Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Brain networks involved in source monitoring in schizophrenia Metzak, Paul 2011

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2011_spring_metzak_paul.pdf [ 7.52MB ]
Metadata
JSON: 24-1.0071664.json
JSON-LD: 24-1.0071664-ld.json
RDF/XML (Pretty): 24-1.0071664-rdf.xml
RDF/JSON: 24-1.0071664-rdf.json
Turtle: 24-1.0071664-turtle.txt
N-Triples: 24-1.0071664-rdf-ntriples.txt
Original Record: 24-1.0071664-source.json
Full Text
24-1.0071664-fulltext.txt
Citation
24-1.0071664.ris

Full Text

BRAIN NETWORKS INVOLVED IN SOURCE MONITORING IN SCHIZOPHRENIA  by PAUL METZAK B.A., Simon Fraser University, 2006  A THESIS SUBMITED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE in THE FACULTY OF GRADUATE STUDIES (Neuroscience) THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) March, 2011 © Paul Metzak, 2011  Abstract  Schizophrenia is characterized by cognitive deficits in many domains. One of the domains in which these deficits are commonly found is in self-other source monitoring. Source monitoring refers to the set of processes by which individuals recall the conditions and contextual details surrounding the encoding of a memory episode, and self-other source monitoring specifically involves differentiating between actions performed by oneself versus those performed by another person. In this study, the goal was to investigate the neural basis of self-other source monitoring, and to discover how this neural activity differs in schizophrenia. The results of this study indicate that schizophrenia patients and healthy control subjects utilize essentially the same neural network for self-other source monitoring, and that this network involves brain areas that have been described as belonging to the task-positive and task-negative networks. Multiple statistical methods were used to analyze this dataset in order to provide a comprehensive set of results, as well as to determine the agreement between them. Although differences exist between the methods employed herein, in both the matrices that are used as input, and the mathematical operations performed on them, the results suggest that all the methods identified a common signal in the data.  ii  Preface This research was approved by the UBC Clinical Research Ethics Board, Certificate Number: C05-0509.  iii  Table of Contents Abstract .......................................................................................................................................... ii Preface ........................................................................................................................................... iii Table of Contents ......................................................................................................................... iv List of Tables ................................................................................................................................. x List of Figures ............................................................................................................................... xi Introduction ................................................................................................................................... 1 Reality Monitoring ...................................................................................................................... 3 The Forward Model ..................................................................................................................... 4 Neural Underpinnings of Source Monitoring.............................................................................. 5 Medial temporal lobes ............................................................................................................. 5 Prefrontal cortex ...................................................................................................................... 6 Parietal lobes............................................................................................................................ 8 Data Analysis Techniques ............................................................................................................ 9 Univariate Techniques............................................................................................................... 10 Statistical parametric mapping .............................................................................................. 10 Multivariate Techniques ............................................................................................................ 15 Principal component analysis ................................................................................................ 16 Constrained principal component analysis ............................................................................ 16 Partial least squares ............................................................................................................... 20  iv  Independent component analysis ........................................................................................... 24 Comparison Between fMRI Data Analysis Techniques .......................................................... 27 Matrices Used ............................................................................................................................ 28 Mathematical Operations .......................................................................................................... 30 Hypotheses ................................................................................................................................... 32 Methods ........................................................................................................................................ 33 Participants ................................................................................................................................ 