The Effects of Age on the Neural Correlates of Recollection Success, Recollection-Related Cortical Reinstatement, and Post-Retrieval Monitoring

Abstract

Functional magnetic resonance imaging (fMRI) was used to investigate whether age-related differences in episodic memory performance are accompanied by a reduction in the specificity of recollected information. We addressed this question by comparing recollection-related cortical reinstatement in young and older adults. At study, subjects viewed objects and concrete words, making 1 of 2 different semantic judgments depending on the study material. Test items were words that corresponded to studied words or the names of studied objects. Subjects indicated whether each test item was recollected, familiar, or novel. Reinstatement of information differentiating the encoding tasks was quantified both with a univariate analysis of the fMRI signal and with a multivoxel pattern analysis, using a classifier that had been trained to discriminate between the 2 classes of study episode. The results of these analyses converged to suggest that reinstatement did not differ according to age. Thus, there was no evidence that specificity of recollected information was reduced in older individuals. Additionally, there were no age effects in the magnitude of recollection-related modulations in regional activity or in the neural correlates of post-retrieval monitoring. Taken together, the findings suggest that the neural mechanisms engaged during successful episodic retrieval can remain stable with advancing age.

Introduction

Recollection of events and their contexts (episodic memory) declines with increasing age (e.g. Park et al. 1996; Nilsson 2003; Old and Naveh-Benjamin 2008; Koen and Yonelinas 2014). This decline has been attributed to impairment of encoding processes that support the capacity to associate or “bind” the different features of an event into a cohesive memory representation (e.g. Naveh-Benjamin and Craik 1995; Naveh-Benjamin 2000; Luo et al. 2007; for review, see Craik and Rose 2012), coupled with reduced ability to engage goal-directed processing of retrieval cues (e.g. Morcom and Rugg 2004; Jacoby et al. 2005) and less effective monitoring and evaluation of the products of retrieval (e.g. Gallo et al. 2006; McDonough et al. 2013; Mitchell et al. 2013).

Recollection can be operationalized in multiple ways. One procedure involves the assessment of source memory—memory for a recognized test item's study context. The findings of numerous studies demonstrate that performance on such tests is markedly reduced in older, compared with younger, individuals [e.g. McIntyre and Craik 1987; Schacter et al. 1991; Henkel et al. 1998; Glisky et al. 2001; Dulas and Duarte 2012, for reviews, see Spencer and Raz (1995) and Cansino (2009)]. Another important means of assessing recollection—the “Remember/Know” (R/K) procedure—relies on subjective report of whether recognition of a test item was accompanied by memory for one or more aspects of its study context, or instead was accompanied solely by a sense of familiarity (Tulving 1985; Gardiner 1988; Gardiner and Richarson-Klavehn 2000). Consistent with the findings from studies of source memory, estimates of recollection derived from the R/K procedure are typically reported to be lower in older than in young adults [e.g. Parkin and Walter 1992; Mäntyla 1993; Perfect and Dasgupta 1997; Schacter et al. 1997; Bastin and Van der Linden 2003; Prull et al. 2006; Bugaiska et al. 2007, see McCabe et al. (2009) and Koen and Yonelinas (2014) for reviews].

Whereas estimates of recollection derived from source memory and R/K judgments both differ according to age, there is evidence that age exerts less of an impact on estimates derived from the latter procedure. This evidence includes the findings of 3 studies in which performance on a source memory and an R/K test was directly contrasted while holding study conditions and experimental materials constant. Mark and Rugg (1998) reported that source accuracy was higher in their young than in their older participants, whereas the proportions of studied test items endorsed as “Remembered” did not differ with age. In 2 other studies (Duarte et al. 2006, 2008), participants made both a source and an R/K judgment on each test trial. In each study, older participants were separated into high- and low-performing subgroups as defined by overall recognition performance. Recollection estimates derived from the R/K procedure did not differ between the young and high-performing older subgroups, but there was a marked disparity in source memory performance in favor of the younger participants. These findings are consistent with more general claims of a disparity in older individuals between their subjective experiences of recollection, and their ability to demonstrate successful recollection on “objective” memory tests such as source judgments. For example, it has been reported that even as older individuals demonstrate impaired performance on tests such as judgments of source memory, they describe recollective experiences that are seemingly as rich and as detailed as those of younger subjects (e.g. Gallo et al. 2011). The findings are also consistent with evidence that older adults are more dependent than young individuals on nonspecific “gist” information when making memory judgments (e.g. Norman and Schacter 1997; Pierce et al. 2005; Carr et al. 2015), perhaps reflecting age-related decline in frontal lobe function (Koutstaal and Cavendish 2006; Aizpurua and Koutstaal 2010).

What underlies this apparent age-related disparity in the sensitivity of “objective” and “subjective” aspects of recollection? As was discussed by McDonough et al. (2014), one possibility is that older individuals recollect less information than young individuals, but that they have “recalibrated” by lowering their criterion for reporting a subjectively rich memory. Alternatively (or additionally), the disparity might reflect the fact that a subjective experience of recollection can be based on less differentiated information than that required to support a source memory judgment. Whereas accurate source memory judgments require retrieval of information diagnostic of a specific contextual feature, “Remember” judgments, and introspective reports about the “richness” of a memory more generally can be based on nondiagnostic information [i.e. on “noncriterial” recollection; Yonelinas and Jacoby 1996; see also McDonough et al. (2014)].

The principal aim of the present study was to assess whether specificity of recollection declines with age by employing functional magnetic resonance imaging (fMRI) to exploit the phenomenon of retrieval-related “cortical reinstatement.” This term refers to the finding that successful recollection is associated with patterns of cortical activity that partially overlap the patterns elicited when the recollected information was initially encoded. For example, Johnson and Rugg (2007) had subjects' study a series of words that were presented against either a scene or a gray background (with the instruction in the latter case to generate a sentence incorporating the word). At test, participants made “Remember/Know/New” judgments on a series of studied and unstudied words. Regardless of their study history, relative to words endorsed “Know,” recollected words elicited enhanced activity in what has been termed the “core recollection network”—a set of regions, including medial temporal lobe (MTL) structures and posterior cingulate, along with medial prefrontal and ventral parietal cortex, that is consistently activated during successful episodic retrieval [see Kim (2010) and Rugg and Vilberg (2013) for reviews]. Outside of this network, recollection-related activity differed according to the study history of the recollected words: Whereas activity exclusively elicited by words encoded in the scene task partially overlapped cortical regions that demonstrated greater encoding-related activity for scenes than for sentences, words studied in the sentence task elicited activity that overlapped regions that were more active during sentence encoding. Similar findings have been reported in several other studies that assessed encoding-retrieval overlap either with univariate analyses similar to those employed by Johnson and Rugg (2007) [e.g. Kahn et al. 2004; Wheeler and Buckner 2004; Woodruff et al. 2005; for a review, see Danker and Anderson (2010)], or with multivoxel pattern analyses [MVPA; e.g. Johnson et al. 2009; Kuhl et al. 2011, 2014; Staresina et al. 2012; Ritchey et al. 2013; Gordon et al. 2014; for a review, see Rissman and Wagner (2012)].

The existence of retrieval-related cortical reinstatement lends support to the proposal that successful episodic retrieval depends on the reactivation of patterns of cortical activity that were encoded as the episode was experienced, the reinstated activity representing the content of the retrieved information [see Rugg et al. (2008) for review]. Crucially, by virtue of how they are operationalized, reinstatement effects reflect only that fraction of encoded content that systematically varies across different classes of study episode. Therefore, if older individuals recollect information that is less diagnostic of a given class of episode than the information recollected by young individuals, their cortical reinstatement effects should be weaker. We assessed this possibility by employing fMRI and a combination of univariate and MVPA analytic approaches. In brief, participants studied a series of pictures and words, performing different study tasks on each class of study item. At test, they made R/K/New judgments on studied and unstudied words. Using the same analytic approach as Johnson and Rugg (2007), we identified regions that demonstrated word- and picture-selective cortical reinstatement effects. We then employed an MVPA classifier that had been trained to discriminate between the 2 study tasks to classify test items according to their study history. We operationalized strength of reinstatement in terms of both the magnitude and extent of univariate reinstatement effects (cf. Johnson and Rugg 2007), and the performance of the classifier (cf. Johnson et al. 2009; Gordon et al. 2014). Since univariate analyses and MVPA do not necessarily converge, we employed both analytic methods to enhance the likelihood of detecting age-related differences in strength of reinstatement [see Davis and Poldrack (2013) for discussion of the rationale and advantages of this approach].

