reports. We have now run additional control analyses to account for the potential influence of individual differences in frequency learning on associative memory. Specifically, for each participant, we computed three metrics: (1) their overall accuracy during frequency learning, (2) their overall accuracy for the last presentation of each item during frequency learning (as suggested by Reviewer 2), and (3) the mean magnitude of the error in their frequency reports. We then included these metrics as covariates in our memory analyses.

When we include these control variables in our model, we continue to observe a robust effect of frequency condition (p < 0.001) as well as robust interactions between frequency condition and linear and quadratic age (ps < 0.003) on associative memory accuracy. We also observed a main effect of frequency error magnitude on memory accuracy (p < 0.001). Here, however, we no longer observe main effects of age or quadratic age on overall memory accuracy. Given the relation we observed between frequency error magnitudes and age, the results from this model suggests that there may be age-related improvements in overall memory that influence both memory for associations as well as learning of and memory for item frequencies. The fact that age no longer relates to overall memory when controlling for frequency error magnitudes suggest that age-related variance in memory for item frequencies and memory for associations are strongly related within individuals. Importantly, however, age-related variance in memory for item frequencies did not explain age-related variance in the influence of frequency condition on associative memory, suggesting that there are developmental differences in the use of knowledge of environmental structure to prioritize valuable information in memory that persist even when controlling for age-related differences in initial learning of environmental regularities. Given the importance of this analysis in elucidating the relation between the learning of environmental structure and value-guided memory, we have now updated the results in the main text of our manuscript to include them. Specifically, on p. 13, we now write:

“Because we observed age-related differences in participants’ online learning of item frequencies and in their explicit frequency reports, we further examined whether these age differences in initial learning could account for the age differences we observed in associative memory. To do so, we ran an additional model in which we included each participant’s mean frequency learning accuracy, mean frequency learning accuracy on the last repetition of each item, and explicit report error magnitude as covariates. Here, explicit report error magnitude predicted overall memory performance, χ2(1) = 13.05, p < 0.001, and we did not observe main effects of age or quadratic age on memory performance (ps > 0.20). However, we continued to observe a main effect of frequency condition, χ2(1) = 19.65 p < 0.001, as well as significant interactions between frequency condition and both linear age χ2(1) = 10.59, p = 0.001, and quadratic age χ2(1) = 9.15, p = 0.002. Thus, while age differences in initial learning related to overall memory performance, they did not account for age differences in the use of environmental regularities to strategically prioritize memory for valuable information.”

In addition, as suggested by the reviewer, we also included the three covariates as control variables in our mediation analysis. When controlling for online frequency learning and explicit frequency report errors, PFC activity continued to mediate the relation between age and memory difference scores. We have now included these results on p. 16 – 17 of the main text:

“Further, when we included quadratic age, WASI scores, online frequency learning accuracy, online frequency learning accuracy on the final repetition of each item, and mean explicit frequency report error magnitudes as control variables in the mediation analysis, PFC activation continued to mediate the relation between linear age and memory difference scores (standardized indirect effect:.56, 95% confidence interval: [.06, 1.35], p = 0.023; standardized direct effect; 1.75, 95% confidence interval: [.12,.3.38], p = 0.034).”

We also refer to these analyses when we interpret our findings in our discussion. On p. 23, we write:

“In addition, we continued to observe a robust interaction between age and frequency condition on associative memory, even when controlling for age-related change in the accuracy of both online frequency learning and explicit frequency reports. Thus, though we observed age differences in the learning of environmental regularities and in their influence on subsequent associative memory encoding, our developmental memory effects cannot be fully explained by differences in initial learning.”

We thank the reviewer for this constructive suggestion, as we believe these control analyses strengthen our interpretation of age differences in both the learning and use of environmental regularities to prioritize memory.

5) It would be helpful to include (supplementary) figures accompanying the behavioral results from the frequency learning phase.

We have now included a supplementary figure (Appendix 2 – Figure 2) showing how accuracy and reaction times during frequency learning vary across age and item appearance counts.

6) Supplementary Figure S2 – providing an (additional) plot of the estimated age effects would help match the displayed results better to the description in the main text.