33 Encoding.................................................................................................................................... 33 Recall ......................................................................................................................................... 34 Image Processing....................................................................................................................... 36 Neuroimaging Data Analysis Procedure ................................................................................... 36 Results .......................................................................................................................................... 39 Behavioural Results................................................................................................................... 39 Neuroimaging Results – Single Component Analyses ............................................................. 39 Single subject, single condition (any stimulus) HRF design matrix analysis ....................... 40 Single subject, single condition (any stimulus) FIR design matrix analysis ......................... 41 21 controls, single condition (any stimulus) HRF design matrix analysis ............................ 43 21 patients, single condition (any stimulus) HRF design matrix analysis............................. 44 21 controls, single condition (any stimulus) FIR design matrix analysis .............................. 44 21 patients, single condition (any stimulus) FIR design matrix analysis .............................. 45 v  21 controls, 8 conditions HRF design matrix analysis .......................................................... 46 21 patients, 8 conditions HRF design matrix analysis........................................................... 47 21 controls, 8 conditions FIR design matrix analysis ............................................................ 48 21 patients, 8 conditions FIR design matrix analysis ............................................................ 50 21 controls and 21 patients, 8 conditions HRF design matrix analysis ................................. 51 21 controls and 21 patients, 8 conditions FIR design matrix analysis .................................. 52 Neuroimaging Results – Multiple Component Analyses .......................................................... 53 21 controls, 8 conditions HRF design matrix analysis .......................................................... 54 21 controls, 8 conditions FIR design matrix analysis ............................................................ 55 21 patients, 8 conditions HRF design matrix analysis........................................................... 56 21 patients, 8 conditions FIR design matrix analysis ............................................................ 57 21 controls and 21 patients, 8 conditions HRF design matrix analysis ................................. 58 21 controls and 21 patients, 8 conditions FIR design matrix analysis .................................. 59 Discussion..................................................................................................................................... 61 Conclusion ................................................................................................................................... 73 References .................................................................................................................................. 111 Appendices ................................................................................................................................. 121 Principal Component Analysis (Mathematics and Description) ............................................. 121 Activated and Deactivated Brain Regions – Single Component Analyses ............................. 123 Single subject, single condition (any stimulus) HRF design matrix analysis ..................... 123 vi  Single subject, single condition (any stimulus) FIR design matrix analysis ....................... 124 21 controls, single condition (any stimulus) HRF design matrix analysis .......................... 125 21 patients, single condition (any stimulus) HRF design matrix analysis........................... 126 21 controls, single condition (any stimulus) FIR design matrix analysis ............................ 127 21 patients, single condition (any stimulus) FIR design matrix analysis ............................ 128 21 controls, 8 conditions HRF design matrix analysis ........................................................ 130 21 patients, 8 conditions HRF design matrix analysis......................................................... 131 21 controls, 8 conditions FIR design matrix analysis .......................................................... 132 21 patients, 8 conditions FIR design matrix analysis .......................................................... 133 21 controls and 21 patients, 8 conditions HRF design matrix analysis ............................... 135 21 controls and 21 patients, 8 conditions FIR design matrix analysis ................................ 136 Activated and Deactivated Brain Regions – Multiple Component Analyses ......................... 138 21 controls, 8 conditions HRF design matrix analysis ........................................................ 138 21 controls, 8 conditions FIR design matrix analysis .......................................................... 139 21 patients, 8 conditions HRF design matrix analysis......................................................... 141 21 patients, 8 conditions FIR design matrix analysis .......................................................... 142 21 controls and 21 patients, 8 conditions HRF design matrix analysis ............................... 144 21 controls and 21 patients, 8 conditions FIR design matrix analysis ................................ 145 Statistics for the Single Component Analyses ........................................................................ 