In addition to permitting assessment of age-related differences in retrieval-related cortical reinstatement, the present study also afforded the opportunity to contrast generic recollection-related activity according to age. Three prior fMRI studies that employed the R/K procedure yielded inconsistent findings with respect to the effects of age (Duarte et al. 2008, 2010; Angel et al. 2013). In the cases where recollection estimates were lower in an older subject group [Duarte et al. 2010, and for the “low-performing” older subjects in Duarte et al. (2008); see also Daselaar et al. (2006)], neural correlates of recollection (operationalized by the contrast between items endorsed as R vs. K) were also reduced. In the 2 cases where recollection performance did not significantly differ according to age the findings diverged. In Duarte et al. (2008), recollection effects in “high-performing” older subjects did not significantly differ from those of the young group. In contrast, the effects in the older subjects of Angel et al. (2013) were reduced in several regions, including left angular gyrus, and in the MTL. Thus, the relationship between age, recollection performance, and the neural correlates of recollection remains to be elucidated.

Finally, the present study also allowed assessment of the effects of age on the neural correlates of “post-retrieval monitoring.” This term refers to cognitive operations that support evaluation of the products of a retrieval attempt with respect to their relevance for current behavioral goals (Rugg 2004). Prior studies investigating the effects of age on neural correlates of monitoring have yielded inconsistent findings [for examples of studies that report age effects, see McDonough et al. (2013); Mitchell et al. (2013); Dulas and Duarte (2014); for examples of studies that report null effects of age, see Mark and Rugg (1998); Li et al. (2004); Giovanello et al. (2010); Dulas and Duarte (2013)]. Following Henson et al. (1999; see also Achim and Lepage 2005), we operationalized post-retrieval monitoring through the contrast between studied items attracting K versus R judgments. Henson et al. (1999) reported that this contrast identified enhanced activity in the right dorsolateral prefrontal cortex (DLPFC) and the anterior cingulate, regions strongly implicated in post-retrieval monitoring in later studies [Chua et al. 2009; Hayama et al. 2009; McDonough et al. 2013; for a review of early studies, see Fletcher and Henson (2001), and see Gallo (2013) for a limited review of more recent work]. Henson et al. (1999; see also Henson et al. 2000) proposed that their findings reflected the greater monitoring demands associated with evaluation of the relatively weak memory signal underlying K judgments than with evaluation of the much stronger signal underlying R judgments. In the present study, we assessed whether, when operationalized in the same way, the neural correlates of post-retrieval monitoring differed according to age.

Materials and Methods

Subjects

Twenty-four older subjects (13 females; mean = 67.9 years) and 24 younger subjects (13 females; mean = 23.8 years) were included in the analyses reported below. Subjects participated in the experiment at a compensation rate of $30 an hour. Young subjects were recruited from the University of Texas at Dallas (UT Dallas) student body. Additional young subjects, along with all of the older subjects, were recruited from the surrounding community. All subjects were right-handed, had normal or corrected-to-normal vision, and scored 26 or more on the Mini-Mental State Examination (MMSE). Exclusion criteria included a history of cardiovascular disease (other than treated hypertension), diabetes, psychiatric disorder, illness or trauma affecting the CNS, substance/alcohol abuse, and current or recent use of psychotropic medication or sleeping aids. Additional exclusion criteria included a score of >1.5 SDs below the age normalized score for any memory test or for 2 or more of the non-memory tests on the neuropsychological test battery described below. Informed consent was obtained in accordance with UT Dallas and University of Texas Southwestern Institutional Review Board Guidelines.

Additional subjects for whom data were collected were excluded for the reasons listed below. Three older subjects were excluded because of excess movement during the scan. One older subject was excluded due to scanner malfunction. Seven older subjects were excluded because of too few trials in a critical response category. One young subject was excluded for excess movement in the scanner, 2 young subjects were excluded for falling asleep in the scanner, and another was excluded because of scanner malfunction. A further young subject was excluded because of a structurally abnormal scan. A final seven young subjects were excluded because of insufficient trials in a critical response category.

Twenty of the older subjects who contributed data to the present study, along with 3 of the younger ones, were also included as subjects in a separate study (de Chastelaine et al. 2015). Twelve of these older subjects, and all 3 of the younger subjects, completed the present study first. A minimum of 2 weeks intervened between any subject's participation in the 2 studies.

Neuropsychological Testing

All subjects were administered a standard test battery on a separate day from the MRI scan session, but no more than 6 months in advance of the session. The battery assessed a broad range of cognitive domains (see Table 1 for details of the tests and subjects' performance).

Table 1.

Young and older subjects' characteristics and neuropsychological test performance

	Young group		Old group		P-value
	Mean (SD)	Range	Mean (SD)	Range
Age	23.8 (2.4)	19–29	67.9 (3.5)	63–74
Years of education	16.3 (2.0)	13–22	16.9 (1.8)	13–21	n.s.
Mini-Mental State Examination	29.4 (1.1)	26–30	29.6 (0.5)	29–30	n.s
CVLTa immediate free recall	12.2 (2.1)	8–16	9.9 (2.6)	7–16	<0.005
CVLTa immediate cued recall	13.1 (1.7)	8–16	12.0 (2.2)	8–16	<0.05
CVLTa delayed free recall	12.8 (2.3)	7–16	11.4 (2.6)	7–16	n.s.
CVLTa delayed cued recall	13.5 (1.9)	8–16	12.0 (2.3)	8–16	<0.05
CVLTa delayed recognition	15.3 (0.9)	13–16	14.4 (1.3)	12–16	<0.05
WMS-IVb immediate recall	30.6 (6.0)	17–46	29.0 (4.9)	22–39	n.s.
WMS-IVb paragraph delayed recall	26.8 (7.0)	15–49	25.8 (5.3)	15–36	n.s.
Forward/backward Digit Span	17.8 (4.3)	11–27	17.5 (3.8)	12–24	n.s
Digit/Symbol substitution test	65.5 (12.6)	48–88	49.6 (8.7)	31–69	<0.001
Trail Making test A (s)	23.3 (6.6)	16–46	33.1 (9.8)	18–58	<0.001
Trail Making test B (s)	48.6 (12.8)	31–78	73.3 (35.7)	37–201	<0.005
Letter fluency	43.0 (11.2)	21–65	44.6 (12.6)	21–70	n.s.
Category fluency	27.1 (4.8)	20–38	21.8 (6.9)	12–37	<0.005
WTARc Raw	43.1 (4.4)	33–49	43.0 (4.6)	34–49	n.s.
Raven's Progressive matricesd	11.0 (1.1)	9–12	9.3 (2.3)	4–12	<0.005

^aCalifornia Verbal Learning Test.

^bWechsler Memory Scale (WMS-IV).

^cWechsler Test of Adult Reading Full Scale Intellectual Quotient.

^dShort version of Raven's Progressive Matrices.

Experimental Materials and Procedure

The experimental items comprised 216 colored pictures of everyday objects and their corresponding names. The pictures were drawn from Hemera Photo Objects 50 000 Volume 2 (http://hemera-technologies-inc.software.informer.com). The names of each picture were between 3 and 12 letters long, with a mean written frequency of 14.3 counts per million (Kucera and Francis 1967). In addition to the picture-word sets described above, 15 additional words and 9 additional pictures were used as fillers for the study and test lists. A further 18 words and 12 pictures were used for 2 study–test practice lists. For each subject, 144 of the 216 picture-name sets were randomly selected as members of the study list. Study items were randomly assigned to 1 of 3 study sessions, with the constraint that for each study block, items were counterbalanced so that there were an equal number of items that would elicit a correct answer for each response category described below. For each study session, half of the study items were shown as pictures and the remaining items were shown as words. At test, the names of the remaining 72 items from the item pool were used as new items, while items that were previously presented as words were re-presented, and items that were previously presented as pictures were denoted by their names (Fig. 1B). The orderings of the study and test lists were pseudorandomized such that there were no more than 3 consecutive presentations of items belonging to the same experimental condition or that were likely to elicit the same response. Study and test items were presented using Matlab2012b (Mathworks, Inc., USA) as implemented in the Cogent software package (http://www.vislab.ucl.ac.uk/cogent.php). The items were back-projected onto a screen and viewed via a mirror mounted on the scanner headcoil.