Thank you for this suggestion. We now include a new figure in the main text of manuscript (Figure 3 on p. 10) that shows neural activation in the parahippocampal cortex in response to all item repetitions, as well as the relation between age and repetition suppression, as suggested.

https://doi.org/10.7554/eLife.69796.sa2

Article and author information

Author details

  1. Kate Nussenbaum

    New York University, New York City, United States
    Contribution
    Conceptualization, Data curation, Software, Formal analysis, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing - original draft, Project administration, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7185-6880
  2. Catherine A Hartley

    New York University, New York City, United States
    Contribution
    Conceptualization, Resources, Supervision, Funding acquisition, Methodology, Writing - review and editing
    For correspondence
    cate@nyu.edu
    Competing interests
    Reviewing editor, eLife
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-0177-7295

Funding

National Science Foundation (1654393)

  • Catherine Hartley

Jacobs Foundation (Early Career Research Fellowship)

  • Catherine Hartley

U.S. Department of Defense (National Defense Science and Engineering Graduate Fellowship)

  • Kate Nussenbaum

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank Daphne Valencia, Jamie Greer, Nora Keathley, and Michael Liu for assistance with data collection, Ali Cohen and Gail Rosenbaum for valuable discussions and help with analyses, and the staff of NYU’s Center for Brain Imaging, especially Pablo Velasco, for technical support and guidance. This project was supported by a National Science Foundation CAREER Grant (1654393) to CAH, a Jacobs Foundation Early Career Fellowship to CAH and a National Defense Science and Engineering Graduate Fellowship to KN.

Ethics

Human subjects: Research procedures were approved by New York University's Institutional Review Board (IRB-2016-1194). Adult participants provided written consent prior to participating in the study. Children and adolescents provided written assent, and their parents or guardians provided written consent on their behalf, prior to their participation. All participants were compensated $60 for the experimental session, which involved a 1-hour MRI scan.

Senior Editor

  1. Christian Büchel, University Medical Center Hamburg-Eppendorf, Germany

Reviewing Editor

  1. Thorsten Kahnt, Northwestern University, United States

Reviewer

  1. Vishnu Murty, Temple University, United States

Publication history

  1. Preprint posted: February 14, 2021 (view preprint)
  2. Received: April 27, 2021
  3. Accepted: August 23, 2021
  4. Version of Record published: September 20, 2021 (version 1)

Copyright

© 2021, Nussenbaum and Hartley

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 290
    Page views
  • 22
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Neuroscience
    Roni O Maimon-Mor et al.
    Research Article Updated

    The study of artificial arms provides a unique opportunity to address long-standing questions on sensorimotor plasticity and development. Learning to use an artificial arm arguably depends on fundamental building blocks of body representation and would therefore be impacted by early life experience. We tested artificial arm motor-control in two adult populations with upper-limb deficiencies: a congenital group—individuals who were born with a partial arm, and an acquired group—who lost their arm following amputation in adulthood. Brain plasticity research teaches us that the earlier we train to acquire new skills (or use a new technology) the better we benefit from this practice as adults. Instead, we found that although the congenital group started using an artificial arm as toddlers, they produced increased error noise and directional errors when reaching to visual targets, relative to the acquired group who performed similarly to controls. However, the earlier an individual with a congenital limb difference was fitted with an artificial arm, the better their motor control was. Since we found no group differences when reaching without visual feedback, we suggest that the ability to perform efficient visual-based corrective movements is highly dependent on either biological or artificial arm experience at a very young age. Subsequently, opportunities for sensorimotor plasticity become more limited.

    1. Cell Biology
    2. Neuroscience
    Shahzad S Khan et al.
    Research Advance

    Activating LRRK2 mutations cause Parkinson's disease, and pathogenic LRRK2 kinase interferes with ciliogenesis. Previously, we showed that cholinergic interneurons of the dorsal striatum lose their cilia in R1441C LRRK2 mutant mice (Dhekne et al., 2018). Here, we show that cilia loss is seen as early as 10 weeks of age in these mice and also in two other mouse strains carrying the most common human G2019S LRRK2 mutation. Loss of the PPM1H phosphatase that is specific for LRRK2-phosphorylated Rab GTPases yields the same cilia loss phenotype seen in mice expressing pathogenic LRRK2 kinase, strongly supporting a connection between Rab GTPase phosphorylation and cilia loss. Moreover, astrocytes throughout the striatum show a ciliation defect in all LRRK2 and PPM1H mutant models examined. Hedgehog signaling requires cilia, and loss of cilia in LRRK2 mutant rodents correlates with dysregulation of Hedgehog signaling as monitored by in situ hybridization of Gli1 and Gdnf transcripts. Dopaminergic neurons of the substantia nigra secrete a Hedgehog signal that is sensed in the striatum to trigger neuroprotection; our data support a model in which LRRK2 and PPM1H mutant mice show altered responses to critical Hedgehog signals in the nigrostriatal pathway.