147 21 controls, single condition (any stimulus) HRF design matrix analysis .......................... 147 vii  21 patients, single condition (any stimulus) HRF design matrix analysis........................... 147 21 controls, single condition (any stimulus) FIR design matrix analysis ............................ 148 21 patients, single condition (any stimulus) FIR design matrix analysis ............................ 148 21 controls, 8 conditions HRF design matrix analysis ........................................................ 148 21 patients, 8 conditions HRF design matrix analysis......................................................... 149 21 controls, 8 conditions FIR design matrix analysis .......................................................... 149 21 patients, 8 conditions FIR design matrix analysis .......................................................... 150 21 controls and 21 patients, 8 conditions HRF design matrix analysis ............................... 151 21 controls and 21 patients, 8 conditions FIR design matrix analysis ................................ 151 Statistics for the Multiple Component Analyses ..................................................................... 153 21 controls, 8 conditions HRF design matrix analysis ........................................................ 153 21 controls, 8 conditions FIR design matrix analysis .......................................................... 154 21 patients, 8 conditions HRF design matrix analysis......................................................... 156 21 patients, 8 conditions FIR design matrix analysis .......................................................... 157 21 controls and 21 patients, 8 conditions HRF design matrix analysis ............................... 160 21 controls and 21 patients, 8 conditions FIR design matrix analysis ................................ 162 Mathematics Underlying Behavioural PLS............................................................................. 166 The Correlations between Behavioural Performance and Brain Activity ............................... 167 Discussion of correlations between brain and behaviour .................................................... 169 Analyses Specific to SPM5 ..................................................................................................... 170 viii  21 controls, 8 conditions HRF design matrix contrasts ....................................................... 171 21 controls, 8 conditions FIR design matrix contrasts ........................................................ 171 21 patients, 8 conditions SPM5 HRF design matrix contrasts ............................................ 172 21 patients 8 conditions FIR design matrix contrasts .......................................................... 172 21 controls and 21 patients, 8 conditions FIR design matrix contrasts ............................... 173 Discussion of Analyses Specific to SPM5 .............................................................................. 173  ix  List of Tables Table 1. Important differences between the statistical analysis methods used in this project. ..... 75 Table 2. Group Demographics (standard deviations in parentheses). .......................................... 76 Table 3. Mean accuracy in percent correct for each task type by group (standard deviations in parentheses)................................................................................................................................... 77 Table 4. Mean response times in milliseconds for each task type by group (standard deviations in parentheses)................................................................................................................................... 78  x  List of Figures Figure 1. Schematic of how the PLS BOLD data vector is obtained from a single subject‟s conventional BOLD data matrix. .................................................................................................. 79 Figure 2. Examples of encoding tasks .......................................................................................... 80 Figure 3. Functional brain images from the single subject, single condition (any stimulus) analysis using an HRF model ....................................................................................................... 81 Figure 4. Functional brain images from the single subject, single condition (any stimulus) analysis using an FIR model ......................................................................................................... 82 Figure 5. Functional brain images from the 21 controls, single condition (any stimulus) analysis using an HRF model ..................................................................................................................... 83 Figure 6. Functional brain images from the 21 patients, single condition (any stimulus) analysis using an HRF model ..................................................................................................................... 84 Figure 7. Functional brain images from the 21 controls, single condition (any stimulus) analysis using an FIR model ....................................................................................................................... 85 Figure 8. Functional brain images from the 21 patients, single condition (any stimulus) analysis using an FIR model ....................................................................................................................... 86 Figure 9. Functional brain images from the 21 controls, 8 condition analysis using an HRF model ....................................................................................................................................................... 87 Figure 10. Functional brain images from the 21 patients, 8 condition analysis using an HRF model............................................................................................................................................. 