Figure 1.

Experimental task. (A) Schematics of picture (i) and word (ii) encoding mini-blocks. (B) Retrieval task.

The experiment comprised 3 consecutive study sessions followed by 3 test sessions, which between them comprised a single study–test cycle. The study and test phases were both scanned. Instructions and practice were administered outside the scanner. The 3 study sessions were separated by brief rest periods (∼2 min) while the scanner was restarted.

During study, subjects were required to perform 1 of 2 tasks, contingent upon item type (see Fig. 1 for a schematic of study and test procedures). For pictures, the task requirement was to judge whether the depicted item would fit into a shoebox. For words, the judgment was whether the denoted object would be more likely to be found inside or outside a house. As is evident from Supplementary Figure 1, the 2 tasks were associated with very different patterns of cortical activity, making them suitable for the study of retrieval-related reinstatement effects. To optimize the training of the MVPA classifier (see below), study items were presented in each study session in “mini-blocks” (cf. Johnson et al. 2009; McDuff et al. 2009). Each mini-block began with a cue question (“Used inside of a house?” or “Fits inside a shoebox?”) presented for 2000 ms, followed by 3 trials of the same type. The first 2 trials of each mini-block began with the presentation of a red fixation cross in the middle of the display frame for 500 ms. This was replaced with an item that was presented for 2000 ms, and replaced in turn by a black fixation cross for 1500 ms. On the third trial, the black fixation cross was presented for 2000 ms. For each task, “yes” and “no” judgments were assigned to the left and right index fingers, with the assignment counterbalanced across subjects. Subjects completed a short practice prior to the experiment proper.

Approximately 5 min after the conclusion of the last study session, a short practice session on the test task was undertaken. Following the practice session, the 3 test sessions were run. In each session, studied words and the names of studied objects were presented, intermixed with new words. Each trial began with a red fixation cross presented in the center of the gray frame for 500 ms, followed by the presentation of a word at fixation for 3000 ms. Following word offset, a black fixation cross was presented for 4500 ms (∼66% of trials), 6500 ms (∼25% of trials), or 8500 ms (∼9% of trials), giving intertrial intervals of 8, 10, and 12 s, respectively. Subjects were instructed to respond on each trial prior to the onset of the red fixation cross. The task was to judge whether a word or the picture denoted by the word had previously been presented, using 1 of 3 response options: “Remember,” “Know,” and “New.” Remember or R judgments were to be used when recognition was accompanied by retrieval of a specific detail or details from the study episode. Know or K judgments were defined as items that were confidently recognized in the absence of retrieval of any detail about the study episode. New judgments were to be given when an item was judged unstudied, or the subject was uncertain as to its study status. Subjects responded with the index and middle fingers of one hand, and the index finger of the opposite hand. Hand and finger assignments were counterbalanced across subjects.

To ensure subjects understood the test instructions, they were required to verbally describe the basis of each R response given in the practice session that preceded the test session. It was emphasized that recollection of any contextual detail, and not only the identity of the study material or study task, justified an R judgment.

fMRI Data Acquisition

Functional and anatomical images were acquired with a Philips Achieva 3-T MR scanner (Philips Medical System, Andover, MA, USA) equipped with a 32-channel parallel imaging head coil. Functional scans were acquired with a T₂*-weighted echo-planar image (EPI) sequence with the following parameters: time repetition (TR) 2 s, time echo (TE) 30 ms, flip angle 70°, field of view (FOV) 240 × 240, matrix size 80 × 79. Each EPI volume comprised 30 3 mm slices (1 mm interslice gap). Slices were acquired in ascending order, oriented parallel to the AC–PC line, and positioned for full coverage of the cerebrum and most of the cerebellum. The functional data were acquired using a sensitivity encoding (SENSE) reduction factor of 2. Data were acquired during both study and test phases (170 and 338 volumes for each study and test sessions, respectively). The first 4 volumes of each block were discarded to allow tissue magnetization to achieve a steady state. For the purposes of the general linear model (GLM) analyses, study and test sessions were concatenated to give one study and one test time-series, respectively. The study and test data were modeled and estimated separately.

A T₁-weighted anatomical image was acquired with a 3D magnetization-prepared rapid gradient echo (MP-RAGE) pulse sequence (FOV = 240 × 240, matrix size 220 × 193, voxel size 1 mm³, 150 slices, sagittal acquisition).

fMRI Data Analysis

fMRI data were analyzed using a mass univariate, GLM, and MVPA. For the purposes of the GLM analyses, EPI images were preprocessed and analyzed using Statistical Parametric Mapping (SPM8, Wellcome Department of Cognitive Neurology, London, UK) implemented under MatlabR2008b (The Mathworks, Inc., USA). Functional volumes were spatially aligned to the mean image across sessions, and then manually reoriented to approximate the orientation of the Montreal Neurological Institute (MNI) reference brain. Volumes were slice-time corrected to the middle slice and normalized to a sample-specific template [based on the MNI (MNI) template brain, see de Chastelaine et al. (2011) and Mattson et al. (2014) for description of the template construction]. Each subject's realigned volumes were normalized with respect to the sample-specific template and resampled into 3 mm isotropic voxels. Normalized volumes were smoothed with an isotropic 8-mm FWHM Gaussian kernel to reduce the effects of residual across-subject and across-group anatomical variation. For each voxel, the EPI time-series was high-pass-filtered to 1/128 Hz and scaled within the session to yield a grand mean of 100 across voxels and scans. T₁-anatomical images were normalized with a procedure analogous to that applied to the EPI images. Normalized T₁ images were resampled into 1 mm isotropic voxels.

Mass univariate statistical analyses were performed in 2 stages of a mixed-effects model. BOLD responses were modeled by convolving delta functions concurrent with the onset of each study and test item with a canonical hemodynamic response (HRF) and its temporal and dispersion derivatives. This procedure yielded GLM regressors that modeled the BOLD response on every trial. The study analysis was confined to 2 events of interest: Words presented in the location task (Word) and pictures presented in the size task (Pic). Other study events (i.e. fillers and missed responses) were modeled as events of no interest. Five events of interest were modeled at test: (1) Accurate R responses to words previously presented as a word (Rword), (2) accurate R responses to words previously presented as a picture (Rpic), (3) accurate K responses to words previously presented as either words or pictures (K), (4) new responses to items previously presented as either words or pictures (MISS), and (5) correctly identified new words (CR). All other events, including false alarms, fillers, and trials where a response was not given, were modeled as events of no interest. Both the study and test analyses also incorporated 6 regressors to model residual movement-related variance (3 rigid-body translations and 3 rotations). Constants modeling means across each session were also entered into each design matrix.

Effects of interest were carried forward to a random-effects analysis that applied linear contrasts to the subject-specific parameter estimates. Unless otherwise specified, effects are reported at an uncorrected threshold of P < 0.001 (one-sided) with a 22-voxel cluster extent threshold, giving a corrected cluster-wise threshold of P < 0.05 across the whole brain. The cluster extent threshold was estimated using Monte Carlo simulations (http://afni.nimh.nih.gov/pub/dist/doc/program_help/AlphaSim.html) with 10 000 iterations.

Regions of overlap between the outcomes of 2 contrasts were identified by inclusive masking of the relevant SPMs. Assuming independence of the 2 contrasts, the conjoint significance of the resulting SPM was estimated with Fisher's method (Lazar et al. 2002). Exclusive masking was used to identify voxels where effects were not shared between 2 contrasts; in these cases, the mask threshold was set at P < 0.05 one-sided. Note that the more liberal the threshold of an exclusive mask, the more conservative is the masking procedure.

