«By Zachary Alexander Rosner A dissertation submitted in partial satisfaction of the requirements for the degree of Doctor of Philosophy in Psychology ...»
42 Chapter 4: Mechanisms Underlying the Generation Effect Introduction As previously stated, psychological theories have suggested that the generation effect is driven by a host of internally mediated, top-down processes such as conceptual analysis (Jacoby, 1983), semantic integration (McElroy, 1987), item distinctiveness (Begg et al., 1989; Hunt & McDaniel, 1993; Kinoshita, 1989), and selective attention (Jurica & Shimamura, 1999; Tyler et al., 1979). Such processes may be defined more distinctly by addressing the neural processes that drive the generation effect. Yet despite extensive behavioral analyses (Bertsch et al., 2007), no published study, to our knowledge, has assessed the neural correlates of the generation effect.
Candidate structures that could potentially drive this active encoding effect include those involved in top–down executive processing. For example, semantic retrieval and conceptual analysis, which lead to elaborative, long-lasting memory traces (Craik & Lockhart, 1972), have been linked to activity in the left inferior frontal gyrus (IFG) (Baker, Sanders, Maccotta, & Buckner, 2001; Bookheimer, 2002; Poldrack et al., 1999). Other prefrontal regions, particularly in the dorsolateral prefrontal cortex (dlPFC), such as the middle frontal gyrus (MFG), have been associated with other executive control processes presumed to interact dynamically with posterior regions (Miller & Cohen, 2001; Shimamura, 2000, 2008). For example, dlPFC regions have been associated with a variety of working memory processes that lead to long–term memory formation (Paller and Wagner, 2002), such as refreshing perceptual features, maintaining items in memory, manipulating information, and selecting items for retrieval (Cohen et al., 1997; D’Esposito et al., 1997; D’Esposito, Postle, Ballard, & Lease, 1999; Johnson et al., 2005; Postle, 2006; Raye, Johnson, Mitchell, Reeder, & Greene, 2002; Thompson-Schill, D’Esposito, Aguirre, & Farah, 1997).
To the extent that the generation effect is mediated by item distinctiveness, it may be that posterior regions involved in verbal or item analysis, such as the left middle temporal gyrus (MTG) and lateral occipital cortex (LOC) (Binder, Desai, Graves, & Conant, 2009; Cabeza & Nyberg, 2000; Malach et al., 1995) also become particularly involved. Additionally, one might predict increased activation in the anterior cingulate cortex (ACC), which is involved in conflict monitoring (van Veen, Cohen, Botvinick, Stenger, & Carter, 2001) and verbal generation (Barch, Braver, Sabb, & Noll, 2000). Finally, with respect to monitoring internally or cognitively mediated processing, the generation effect may map onto activation related to the so-called default mode network (DMN), initially observed during periods of "rest," such as between stimulus presentations (Raichle et al., 2001). The DMN is a set of brain regions that includes the dorsal medial prefrontal cortex (dMPFC), ventral medial prefrontal cortex (vMPFC), posterior cingulate cortex (PCC), inferior parietal lobule (IPL), precuneus (PrC), retrosplenial cortex (Rsp), lateral temporal cortex (LTC), and hippocampal formation. Upon further analysis, this network has been associated with various internally mediated processes, such as episodic recollection, prospective memory, and perspective taking (Buckner, Andrews-Hanna, & Schacter, 2008; Buckner & Carroll, 2007; Spreng, Mar, & Kim, 2009). Given the view that the generation effect is involved in internally mediated processing, one might expect greater DMN activation during encoding for generate versus read items.
With respect to long-term memory processes, activity in the IFG during encoding has been particularly associated with successful retrieval (Brewer, Zhao, Desmond, Glover, & Gabrieli, 1998; Paller & Wagner, 2002; Wagner et al., 1998). Specifically, the IFG is more 43 active during encoding for items subsequently remembered compared to those subsequently forgotten. This effect is robust and has been observed in a variety of tasks and conditions (Paller & Wagner, 2002). In addition to the IFG, generation may increase activity in other areas also associated with this subsequent memory effect, including the frontal operculum (FOP), fusiform gyrus (FG), inferior temporal gyrus (ITG), cingulate gyrus, dorsal posterior parietal cortex (dPPC), and LOC (Cansino, Maquet, Dolan, & Rugg, 2002; Kirchhoff, Wagner, Maril, & Stern, 2000; Uncapher & Wagner, 2009; Wagner et al., 1998).