88 Figure 11.Functional brain images from the 21 controls, 8 condition analysis using an FIR model ....................................................................................................................................................... 89  xi  Figure 12. Functional brain images from the 21 patients, 8 condition analysis using an FIR model ....................................................................................................................................................... 91 Figure 13. Functional brain images from the 21 controls and 21 patients, 8 condition analysis using an HRF model ..................................................................................................................... 93 Figure 14. Functional brain images from the 21 controls and 21 patients, 8 condition analysis using an FIR model ....................................................................................................................... 94 Figure 15. 21 control multicomponent solution with an HRF design matrix. Each method lists the number of components used to create the functional brain image................................................ 96 Figure 16. 21 control multicomponent solution with an FIR design matrix. Each method lists the number of components used to create the functional brain image................................................ 98 Figure 17. 21 patient multicomponent solutions with an HRF design matrix. Each method lists the number of components used to create the functional brain image. ....................................... 101 Figure 18. 21 patient multicomponent solutions with an FIR design matrix. Each method lists the number of components used to create the functional brain image.............................................. 103 Figure 19. 21 control & 21 patient multicomponent solutions with an HRF design matrix. Each method lists the number of components used to create the functional brain image. .................. 106 Figure 20. 21 control & 21 patient multicomponent solutions with an FIR design matrix. Each method lists the number of components used to create the functional brain image. .................. 108 Figure 21. Functional brain images and activity time courses from the 21 control 21 patient correlation with behavioural accuracy analysis. ......................................................................... 181 Figure 22. Functional brain images of significant contrasts from the SPM5 - 21 control, 8 condition analysis using an HRF design matrix. ........................................................................ 184  xii  Figure 23. Functional brain images of significant contrasts from the SPM5 - 21 control, 8 condition analysis using an FIR design matrix. .......................................................................... 185 Figure 24. Functional brain images of significant contrasts from the SPM5 – 21 patient, 8 condition analysis using an HRF design matrix. ........................................................................ 186 Figure 25. Functional brain images of significant contrasts from the SPM5 – 21 patient, 8 condition analysis using an FIR design matrix. .......................................................................... 187 Figure 26. Functional brain images of significant contrasts from the SPM5 – 21 controls and 21 patients, 8 condition FIR analysis. .............................................................................................. 188  xiii  Introduction Source monitoring refers to the ability to distinguish the wide range of variables that specify the context and conditions in which a memory episode was encoded (Johnson et al., 1993). The conditions can include the sensory modality through which the memory was encoded, the media or agent through which the information was presented, the social, spatial, temporal, or affective context of the memory, or any other features which can serve to distinguish the origin of a memory. Although the source monitoring framework is very general, its core thesis is that memories do not come „tagged‟ or „labeled‟ as resulting from a particular source, but rather that the source information is retrievable on the basis of the qualitative and quantitative characteristics of the memories themselves (Lindsay, 2008). Thus source monitoring involves both memory of a particular episode, and a judgment process whereby the characteristics of the memory are interpreted and evaluated. For instance, it is likely that remembering a story you heard on the radio would be associated with a great deal of auditory information but very little visual information or conceptual elaboration, which would help to correctly identify its modality and/or media source. Source monitoring has been of interest to many fields, and has been studied in many contexts including spatial (Ross and Slotnick, 2008) and temporal (Duarte et al., 2009) source monitoring, task monitoring, source monitoring throughout development (Poole and Lindsay, 2002) and aging (Henkel et al., 1998), source monitoring in legal contexts (Zaragoza and Lane, 1994; Hekkanen and McEvoy, 2005), source monitoring in animals (Squire and Zola-Morgan, 1991), and emotional source monitoring (Mather and Knight, 2008).  