MVPA Preprocessing

For the purposes of the MVPA, functional MRI data were preprocessed as described above, but excluding the smoothing step. The data were then de-trended to remove linear and quadratic trends, and z-scored across voxels within each scanning session.

MVPA

MVPA was conducted using the Princeton MultiVoxel Pattern Analysis toolbox (Computational Memory Laboratory, Princeton, NJ, USA; www.pni.princeton.edu/mvpa) and SPM8 (Wellcome Department of Cognitive Neurology). A classifier was trained on TRs 3 through 8 of each study mini-block to discriminate between word and picture encoding trials (see below) and was then employed to classify the test trials associated with different classes of memory judgment. The accuracy of the classifier in identifying a test item's encoding history (i.e. word or picture) was used to operationalize the strength of cortical reinstatement. Classification was implemented with regularized logistic regression (L2 penalty).

MVPA was used to quantify the strength of cortical reinstatement rather than to demonstrate the existence of the phenomenon [cf. Johnson et al. 2009; see Gordon et al. (2014) for a similar approach]. Accordingly, the voxels selected for input to the classifier were constrained to belong to cortical regions demonstrating reinstatement effects as operationalized through the parallel univariate analyses (see below). To ensure that the voxels were restricted to gray matter, word and picture reinstatement effects were inclusively masked with the default SPM gray matter probability map (thresholded at P < 0.2). To determine which voxels would enter the study phase classifier, we selected the 500 voxels with the highest T-values for the word and picture reinstatement effects, respectively, and aggregated the 2 populations to create a 1000 voxel region of interest (ROI) that served as the “feature selection ROI” for classifier training. Importantly, feature selection was performed in a manner unbiased with respect to any given subject. For each of 24 sets of randomly paired older and younger subjects, the feature selection ROI was determined by a group-level analysis of the cortical reinstatement effects on the data from the remaining 46 subjects (23 older and 23 young). Therefore, for any given subject, the classifier was trained and tested on voxels that were selected using data independent of those that constituted the training and test sets.

Calculation of Classification Accuracy

For any given test trial, classification was deemed accurate if the classifier returned a probability estimate for the correct condition that exceeded 0.5 (chance classification). Accuracy was binarized to give a score of 1 for each correct classification, and 0 for each incorrect classification. Accuracy was assessed for the TR corresponding to the onset of each item (TR 1) and for the 6 subsequent TRs.

Study Phase Classification

The study phase classification was implemented using a 3-fold cross-validation procedure (Hastie et al. 2001). Thus, data from 2 of the 3 study blocks were used to train the classifier, while the remaining study block was used to test it. The process was repeated 3 times, so that the classification accuracy represented average accuracy across the 3 iterations. Following Johnson et al. (2009), we designated the TR corresponding to the onset of the first item of a study mini-block as TR 1. The classifier was trained on TRs 3–8, with a penalty of 0.05.

Visualization of Results

For the purposes of visualizing the univariate findings, Caret software (Van Essen et al. 2001) was used to map fMRI effects of interest onto inflated fiducial brains derived from the PALS-B12 atlas (Van Essen 2002, 2005) in SPM5 space. Results were also visualized using sections from the across-subjects averaged T₁-structural image.

Correction for Nonsphericity and Multiple Comparisons

Nonsphericity between the levels of repeated-measures factors in the ANOVAs reported below was corrected with the Greenhouse–Geisser procedure (Greenhouse and Geisser 1959). Correction for multiple comparisons of the significance levels of t-tests and correlation coefficients was effected with the Holm–Bonferroni procedure (Holm 1979).

Results

Neuropsychological Data

Demographic information and neuropsychological test performance for older and younger subjects are summarized in Table 1. The 2 groups of subjects were equated for years of education and performed equally well on tests typically found to be relatively insensitive to age (e.g. Digit Span and Letter Fluency). Long-term memory for word lists as measured by the California Verbal Learning Test was significantly lower in older subjects, whereas performance on the WMS-IV logical memory tests did not differ with age. As is typical, older subjects demonstrated poorer performance on tests of speeded cognition, such as Trail Making A&B and Digit Symbol substitution. Finally, whereas the 2 age groups did not differ on a putative measure of crystallized intelligence (written word pronunciation), fluid intelligence, as assessed by Ravens Progressive Matrices, was significantly lower in the older group. In summary, the neuropsychological profiles of our young and older subject groups are typical of those reported in numerous previous studies of individuals drawn from similar populations (e.g. Park et al. 1996; Duarte et al. 2010; Salthouse 2010; de Chastelaine et al. 2011). Therefore, there is no reason to presume that the behavioral or fMRI findings described below reflect idiosyncratic characteristics of our subject samples.

Behavioral Results

Study Phase

Accuracy was defined as the consistency between the study judgments of individual subjects with that of the judgments of 3 raters (if there were disagreements among the raters, the majority judgment was considered correct). In light of the subjective element involved in these judgments, the accuracy data should be treated with caution. Nonetheless, accuracy of the young subjects was high for both picture (0.95, SD = 0.04) and word (0.85, SD = 0.03) tasks. Accuracy on the study task was also high for the older subjects, albeit slightly lower than in the case of the young subjects (0.91, SD = 0.06, and 0.81, SD = 0.07, for the word and picture tasks, respectively). In light of the foregoing caveat about the subjectivity of the accuracy scores, we did not subject the data to inferential statistics. Given their high scores, it is clear, however, that both young and older subjects complied with the study instructions.

Mean study phase reaction times (RTs) for accurate trials were segregated by study material. To assess whether RT varied with material or age, a 2 × 2 ANOVA with factors of material (word and picture), and age group (young and old), was performed. The ANOVA revealed an interaction between the 2 factors (F_1,46 = 6.46; P < 0.05). Follow-up contrasts revealed that young subjects responded significantly more slowly on word (1322 ms, SD = 286 ms) than on picture trials (1264 ms, SD = 298 ms; P < 0.01), whereas older subjects did not demonstrate a significant RT difference between material types (Words: 1295 ms, SD = 266 ms; Pictures: 1314 ms, SD = 269 ms).

Test Phase

The proportions of studied and unstudied items endorsed as R, K, and New are given in Table 2. Strength of recollection (pR) was estimated as the probability of an R hit minus that of an R false alarm. Familiarity strength was estimated under the assumption that R and K judgments are independent, giving the formula: pF = (pKhit/1 − pRhit) – (pKfalse alarm/1 − pRfalse alarm) (Yonelinas and Jacoby 1996). Separate estimates were obtained for items studied as words and pictures (Table 3). A 2 × 2 ANOVA of the recollection estimates revealed main effects of age (F_1,46 = 4.70; P < 0.05) and study material (F_1,46 = 5.71; P < 0.05), with no interaction. The results reflected higher pR for young subjects (0.59 vs. 0.49 for young and older groups, respectively), and for items studied as pictures rather than as words (0.57 vs. 0.51, respectively). ANOVA of the familiarity estimates revealed a main effect of study material (F_1,46 = 74.03; P < 0.001), but no effect of age or the age × material interaction. The material effect reflected greater familiarity for items studied as words than as pictures (0.57 vs. 0.38, respectively).

Table 2.

Mean (and standard deviation) RTs (ms) of R, K, and New responses to studied and unstudied items in young and older subjects, respectively

Study status	Studied words			Studied pictures			New items
Response	R	K	New	R	K	New	R	K	New
Young	0.59 (0.14)	0.30 (0.13)	0.11 (0.07)	0.65 (0.14)	0.18 (0.10)	0.17 (0.10)	0.03 (0.03)	0.11 (0.07)	0.86 (0.09)
Older	0.51 (0.24)	0.33 (0.18)	0.16 (0.11)	0.58 (0.20)	0.20 (0.11)	0.22 (0.13)	0.05 (0.04)	0.12 (0.10)	0.83 (0.13)

Table 3.