In the present study, we employed a prototypical memory paradigm used to assess the generation effect. Participants were shown related word pairs in the form of a cue word and word fragment (e.g., QUARREL–F_GHT) and asked to complete the second word in each pair. These encoding trials were compared to trials in which participants simply read related pairs (e.g., QUARREL–FIGHT) (Figure 3.1A). At test, old/new recognition memory for the second word in each pair was assessed with confidence ratings (high vs. low) (Figure 3.1B). Participants were scanned during both study and test phases to identify the neural substrates underlying the generation effect.
Materials and Methods Participants Twenty-four healthy individuals (13 females, 11 males, mean age = 23 years, range = 18– 32 years; all right-handed, native English speakers) participated in the study. Informed consent was obtained according to guidelines approved by the UC Berkeley Office for the Protection of Human Subjects. No participants reported any history of neuropsychiatric disorder or recent use of psychoactive medication. Participants were compensated $12 per hour.
Design and Materials A total of 200 cue-target synonym word pairs were constructed (e.g., GARBAGE– WASTE). One hundred items were presented at study and again at test, while the other 100 items were used as lures at test. Target words were obtained from the MRC Psycholinguistic Database (Wilson, 1988) and consisted of a mean word length of 5.39 letters (range = 3–8 letters), and a mean frequency of 54.32 (range = 1–314) (Kucera & Francis, 1982). During encoding, target words were presented in fragmented form (generate condition; e.g., GARBAGE–W_ST_) or in complete form (read condition; e.g., GARBAGE–WASTE). Fragments were created by removing each vowel (unless it began a word) and replacing it with an underline score. The encoding strategy (read vs. generate) and mnemonic status (old vs. new) of each word were counterbalanced across participants.
Behavioral Procedure The study phase was presented in 2 separate scanning blocks, each consisting of a randomized presentation of 25 generate and 25 read trials. For each study trial, the stimulus (either intact or fragmented pairs) was shown for 3 seconds which was followed by a 500millisecond blank screen and a jittered fixation cross (4–8 seconds). Participants were instructed to make a keypress response when they could identify the second word in each pair (i.e., the target word). This procedure encouraged comparable processing across study conditions, except that fragmented items had to be generated (Figure 3.1A).
Following the study set, a 3–minute filled retention interval was presented. During this interval, participants were shown 24 simple math equations (e.g., 3 + 5 = 8) and determined 44 whether the answer was true or false. Thereafter, old/new recognition memory was assessed using the 50 target items and 50 new items. New items were target words from unused word pairs. For each test trial, a word was presented for 500 milliseconds, followed by a 3-second blank screen, and a jittered response interval (4-8 seconds) (Figure 3.1B). Participants determined whether a test word was old or new while indicating their confidence (high or low) for each response during the inter-trial interval (ITI). They were instructed to respond old with high confidence (HC) only if they were absolutely certain that the test item was presented during the study phase. Thus, we interpret such HC hits to reflect strong recollective responses. Upon completion of the first study-test block, the behavioral procedure was repeated with a different set of cue-target pairs.
fMRI Acquisition A 3T Siemens (Erlangen, Germany) Trio scanner housed at the UC Berkeley Brain Imaging Center was used to acquire T1-weighted anatomical images and T2*-weighted echoplanar images (EPIs) [repetition time (TR) = 2000 milliseconds; echo time (TE) = 22 milliseconds; flip angle = 90º; matrix = 128x128; FOV = 220mm; 1.7x1.7 in-plane resolution] with GRAPPA [acceleration factor3]. For functional scans, EPIs consisted of 37 axial slices,
2.5mm thick, oriented to the anterior–posterior commissure (AC–PC), and were acquired in an interleaved order which resulted in whole brain coverage. A total of 155 volumes (run duration = 310 seconds) were collected during each of 2 encoding runs and 255 volumes (run duration = 510 seconds) were collected during each of 2 retrieval runs. The first 5 volumes of each run were used for magnetization preparation and were removed from future analyses, resulting in 150 and 250 volumes for each encoding and retrieval session, respectively. For registration purposes, a high resolution magnetization-prepared rapid-acquisition gradient echo (MPRAGE) volume [TR = 2300 milliseconds; TE = 2.98; matrix = 256x256; FOV = 256; sagittal plane; slice thickness=1 mm; 160 slices] and a gradient-echo multislice (GEMS) volume [TR = 250 milliseconds; TE = 3;
matrix = 256x256; FOV = 220; 3mm slice thickness, 28 slices] were collected. Due to movement artifacts, 8 of the 96 runs were excluded from data analysis.