1  While most of the source judgments are made quickly and automatically, the decision making process underlying the distinguishment of source can also be slow and deliberative. These instances of strategic source monitoring usually involve both the recollection of perceptual details as well as the reasoning processes regarding the likelihood of the memory arising from different sources. In these latter cases the memories are evaluated for consistency with the sources from which they are thought have originated. These strategic source monitoring decisions are more common when: 1) there is little contextual information in the memory event, 2) there is much overlap (or interference) between the memory characteristics surrounding different events, and 3) the judgment processes underlying source attribution are poor (Johnson et al., 1993). The amount of contextual information available may be influenced by factors such as poor attention at encoding or using cues at retrieval. For instance, the performance of a fingertapping task or visual reaction time task during encoding was found to impair source memory to a greater extent than recognition memory (Troyer et al., 1999). In an experiment examining the effect of overlap between memory characteristics, Johnson et al. (1988) found that participants had more difficulty in distinguishing words they heard the experimenter say aloud versus words they imagined being spoken if they imagined the words being spoken in the experimenter‟s voice, as this heightens the degree of overlap between the two conditions. Additionally, Lindsay (1990) found that encoding a slide show and a narrative depicting the same event but with conflicting details during the same experimental session led to a greater number of errors than encoding the slide show and narrative during different experimental sessions. He argues that this is due to the high degree of temporal overlap between the two encoding events. The judgment processes underlying source monitoring can be poor due to the strategies employed by the participants, which can be influenced by the types of questions asked by the 2  researchers (Lindsay and Johnson, 1989; Dodson and Johnson, 1993). In these studies, the number of source monitoring errors varied as a factor of how the response options and instructions were presented to participants, respectively. Reality Monitoring The source monitoring framework was developed from earlier studies focusing on reality monitoring, which was originally proposed as the process of discriminating between external perceptual events, and internal thoughts and imaginings (Johnson and Raye, 1981; Johnson et al., 1988; Johnson et al., 1994). Although both perceptual events and thoughts can serve to generate „real‟ memories, in this context, reality refers to events that occur outside of oneself. In the same way as was mentioned for source monitoring above, in reality monitoring, the source of the memories should be able to be distinguished on the basis of their contents. For instance, one would expect that, ceteris paribus, memories of external events would have more temporal, spatial, and sensory detail than do imagined or thought events. Also, they are also likely to contain more specific information that imagined events (Johnson and Raye, 1981). In contrast, memories formed solely from thought or imagination are likely to contain more traces of effortful cognitive operations, as imagining and thinking require more effort than mere perception (Hasher and Zacks, 1979). That reality monitoring forms a distinct and distinguishable subclass of source monitoring in general is supported by findings that performance on reality monitoring tasks is selectively impaired in certain populations (Hashtroudi et al., 1989; McNally and Kohlbeck, 1993; Barnes et al., 2003). Notably, schizophrenia patients have been found to display problems with reality monitoring (Harvey, 1985; Brébion et al., 1996; Brebion et al., 2000). Furthermore, these reality monitoring deficits in schizophrenia have been linked to specific positive 3  symptoms: hallucinations (Aleman et al., 2003; Ditman and Kuperberg, 2005; Woodward et al., 2007), delusions (Brebion et al., 2000), and thought disorder (Harvey, 1985); although it should be noted that other researchers have failed to find any association between symptomatology and reality monitoring deficits (Allen et al., 2007; Versmissen et al., 2007). It should be noted that some researchers would argue that the experiment to be described in these pages would not fall under the rubric of reality monitoring as both of the source monitoring conditions involve responding to an external stimulus (e.g. perceiving an event), whereas a true reality monitoring experiment requires the generation of internal stimuli (e.g. imagining an event). The Forward Model It has been theorized that these failures in reality monitoring are illustrative of a core deficit in schizophrenia: a failure to recognize self-initiated actions (Frith et al., 1998; Frith, 2005). This theory is grounded in the observation that delusions of passivity in schizophrenia (i.e. delusions of thought insertion, thought broadcasting, and alien control) involve a loss of sense of agency. Studies have found that patients suffering from these delusions also show deficits in the monitoring of self-produced actions (Frith and Done, 1989). Similarly, auditory verbal hallucinations (AVHs) in schizophrenia have been associated with sub-vocal speech (Stephane et al., 2001) as well as dysfunctions in brain regions that support speech production (McGuire and Shah, 1993) which implies that patients exhibiting these symptoms may be suffering from an inability to monitor their actions and therefore attribute them to an external agent. In the non-clinical population, the monitoring of internally generated thoughts and actions is hypothesized to occur via a forward model, whereby the motor command to perform an action leads to the production of an efference copy of the command that contains the predicted outcome 4  of that action (Miall and Wolpert, 1996). This model explains how the brain can differentiate between self and other generated actions in many situations including tickling oneself (Blakemore et al., 2000a). Using this model, it has been shown that schizophrenia patients suffering from AVHs and delusions of passivity do exhibit differences in how they react to selfgenerated movements relative to schizophrenia patients who