Mean (and SD) of recollection and familiarity estimates by study material and age group

	Word		Picture
	Recollection	Familiarity	Recollection	Familiarity
Young	0.56 (0.14)	0.61 (0.19)	0.62 (0.14)	0.40 (0.19)
Older	0.46 (0.23)	0.53 (0.15)	0.53 (0.19)	0.36 (0.14)

RTs for R, K, and New responses for studied and unstudied items are summarized in Table 4. These data were subjected to a three-way ANOVA with factors of material (picture and word), memory judgment (R and K), and age group (young and old). The ANOVA revealed a main effect of study material (F_1,46 = 18.53; P < 0.001), a main effect of memory judgment (F_1,46 = 90.98, P < 0.001), an interaction between study material and judgment (F_1,46 = 9.26, P < 0.005), but no main effect of age or any interactions with age. Follow-up contrasts revealed that these effects reflected longer RTs for items eliciting a K than an R response (3002 vs. 2118 ms; P < 0.001), and longer RTs for items that had been studied as pictures than words (2623 vs. 2497 ms; P < 0.001). The study material × memory judgment interaction took the form of roughly equivalent RTs for remembered items across study material (2110 and 2126 ms for words and pictures, respectively), but significantly shorter RTs for familiar items studied as words than as pictures (2883 vs. 3120 ms, respectively; P < 0.001). (Note that although for simplicity we refer to the 2 encoding conditions as the “word” and “picture” tasks, study material was perfectly confounded with study task (indoor/outdoor and size judgments). Thus, we cannot separate the influences of material, task, or the interaction between these 2 variables, on the encoding-related or reinstatement effects reported below.)

Table 4.

Mean (and SD) RTs (ms) of R, K, and New responses to studied and unstudied items in young and older subjects, respectively

Study status	Studied words			Studied pictures			New items
Response	R	K	New	R	K	New	R	K	New
Young	2126 (719)	2969 (859)	3246 (855)	2155 (706)	3152 (773)	3260 (958)	2155 (993)	3475 (834)	2921 (849)
Older	2095 (707)	2797 (761)	3263 (902)	2097 (679)	3089 (756)	3211 (895)	1656 (748)	3124 (739)	2882 (891)

fMRI Data: Generic Recollection Effects

To identify recollection effects common to the 2 classes of study conditions [“material-independent effects,” we inclusively masked the contrast between items endorsed R and K (R > K; thresholded at P < 0.001)] with the separate contrasts for each class of study material (Rword > K and Rpic > K, each thresholded at P < 0.05; due to low trial numbers, K judgments were collapsed across material to give a single estimate of the activity associated with these judgments). The outcome of this procedure is an SPM where recollection effects collapsed across study material are both reliable and are not driven exclusively by only one class of material.

The resulting SPM (Fig. 2A), which is unbiased with respect to age group, was used to extract the parameter estimates employed in the ANOVA. Since we were unable to identify any clusters where these recollection effects were modified by age (see below), the figure also depicts the regions where the effects did not differ in magnitude between the 2 age groups. Figure 2B depicts the parameter estimates for trials associated with R and K judgments derived from selected regions within the core recollection network (see Table 5 for a list of other age-invariant generic recollection effects). The estimates are across-voxel averages within 3 mm radius spheres centered on the peak of the recollection effect in each ROI [left angular gyrus, posterior cingulate/retrosplenial cortex, hippocampus and medial prefrontal cortex (mPFC); hippocampal estimates were collapsed across anterior and posterior peaks and hemispheres].

Figure 2.

(A) Generic age-invariant recollection effects (R > K). Roman numerals identify regions from where the parameter estimates depicted in B were extracted. (B) Parameter estimates from the 4 regions identified in A above. (i) left angular gyrus (−45, −76, 28), (ii) left posterior cingulate (−15, −55, 34), (iii) mean of the parameter estimates from hippocampal peaks (−27, −22, −20; −24, −10, −23; 36, −16, −23, and 21, −7, −23), and (iv) mPFC (−6, 65, 13).

Table 5.

Regions demonstrating material-independent age-invariant recollection effects^a

Region	Peak voxel [X Y Z]	Z-value
L mPFC	−6 65 13	Inf
L posterior cingulate	−15 −55 34	Inf
L angular gyrus	−45 −75 28	Inf
L hippocampus	−27 −22 −20	7.61
L SMA	−15 −28 70	7.04
L inferior temporal gyrus (posterior)	−63 −46 −14	6.99
R SMA (including middle cingulate cortex)	15 −19 67	6.86
R cerebellum	27 −73 −44	6.64
L precentral gyrus	−42 5 49	6.50
R hippocampus	21 −7 −23	6.33
R postcentral gyrus	51 −7 34	6.29
L middle orbital gyrus	−33 41 −14	6.20
L inferior temporal gyrus (anterior)	−54 2 −35	6.16
L postcentral gyrus	−51 −10 37	6.14
R postcentral gyrus	66 −4 16	6.03
R inferior temporal gyrus	57 −52 −5	5.85

^aTo simplify the table, peaks are reported only for effects where Z > 5.13.

A reviewer noted that because R endorsements were more common in the picture condition, and K endorsements were more common in the word condition, collapsing across the 2 encoding conditions as described above biased the resulting R and K effects in favor of the picture and word conditions, respectively. Therefore, we repeated these analyses after removing (by random selection) from the first-level GLMs “excess” R and K trials from the picture and word study conditions, respectively (limitations on trial numbers meant that, for K trials, this was possible for only 19 young and 23 older subjects). The outcomes of these analyses were very similar to those reported below for the full subject samples other than for the emergence of a single small cluster (55 voxels) on the border of occipital and posterior cingulate cortex, where recollection effects were greater in the older group.

Age-Related Differences in Recollection-Related Effects

We searched at the whole-brain level for clusters demonstrating age-related differences in recollection-related effects using the age × recollection interaction contrast (P < 0.001 for each side of the contrast). No significant clusters could be identified. This remained the case even when the threshold for this contrast was lowered to P < 0.05 (two-sided).

Inspection of the parameter estimates illustrated in Figure 2B suggests a trend toward more positive-going estimates in older than in younger subjects. To assess the reliability of this trend, we performed a 2 (group) × 4 (region) × 2 (memory judgment) ANOVA. While the presence of a reliable memory judgment effect is a foregone conclusion given how these regions were selected, the main effect of group and its interaction with region are both free to vary. Both of these effects were significant (group: F_1,46 = 11.25, P < 0.005; group × region: F_{2.39, 115.20} = 4.17, P < 0.05). Follow-up ANOVAs revealed reliable group effects in the mPFC and hippocampus (Fs = 16.19 and 5.59, respectively), but not in the posterior cingulate or left angular gyrus.

Finally, we assessed whether the magnitude of the recollection effects in regions belonging to the core recollection network (see Introduction) correlated across subjects with recollection performance, as this is operationalized by pR. The regions and the respective parameter estimates included in this analysis are detailed in Figure 2. Correlations were performed employing age as a covariate. In only 1 case (the composite hippocampal effect), was there any evidence for a relationship with performance (partial r = 0.295, P < 0.05); however, the correlation did not survive correction for multiple comparisons across the 4 regions.

Reinstatement Effects

Cortical reinstatement effects were identified by searching for regions where material-specific encoding and recollection effects overlapped (cf. Johnson and Rugg 2007). To achieve this, we first exclusively masked each material-specific recollection-contrast (Rpic > K and Rword > K, each thresholded at P < 0.01) with the contrast defining the effects for the alternate material (P < 0.05), thus eliminating voxels sharing the 2 effects. The resulting SPMs were then inclusively masked with their respective material-specific encoding contrasts (i.e. Pic > Word, and Word > Pic, each thresholded at P < 0.001; see Supplementary Fig. 1). To eliminate any voxels where reinstatement effects varied according to age group, we exclusively masked each effect with the interaction between the corresponding material-selective recollection effects and age (in each case, thresholding the interaction contrast at P < 0.05 two-sided). The results of these analyses are illustrated in Figure 3 and detailed in Table 6. As is evident from the figure, age-invariant word-specific reinstatement was evident in posterior parietal cortex, posterior cingulate, and lateral prefrontal cortex, whereas picture-specific reinstatement was evident primarily in the fusiform cortex. (Note that the posterior cingulate and the posterior parietal regions that demonstrated word-specific reinstatement did not overlap with adjacent regions that demonstrated material-independent effects.)

Figure 3.