fMRI Data Analysis All data processing and analyses were performed using the FMRIB Software Library (FSL) toolbox v4.1.4 (http://www.fmrib.ox.ac.uk/fsl; S. M. Smith et al., 2004). During preprocessing, BET (brain extraction tool) was applied to each participant’s data to separate brain tissue from skull and dura using a mask of the brain from the first volume, which was used for subsequent volumes. Images were then spatially smoothed using a 5mm full width at half maximum (FWHM) Gaussian kernel. To remove low frequency artifacts, highpass temporal filtering was performed with the local Gaussian-weighted fit of a running line. Motion Correction using FMRIB’s Linear Image Registration Tool (MCFLIRT) corrected for motion by aligning images to the middle slice with rigid body transformation. Sinc interpolation (Hanning windowed) shifted each slice in the volume in reference to the middle of the TR period. Next, FLIRT (FMRIB’s Linear Image Registration Tool) registered subject’s EPIs to their skullstripped high resolution T1-weighted images, which were then registered to standard Montreal Neurological Institute (MNI) space (FSL’s MNI152 template), both of which were combined to transform the EPI's and statistical maps into standard space.
At the first level of analysis, a multilevel, mixed effects general linear model was run using FILM (FMRIB’s Improved Linear Model). Each individual run (2 encoding and 2 retrieval 45 runs per participant) was modeled in individual subject space. Next, each resulting statistical map was registered to standard space. Regressors of interest were obtained by convolving stimulus onset times with FSL’s canonical (gamma) hemodynamic response function and temporal derivative. Trials in which the participant failed to respond, including those trials in which the participant was unable to identify the target word at encoding, were included in the model as regressors of no interest. Finally, motion parameters were added as a confound variable and temporal autocorrelation was removed through prewhitening.
At the second level of analysis, each subject’s 2 encoding runs were combined, as were each subject’s 2 retrieval runs, using one-sample t-tests. These runs were treated as fixed effects.
At the third level, statistical maps were created at the group level for each contrast using FLAME (FMRIB's Local Analysis of multilevel GLM Mixed Effects). The whole-brain family-wise error was corrected to p.05 using Gaussian Random Field theory with a cluster forming threshold of Z 2.3. To assess the relationship between behavioral performance and neural activity, we applied 2 separate subject-specific covariate analyses. First, we used individual generation effect recollection benefit (generate HC hit rate – read HC hit rate) as a covariate of interest in an analysis of generation effect recollection activity (generate read, HC hits). We used HC responses to isolate recollection responses and eliminate the confound of varying memory strength and remove possible guess trials. In addition, we used individual memory performance (hit – false alarm score) as a covariate of interest in an analysis of overall generation effect activity (generate read, all items). Localizations of peak activations were identified by mapping images onto the Harvard-Oxford Cortical Atlas.
Results and Discussion Behavioral Data We confirmed the robust benefit afforded by the generation effect. Specifically, the generate condition produced a hit rate that was 22% greater than that for read items (generate hits = 87%, read hits = 65%, t(23) = 9.97, p.001, false alarm rate = 21%; see Figure 3.1C, Table 3.1). The difference between the two conditions was even greater when performance was based only on high-confident hits (generate HC hits = 74%, read HC hits = 42%, t(23) = 11.61, p.001, HC false alarm rate = 7%; see Figure 3.1C, Table 3.1). As mentioned above, an HC rating was made only when participants were absolutely certain that they had seen a test item during the study phase. Given our findings for HC hits, we can assert that the generation effect is particularly potent in driving strong recollective responses. During encoding, the ability to identity targets was high and not significantly different between generated and read targets (generated targets = 98%, read targets = 99%, t(23) = 1.89, p =.07). Mean latency to identify a target was longer for generated items than read items (generate = 843 milliseconds, read = 670 milliseconds, t(23) = 6.58, p.001).
fMRI Data We first assessed memory-related effects by contrasting activations during encoding for items that were subsequently remembered with those that were subsequently forgotten, collapsed across encoding condition. This contrast revealed significant activation in the left LOC. In a second analysis, we assessed items that elicited HC (i.e., strongly recollected) ratings. This contrast revealed significant activation in the left LOC, IFG, ITG, and right precentral gyrus.
Thus, with respect to encoding effects, memory-related activity was particularly observed for items remembered with HC (i.e. strong recollections).