do not exhibit these symptoms and healthy control subjects (Blakemore et al., 2000b; Frith et al., 2000). The goal of this study was to examine the neural underpinnings of source monitoring in schizophrenia patients and healthy controls using a novel source monitoring paradigm. Neural Underpinnings of Source Monitoring Since the advent of functional MRI in the early nineties (Ogawa et al., 1992), hundreds of fMRI investigations of various types of source memory have been performed (Mitchell and Johnson, 2009). These experiments have revealed that there are multiple brain regions that underlie successful source monitoring performance. These regions are concentrated in the medial temporal lobes (MTL), prefrontal cortex (PFC), and parietal lobes. The contributions from each of these regions will be discussed in turn. Medial temporal lobes Evidence of the crucial involvement of the medial temporal lobes (MTL) in memory comes from animal studies (Squire and Zola-Morgan, 1991), as well as human patients that have undergone temporal lobe resection (Scoville and Milner, 1957). The primary anatomical structures of interest to memory researchers in the medial temporal lobes include the hippocampal formation, the entorhinal, perirhinal, and parahippocampal gyri (Squire et al.,  5  2004). The central role of the MTL structures appears to be the binding of memory items and context (Eichenbaum et al., 2007). The results from several fMRI experiments using source monitoring paradigms also support this functional mapping of the MTL. For instance, significant increases in hippocampal activity are often found at both encoding and test in source monitoring studies when correctly recalled source items and incorrectly recalled source items are contrasted (Diana et al., 2007). Prefrontal cortex Prefrontal cortex (PFC) activity has been widely reported in studies of source monitoring (Janowsky et al., 1989; Mitchell et al., 2004), and this activity has been associated with retrieval strategies (Dobbins et al., 2002; Dobbins et al., 2003). Furthermore, some have argued that the PFC activity lateralizes on the basis of the retrieval strategy employed. For instances, studies have found that retrieval of source information activates left PFC preferentially when compared to activations resulting from retrieval of new/old judgments (Johnson et al., 1997; Nolde et al., 1998). In contrast, significant right PFC activity is found when performing quick heuristic judgments of source that do not require systematic evaluation (Dobbins et al., 2003; Mitchell and Johnson, 2009). The PFC has been further subdivided into dorsolateral (DLPFC) and ventrolateral (VLPFC) on the basis of cytoarchitecture (Petrides and Pandya, 2002), as well as the functional roles that each of these areas is hypothesized to carry out. VLPFC has been found to be more involved in the selection process for allocating attention to the goal-relevant features of a memory stimulus, whereas the DLPFC appears to be more involved in encoding between-item associations and organizing information in memory (Blumenfeld and Ranganath, 2007). Support  6  for this functional subdivision in the PFC comes from studies that have demonstrated that VLPFC is significantly more active when attempting to resolve proactive interference (Jonides and Nee, 2006) and when making semantic judgments where response competition between the alternatives is varied (Thompson-Schill et al., 1997), whereas the DLPFC is more active when organizing (Postle et al., 1999) or manipulating items in memory (Bor et al., 2003). Mitchell and Johnson (2009) argue that this division implies that DLPFC should be more important when performing source memory tasks, whereas VLPFC memory should be more important when performing item memory tasks. In addition to DLPFC and VLPFC distinctions, another region of prefrontal cortex that has been found to be involved in source monitoring is the medial prefrontal cortex (also known as fronto-polar cortex, rostral prefrontal cortex, or anterior prefrontal cortex). The medial prefrontal cortex (mPFC), which largely overlaps with Brodmann‟s area 10, has been argued to be active when performing memory tasks that require contextual recollection. However, not all context memory studies find significant activations in this region; one theory posits that memory studies that use perceptual features as context information are less likely to elicit mPFC activity than memory studies in which the contextual recollection involved recalling which study task had been performed at encoding (Simons et al., 2005a). The authors of this study argue that the reason for this difference in mPFC activity is that recalling which task was performed involves using internally-generated, rather than externally available, contextual information to perform the memory task. This internally generated contextual information is made up of the thought processes and cognitive operations required to mentally disambiguate the task conditions during encoding. As memory processes often re-activate areas used during encoding (Wheeler et al., 2000), the mPFC activity found in multiple trial type experiments results from re-instantiating  7  the mental processes occurring at encoding in an attempt to use this information to identify the source of the memory. This region has also been found to be active when making inferences about the mental states of other people (Frith and Frith, 2003), distinguishing between imagined and perceived events (Vinogradov et al., 2008), or the evaluation of internally generated information (Christoff et al., 2003), all of which support the idea that this region is involved in self-referential thinking. Parietal lobes The parietal cortices are also often found to be active in many source memory experiments as they are hypothesized to be involved in both the initial bindings of disparate stimulus features (Uncapher et al., 2006), and voluntary attention (Husain and Nachev, 2007). The binding of different aspects of a stimulus into a unified episode is a requisite step in successful source monitoring, and fMRI studies have hypothesized that the intraparietal sulcus (IPS) may play a key role in this process. Significant activity has been found in the IPS when stimulus features are bound together in the visual (Donner et al., 2002) and auditory modalities (Cusack, 2005), as well as across sensory modalities (Calvert, 2001). Taken together, these findings imply that one of the roles of the IPS is to provide a structural organization for incoming sensory information, which is a critical feature for successful source monitoring.  