Age-invariant reinstatement effects for recollected words (A) and recollected pictures (B).

Table 6.

Regions demonstrating age-invariant reinstatement effects

Region	Peak voxel [X Y Z]	Z-value	Cluster size
Word reinstatement
Posterior cingulate	0 −46 28	5.76	308
L angular gyrus	−45 −70 40	5.50	102
L orbital frontal gyrus	−48 32 −20	4.68	25
L middle frontal gyrus	−30 20 43	4.46	116
L dorsal medial frontal gyrus	−9 29 55	4.23	56
L orbital frontal gyrus	−57 29 −8	4.09	43
R angular gyrus	51 −76 34	3.92	92
R middle temporal gyrus	66 −25 −11	3.23	27
Picture reinstatement
L occipito-temporal	−54 −61 −11	6.26	359
L precentral gyrus	−48 −1 25	5.31	44
R inferior frontal gyrus	45 38 7	5.18	74
R inferior temporal	54 −49 −20	5.17	122
L insula	−36 −4 10	4.94	22
L middle occipital gyrus	−24 −67 37	4.51	116
R middle occipital gyrus	30 −64 37	4.46	61
L inferior parietal cortex	−66 −31 31	4.43	89
R inferior frontal gyrus	48 8 19	4.28	179
R hippocampus	18 −19 −11	3.51	91
R superior occipital gyrus	21 −58 49	3.12	30

Effects of Age on the Magnitude of Reinstatement Effects

We searched for regions demonstrating differences between the age groups in the magnitude of reinstatement-related BOLD signal change by inclusively masking the interactions between material-selective recollection effects (thresholded at P < 0.01, two-sided) and age with the respective encoding contrasts (Pic > Word and Word > Pic, each thresholded at P < 0.001). In neither case were any significant clusters identified, and thus no further analyses were warranted. Below, we describe how MVPA was employed to further address the question whether reinstatement effects differed according to age.

MVPA

As was noted in the Methods section, an MVPA classifier was employed on each subject's data to quantify strength of reinstatement (see Introduction). As already described, the voxels from the study phase employed to train each subject's classifier were constrained to regions demonstrating recollection-related reinstatement effects, but were selected on the basis of across-subjects analyses that excluded the subject's own data. Mean accuracy of the study classifier was high, reaching 0.85, and exceeding chance (0.5) at all TRs (one-tailed t-test, corrected for multiple comparisons; see Supplementary Fig. 2). A 2 (age group) × 7 (TR) ANOVA on the study classification data revealed a main effect of TR (F_{3.0, 139.5} = 249.40; P < 0.001), but no effect of age or an age × TR interaction.

For each subject, the trained study classifier was first employed to classify each R judgment according to the study history of the test item. The outcome of this procedure is summarized in Figure 4A. As is evident from the figure, classifier accuracy in both age groups was above chance for TRs 3–6, and also showed above-chance accuracy at TR 7 for the older group. The peak accuracy for both groups was at TR 4. A 2 (age group) × 7 (TR) ANOVA revealed a main effect of TR (F_{4.6, 213.3} = 33.16; P < 0.001), but no effect of age group and no interaction between age group and TR (min P > 0.1). Thus, there was little evidence that recollection-related reinstatement effects, as operationalized by classifier accuracy, differed according to age. [It should be noted, however, that when classifier accuracy was contrasted across groups for each TR separately, there was a significant difference at TR 3 in favor of the young subjects; t₍₄₆₎ = 2.01, P < 0.05. This effect did not survive correction for multiple comparisons.] Equivalent findings were obtained when the continuous variable of classifier evidence, rather than binarized accuracy, was analyzed (see Supplementary Fig. 3).

Figure 4.

MVPA classification accuracy for test items [(A) old items endorsed Remember, (B) old items endorsed Know, and (C) Missed items] according to TR and age group (blue and red bars denote young and older groups, respectively). Dotted line denotes chance performance. Asterisks signify above-chance classification accuracy after correction for multiple comparisons.

Following Johnson et al. (2009), we also assessed whether the study classifier could discriminate K trials according to their study history (as already noted, because of limitations on trial numbers, this analysis could be conducted on only a subset of the subjects in each age group). As is illustrated in Figure 4B, classifier performance was significantly above chance for TRs 4 and 6 for the young subjects, and for TRs 5 and 6 for the older group. An ANOVA gave rise to a main effect of TR (F_{5.1, 234.8} = 6.08, P < 0.001), no main effect of group, and a near- significant group × TR interaction (F_{5.1, 234.8} = 2.23, P < 0.06). Pairwise contrasts between the groups at each TR revealed no significant differences (min P = 0.09). As for the R judgments, results did not change when classifier evidence, rather than accuracy, was employed as the dependent variable.

We also assessed whether the classifier was able to discriminate the encoding class of studied items given a new response (“Miss” trials). As is illustrated in Figure 4C, the classifier failed to perform above chance at all TRs.

Finally, to determine whether reinstatement varied according to memory judgment, we directly contrasted classifier accuracy for R and K judgments using a 2 (age group) × 2 (response category) × 7 (TR) ANOVA. The ANOVA revealed a significant main effect for TR (F_{4.4, 202.8} = 22.3, P < 0.001), and significant interactions between group and TR (F_{4.4, 202.8} = 2.47, P < 0.05) and response category and TR (F_{5.2, 237.4} = 5.12, P < 0.001). The three-way interaction did not approach significance (P > 0.1). Follow-up contrasts revealed no group differences at any TR, but significant differences between response categories (greater accuracy for R judgments) at TRs 3 and 4 (P < 0.025 and <0.001, respectively). The latter of these effects survived correction for multiple comparisons. An identical pattern of findings was revealed when classifier output was subjected to the same analysis.

The relationship between recollection performance and classifier accuracy for R judgments was assessed at each TR with partial correlation, employing age as a covariate. The 2 highest correlations were for TRs 5 and 6 (r = 0.316 and 0.299, respectively, P < 0.05); neither correlation survived correction for multiple comparisons.

Post-Retrieval Monitoring

In a final analysis, we identified and quantified putative neural correlates of post-retrieval monitoring–cognitive operations that support evaluation of the products of an episodic retrieval attempt with respect to their relevance for current behavioral goals (Rugg 2004). As noted in the Introduction, we operationalized monitoring by the contrast between studied items attracting K versus R judgments, that is, between test items eliciting weak and strong evidence diagnostic of recollection, and hence varying in their monitoring demands. [Note that the absence of reliable interaction effects between the factors of memory judgment (R vs. K) and age means that this analysis also identified regions where these putative monitoring effects were age-invariant.] To ensure that the effects identified by the contrast were not driven exclusively by only one class of study material, the contrast was masked by the 2 simple effects (i.e. K > Rword and K > Rpic, thresholded at P < 0.05). The procedure identified 4 clusters (Table 7 and Fig. 5A), including 2 (right RDLPFC and anterior cingulate) with peaks that were within a few mm of those reported by Henson et al. (1999).

Table 7.

Regions demonstrating age-invariant effects for the K > R contrast

Region	Peak voxel [X Y Z]	Z-value	Cluster size
R insula	33 26 −5	5.53	149
R anterior cingulate	9 29 31	5.04	122
R inferior frontal gyrus	24 53 10	5.02	73
R dorsolateral PFC	51 35 34	4.89	107

Figure 5.

(A) Regions demonstrating greater activity for items endorsed K than for those endorsed R. (B) Partial Plot of the relationship between recognition performance and the K–R differences in the anterior cingulate, controlling for age group, and the differences in RT for the respective response categories. Blue and red dots denote young and older subjects, respectively.

The relationship between the aforementioned “monitoring effects” and recollection performance was assessed with partial correlation, employing age group as a covariate. The difference in RT between K and R judgments was employed as an additional covariate since there was a marked difference in RT for the 2 classes of judgments (see Behavioral Results) and these prefrontal regions—the anterior cingulate in particular—have been implicated in processes likely to vary with RT, such as response selection [e.g. for a review, see Rushworth et al. (2007); Holroyd and Coles (2008); Chaudhry et al. (2009); Zhang et al. (2012)] and “time on task” (e.g. Grinband et al. 2011; Weissman et al. 2013). After correction for multiple comparisons (correlations were performed for each of the 4 regions identified by the contrast), only the correlation between pR and the anterior cingulate cortex was reliable [r₍₄₄₎ = 0.541, P < 0.001; Fig. 5B]. When this correlation (controlling for RT differences) was performed for each age group separately, it remained reliable after correction in both cases [old: r₍₂₁₎ = 0.609, young: r₍₂₁₎ = 0.516, P < 0.0125].

Discussion

Successful recollection was associated with enhanced activity in several regions belonging to the “core recollection network” (Rugg and Vilberg 2013). There were no differences in the magnitude of any of these recollection effects according to age. Material-specific cortical reinstatement effects associated with successful recollection were also robust but age-invariant, regardless of whether the effects were contrasted in a univariate analysis, or with MVPA. Finally, we identified putative correlates of post-retrieval monitoring in frontal regions implicated previously in monitoring processes; these effects also did not differ reliably across age groups. Below, we discuss the implications of these findings for the understanding of the effects of age on episodic memory retrieval, as well as for the understanding of the neural correlates of episodic retrieval more generally.

Behavioral Results

As detailed in the Introduction section, it has been reported that the R/K procedure is less sensitive to age-related memory decline than “objective” indices of recollection. Nonetheless, our older sample demonstrated significantly lower recollection estimates than our young subject group, as has been reported previously for this test procedure [for reviews, see McCabe et al. (2009); Koen and Yonelinas (2014)]. Older and younger subjects did not, however, differ in their familiarity estimates, replicating some previous reports (e.g. Jennings and Jacoby 1993, 1997; Titov and Knight 1997; Caldwell and Masson 2001; Howard et al. 2006), but contrary to others (e.g. Prull et al. 2006; Toth and Parks 2006; Parks 2007; Duarte et al. 2010; Duzel et al. 2011; Wang et al. 2012). It has been argued that an important factor explaining this disparity is the test procedure employed to obtain the familiarity estimates, with the R/K procedure more likely to lead to reduced estimates for older individuals than other procedures (Koen and Yonelinas 2014). This cannot be the sole explanation, however; for instance, an R/K procedure was employed in both Wang et al. (2012) and the present study, yet familiarity demonstrated a reliable age effect in the former study only.

Although not relevant to the main focus of the experiment, a further aspect of the present behavioral findings is worthy of note. Regardless of age group, recollection estimates were higher for test items initially studied as pictures, whereas familiarity estimates demonstrated the reverse pattern (an ANOVA employing the factors of study material and memory type revealed a reliable cross-over interaction; F_1,46 = 50.15; P < 0.001). These findings presumably reflect the combination of a picture superiority effect for recollection (e.g. Ally et al. 2008; Curran and Doyle 2010), and the benefit to familiarity that arises when the perceptual overlap between study and test items is high rather than low (e.g. Jacoby and Dallas 1981). The interaction reflects the type of dissociation that is arguably incompatible with models of recognition memory in which familiarity and recollection depend on a common memory signal (Dunn 2004). Thus, the findings lend support to the validity of the two-process model adopted here.

A limitation of our experimental design is that we did not assess source memory in our subjects, allowing a comparison of indices of source recollection with the recollection estimates obtained from the Remember/Know procedure. We considered and rejected the option of requiring both types of judgment on each trial (cf. Duarte et al. 2006, 2008), since we did not wish to complicate our test procedure with a double response requirement. Such a requirement would have approached and, in some cases, exceeded the capacity of our older subjects in what was already a lengthy experimental session. An alternative to this option would have been to run a separate behavioral session that directly paralleled the fMRI study, but that included source memory rather than R/K judgments. Limitations on the availability of suitable experimental materials (we would have needed to double the present number of unique items) made this option impractical, however.

fMRI Findings

Material-Independent Recollection Effects

Material-independent recollection effects (greater activity for items endorsed R than K) were evident in numerous cortical regions, as well as in the MTL (Fig. 2A and Table 5). The same regions have been consistently identified in prior studies of the neural correlates of successful recollection that employed a wide variety of different experimental materials and test procedures [for reviews, see Spaniol et al. (2009) and Kim (2010)]; several of these regions are held to constitute a “core” recollection network (see Introduction and Rugg and Vilberg 2013).

As in the present experiment, prior aging studies employing the R/K test procedure have also reported reliable recollection effects in older adults in regions belonging to the core recollection network (Duarte et al. 2008, 2010; Angel et al. 2013). As was noted in the Introduction section, with one exception (Duarte et al. 2008: high-performing older subgroup), the findings from these prior R/K studies largely took the form of a relative reduction in the magnitude of older subjects' recollection effects relative to the effects identified in young subjects. Although it is tempting to attribute these results to the age-related differences in recollection accuracy that were evident in these studies (cf. Rugg and Morcom 2005), this seems unlikely to be a sufficient explanation. Angel et al. (2013) employed a study manipulation that eliminated age-related differences in recollection performance, but they nonetheless reported attenuated recollection effects in their older group. In addition, the present findings of age-invariant recollection effects were obtained despite a significant age-related difference in recollection performance. It is presently unclear whether these divergent findings reflect subtle differences between the studies in the characteristics of the subject samples, or in such factors as the experimental procedures or materials.

The finding that recollection effects were age-invariant in the present study is relevant to the question of whether young and older individuals typically retrieve different amounts of information when endorsing test items as Recollected. As was noted in the Introduction section, a possible explanation for the reported disparities between the effects of age on subjective and objective indices of recollection is that older subjects recollect less information than young subjects do, but they employ a different internal scale when required to report their recollective experiences. There is evidence that the magnitude of recollection-related enhancement of BOLD activity in components of the recollection network, most notably, in the left angular gyrus (e.g. Vilberg and Rugg 2009; Yu et al. 2012a) and the hippocampus (e.g. Rugg et al. 2012; Yu et al. 2012b) covaries with the amount of recollected information. The present finding that recollection effects in these regions did not vary according to age therefore argues against the possibility that our young and older subjects recollected different amounts of information on test trials attracting an Remember endorsement.

Whereas the present recollection effects were age-invariant, the overall magnitude of the BOLD responses (relative to the implicit baseline of the GLM) elicited by test items in the mPFC and hippocampus did differ with age. As is evident in Figure 2B, the parameter estimates for the young subjects were, on average, more negative with respect to baseline than were those of the older group (in arbitrary units, mean estimates in the young were −1.61 and −0.69 for the mPFC and hippocampus, respectively, compared with −0.99 and −0.10 for the older subjects). These findings are reminiscent of prior reports that “task-negative effects” are attenuated in older relative to young individuals [e.g. Lustig et al. 2003; Andrews-Hanna et al. 2007; Persson et al. 2007; Miller et al. 2008; Grady et al. 2010; de Chastelaine et al. 2015; for review, see Grady (2012)]. The present findings suggest that age-related attenuation of task-negative activity in a given region does not necessarily mean that the region also demonstrates smaller differences between the responses elicited by items engaging distinct cognitive processes. These findings may have implications for the functional interpretation of between-group differences in item-related BOLD responses that are measured with respect to baseline rather than to responses elicited by other classes of experimental items (e.g. Huijbers et al. 2014).

Cortical Reinstatement

A major focus of the present study was assessment of the prediction that, in comparison with young individuals, older adults would demonstrate reduced specificity of recollection, as this is operationalized either by the magnitude of reinstatement-related BOLD activity or by the performance of an MVPA classifier. We were unable to find evidence to support this prediction: whether assessed with univariate analyses or MVPA, older and young subjects demonstrated statistically equivalent cortical reinstatement effects.