8  Data Analysis Techniques Throughout the history of brain research from antiquity to present, there has been a divide between methods that have sought to understand the brain as a series of discrete, functionally specialized subunits, and those who have sought to understand it as single working unit comprised of the coordinated activity of integrated brain structures (Finger, 1994). One pertinent example of this conflict in the history of neuropsychology comes from the study of aphasia, where Broca and Wernicke sought to localize specific types of aphasia to lesions in discrete brain regions. However, their work was questioned by Marie, Head, Freud, and others who argued that the variability they saw in their patients, in terms of both behaviour and brain lesions, suggested that aphasia involved many more areas of the brain than were incorporated by the localizationist‟s models (Grabowski et al., 2000). Although today‟s neuroscientists would consider this contrast to be a false dichotomy, the conflict between these two viewpoints is often played out in the choice of statistical techniques used for data analysis, as the answers provided by the data analysis methods often rely on assumptions about what constitutes a valid answer. This can lead to additional difficulties in interpreting results as the answers provided by various methods often disagree, and there exists little consensus as to appropriate ways to compare results. Currently, the techniques used to analyze neuroimaging data can be sorted into two main types that map onto this dichotomy: univariate and multivariate statistics. Univariate data analysis methods involve analyzing the activity in each voxel independently of the analyses performed on every other voxel using a general linear model (Friston, 1994a). The results from each analysis (i.e., the level of statistical significance for each voxel) is then mapped onto a brain  9  template image and thresholded to correct for the increased false-positive rate resulting from performing thousands of discrete non-independent statistical tests (Friston et al., 1991). This mass-univariate approach is best-suited to provide information about the functional specificity of brain regions (Frackowiak et al., 2004a). Univariate techniques using the general linear model require a design matrix, or a model of the timing of presentation of stimuli of interest. There are multiple methods to perform multivariate analyses on neuroimaging data but essentially, they all involve investigating relationships between voxel activity over time. Some assess the relationship of each voxel in the brain to every other voxel in the brain, some assess the relationship between voxels in a pre-specified region of interest (ROI) to every other voxel in the brain, and others require prespecification of every node in the putative network. Multivariate statistical techniques can be used to examine either correlational or causal relationships between voxels in the brain, but the methods that examine causal relationships can only be used in ROIbased multivariate analyses. Furthermore, models of stimulus timing are required for the majority, but not all multivariate data analysis techniques. Univariate Techniques As all univariate techniques for the analysis of neuroimaging data are variants of the same method, only a single univariate technique will be discussed.  Statistical parametric mapping Statistical Parametric Mapping is both the method and the name of the most widely used statistical analysis software package for the analysis of neuroimaging data. Statistical Parametric Mapping (SPM) has come to refer to the joint use of a General Linear Model (GLM) and Gaussian Random Field (GRF) theory in analyzing neuroimaging data (Frackowiak et al., 2004b). The GLM is a statistical model in which an obtained response variable y can be 10  explained via a linear combination of explanatory variable(s) x, a constant c and an error term e. Each explanatory variable is multiplied by a parameter estimate β which serves to weight each explanatory variable to provide a better fit of the model to the obtained response variable. (1)  Where n is the number of timepoints (scans), m is the number of voxels, and p is the number of explanatory variables used to model the data. Note that the constant c does not appear in the formula, as the constant term can be removed from the model if the mean BOLD signal value from each voxel is subtracted from every scan in that voxel‟s time series. In fMRI studies, the response variable is the BOLD response measured in each voxel, each column in the design matrix is one explanatory variable, the constant is the baseline BOLD response in each voxel, and the error term accounts for the residual error between the fitted model and the response variable (Smith, 2001). In the case of an experiment with two conditions, there would be two explanatory variables in the GLM that correspond to the two columns in the design matrix coding for the timing of stimulus presentation. If a voxel responds strongly to the stimuli presented in condition 1, and the voxel‟s response is well matched to the modeled explanatory variable, then the β value for condition 1 for that voxel would be high. If the β