To our knowledge, 3 prior studies have been reported in which reinstatement effects were contrasted according to age. Dulas and Duarte (2012) had subjects study words and objects, making 1 of 2 semantic judgments to each item. Subsequently, memory was tested for the items and the associated study judgment. There was little evidence of material-selective recollection effects even in the young subjects, and no differences in such effects according to age. In a second study (St-Laurent et al. 2014), MVPA was employed to assess reinstatement effects associated with the encoding and subsequent recall of 11 short video clips. It was reported that reinstatement effects were weaker in the older subject sample, consistent with the proposal that recollection declines in specificity with increasing age. In this experiment, however, each of the video clips was encoded and recalled multiple times (for a total of 21 study–test pairings per clip). Thus, as was acknowledged by the authors, it is difficult to know whether the findings reflect age-related differences in reinstatement per se, or in the benefit accruing from multiple encoding opportunities. In the third study, McDonough et al. (2014) contrasted retrieval-related reinstatement of pictorial information in older and younger subjects. These researchers reported that picture recollection effects overlapped partially with regions that were more active for pictures than words in a separate localizer task, and identified 4 clusters where these overlapping effects were greater in magnitude in the young group. Whether these findings should be interpreted as evidence for age-related attenuation of retrieval-related reinstatement is unclear, as McDonough et al. (2014) did not identify reinstatement effects by using a study condition manipulation. Thus, it is not possible to distinguish between retrieval effects that were selective for the retrieval of pictorial information (i.e. picture reinstatement effects), and nonselective retrieval effects that happened to overlap regions that were also sensitive to picture encoding [compare Fig. 2A and Supplementary Fig. 1, and see Johnson et al. (2013) for examples of overlap between encoding effects selective for one study condition and generic recollection-related activity].

The present findings suggest that recollection in older individuals is not necessarily associated with a reduction in the specificity of retrieved information, at least as specificity was operationalized in the present experiment. That said, it would be premature to reject the proposal that there are age-related differences in the specificity of recollected information solely on the basis of these findings. The 2 classes of study episode employed in the present experiment were selected so as to engender markedly different patterns of encoding-related activity, thereby maximizing the likelihood that highly differentiated reinstatement effects would be evident, at least in our young subjects. It remains to be seen whether the findings extend to circumstances where different classes of study episode are less distinctive, and reinstatement effects are more subtle. Relatedly, as in most prior studies of cortical reinstatement, the approach adopted here assessed reinstatement effects shared across an entire class of study items, rather than effects that were selective for individual items. Thus, it also remains to be determined whether the present findings extend to designs permitting assessment of item-specific reinstatement effects (e.g. Rissman et al. 2010; Staresina et al. 2012; Ritchey et al. 2013).

In addition to examining recollection-related reinstatement, we used the same MVPA classifier to assess reinstatement effects elicited by test items endorsed with a K judgment. (We thank an anonymous reviewer for suggesting this analysis.) In agreement with the findings of Johnson et al. (2009), the classifier was able to assign these items to the correct study task at above-chance levels, albeit over a more restricted range of TRs than was the case for the items endorsed with R judgments, and with significantly lower accuracy. As in the case of R judgments, there was little or no evidence to suggest that reinstatement effects associated with K judgments differed according to age group (although there is perhaps a hint that the effects in the older subjects were delayed relative to those in the young, see Fig. 4B).

As was discussed by Johnson et al. (2009), the finding that test trials on which recollection ostensibly failed can nonetheless be associated with cortical reinstatement is consistent with the proposal that the memory signal supporting recollection is continuous (e.g. Donaldson 1996; Slotnick and Dodson 2005; Wixted and Mickes 2010). By this account, subjects only endorse an item as R when this signal exceeds a response criterion; when the signal is below criterion, a K judgment is given. Thus, neural correlates of recollection, such as cortical reinstatement, will be evident for K judgments, although to a lesser extent than when an R endorsement is given. It is important to note that, under this account, the present findings remain fully compatible with “dual-process” models of recognition memory, in which recognition is supported both by retrieval of qualitative information (recollection) and a separate, scalar familiarity signal (e.g. Yonelinas 2002; Wixted and Mickes 2010). The findings are incompatible, however, with the commonly held assumption that, unlike the signal supporting familiarity, the signal supporting recollection is thresholded (cf. Wixted et al. 2010 and Yonelinas et al. 2010).

Post-Retrieval Monitoring

Post-retrieval monitoring refers to processes engaged when the outcome of a retrieval attempt must be evaluated in relation to current task goals (Rugg 2004). As already noted (see Introduction), here we followed Henson et al. (1999; see also Henson et al. 2000, and Achim and Lepage 2005) in employing a contrast (K > R) between test items assumed to impose high versus low demands on monitoring processes. According to this assumption, whereas items eliciting strong recollection can be unambiguously judged R, familiar items require additional monitoring to ensure that the sense of familiarity is not accompanied by an above-criterion recollection signal (cf. Rhodes et al. 2008). The finding that RTs for R judgments were some 800 ms faster than those for K judgments is consistent with this assumption.

As in the original Henson et al. (1999) study, monitoring-related effects were identified in right DLPFC and the anterior cingulate. Both of these regions are held to play key roles in large-scale brain networks supporting cognitive control [for reviews, see Cocchi et al. (2013) and Powers and Petersen (2013)]. The right DLPFC in particular has been implicated in post-retrieval monitoring in numerous prior studies [see Fletcher and Henson (2001) for a review of early work, and Han et al. (2009) for an alternative perspective]. In light of the evidence that cognitive control processes supported by the PFC are especially vulnerable to aging (e.g. Braver and Barch 2002; Hedden and Gabrieli 2004; Craik and Bialystok 2006), age-related differences in the neural correlates of post-retrieval monitoring would be unsurprising. Such differences have been reported in some studies (e.g. McDonough and Gallo 2013; McDonough et al. 2013; Mitchell et al. 2013; Dulas and Duarte 2014), but not in others (e.g. Mark and Rugg 1998; Li et al. 2004; Giovanello et al. 2010; Dulas and Duarte 2013). The present findings converge with these latter results to suggest that monitoring processes supported by the PFC are not always compromised by age. We conjecture that whether monitoring is found to decline in efficacy with age will depend on such factors as the complexity of the retrieval task (e.g. source memory and “exclusion” tasks are arguably more complex than the R/K task employed here), and hence the overall demands placed on the control processes engaged in support of goal-directed behavior.

Relationship Between fMRI Effects and Recollection Performance

The only fMRI effect to demonstrate a robust relationship with recollection performance was related not to generic recollection or cortical reinstatement effects, but to post-retrieval monitoring. Retrieval monitoring operations are thought to be important for the setting of decision boundaries (in this case, between items to be endorsed as R or K), in evaluating mnemonic evidence with respect to such boundaries, and in initiating additional retrieval attempts when appropriate. Significantly, the relationship between the present neural correlates of monitoring and recollection performance was carried entirely by the R hit rate (partial r = 0.512, P < 0.001, and 0.312, P < 0.05, for the anterior cingulate and right DLPFC, respectively), the R false alarm rate demonstrating a negligible correlation (max r = 0.098). A plausible explanation for these findings is that engagement of the monitoring operations reflected by these fMRI effects was associated with a tendency to extend or iterate a retrieval attempt in the face of weak evidence as to whether a familiar test item elicited sufficient qualitative information about its study episode to warrant an Remember endorsement. Clearly, while there were marked individual differences in the ability to engage these control processes effectively, as evidenced by the strong correlations between the fMRI effects and recollection performance, the processes were no less available to older than to young participants.

Concluding Comments

The 3 neural effects linked to successful recollection that were identified here—enhanced activity within the core recollection network, material-selective cortical reinstatement, and engagement of post-retrieval monitoring—were remarkably comparable in the 2 age groups. Thus, in at least some circumstances, the neural systems supporting successful episodic retrieval can maintain their functional integrity into later life [see also Mark and Rugg (1998); Li et al. (2004); Duarte et al. (2008)]. Notably, the present findings suggest that, contrary to our pre-experimental hypothesis, the diagnostic specificity of the information retrieved on test trials associated with a phenomenal report of successful recollection does not differ with age. Thus, the findings are consistent with a long-standing proposal (Mark and Rugg 1998) that the effects of age on episodic memory performance owe more to age-related differences in encoding operations (Malliet and Rajah 2014) and retrieval cue processing (Morcom and Rugg 2004; Jacoby et al. 2005) than they do to differences in the nature or the processing of recollected information.

Supplementary Material

Supplementary material can be found at http://www.cercor.oxfordjournals.org/ online.

Funding

This work was supported by the National Institute on Aging (grant no. 5R01AG039103).