value for this voxel is high in condition 1 for all (or most) subjects then it would seem likely that the stimuli presented in condition 1 are related to the BOLD activity in that voxel, although the relationship is not necessarily a causal one. Given the mass-univariate approach implemented in SPM, whereby each voxel in the brain is tested individually using an independent general linear model, the problem of correcting for multiple comparisons becomes an important issue. For a single comparison, we compare the obtained value from the GLM in a given voxel to the null distribution, and if the value is more 11  extreme than a predetermined cutoff value (usually 5% or 1%) corresponding to the probability of making a type I error, then the voxel is considered to be significantly activated (or deactivated). However, if this same value is used as a threshold for an entire family of comparisons, the probability of making a type I error multiplied by the number of independent statistical tests that are performed so the probability of obtaining a type I error becomes distressingly large. A simple method to control for the problem of multiple comparisons is to use a Bonferroni correction whereby the probability of a type I error for a single test statistic is divided by the number of tests to be performed to obtain a new threshold that maintains the probability of a type I error for the entire family of tests. However, the Bonferroni inequality is intended for use with a family of tests that are independent, whereas fMRI data contains correlations between spatially adjoining voxels, which results in the Bonferroni correction being too conservative. The correlations between adjacent voxels are due to many factors including the spatial extent of the hemodynamic responses, the poor mapping of voxels to discrete neural populations (which is compounded by the re-slicing process during pre-processing), physiological processes, and spatial smoothing performed during pre-processing (Frackowiak et al., 2004b). The spatial smoothing performed during pre-processing uses a Gaussian kernel to replace the raw voxel activity values with a weighted average of its activity values and those of its neighbors. The size of the kernel dictates both how many voxels will be included in each weighting, and how large the central value of the kernel will be. This smoothing kernel is used during pre-processing for three reasons; firstly, it helps compensate for any remaining variability in brain shape between participants following the non-linear warping conducted during the spatial normalization pre-processing step, secondly, it increases signal to noise ratio, and thirdly, 12  it allows the use of Gaussian Random Field Theory to determine the appropriate statistical significance threshold for the family of comparisons. The increase in signal to noise ratio results from the fact that a proportion of the noise in each voxel is uncorrelated from that of its neighbors but, given the reasons mentioned above, the signal is expected to extend over several voxels (Smith, 2001). Given that we know the size of the kernel used in the smoothing process, Gaussian Random Field Theory can be used in determining the expected Euler characteristic for each statistical map. The Euler characteristic uses the number of resolution elements (or resels) in the statistical map after smoothing to determine how many significant clusters of activation are to be expected at different Z-value thresholds (Worsley et al., 1992). Therefore, the Zthreshold corresponding to an expected Euler characteristic of 0.05 can be used as the statistical significance threshold that appropriately controls for the family-wise error resulting from making multiple comparisons (Frackowiak et al., 2004b). fMRI data is also subject to correlations between adjacent scans, which must be corrected for two reasons: firstly, the GLM implemented by SPM relies on the assumption that each time point represents an independent observation which implies that the error at each time point should be uncorrelated with the error at preceding and subsequent time points; and secondly, to maximize the difference between signal and noise during the analysis process (Turner et al., 1998). SPM deals with this problem via the use of a high-pass filter in conjunction with an autoregressor. The high-pass filter is used to eliminate low frequency sources of variance, as these low frequency sources of variability are often related to respiration, heartbeat, and scanner drift (Bianciardi et al., 2009). Not only do these low frequency signals add noise to the data but they can also appear to be signals of experimental interest (aliasing) in experiments with sub-optimal designs (Chuang and Chen, 2001). 13  The auto-regressor is intended to remove the correlations in the BOLD signal at adjacent time points in order to render the GLM statistically valid. This process, known as „whitening‟ of the data, is performed via an iterative process whereby a GLM is performed on the data, and the error term in the model is examined for correlations. Then that correlation is used to create a filter to „pre-whiten‟ the fMRI data before performing a new GLM analysis on the whitened data (Frackowiak et al., 2004c). Although this method of correcting for correlations between scans is implemented in SPM5, it has not yet been implemented in the Constrained Principal Component Analysis (i.e., the other statistical method that implements the GLM) software as its effects are currently being tested. Worsley has argued that this lack of independence between scans affects the standard errors of the estimated betas, rather than the betas themselves (Worsley, 2001) so it is unclear whether, or to what degree, pre-whitening of the data will influence the results of a Constrained Principal Components Analysis. Design Matrices Implemented in SPM5 In this report, two different types of design matrices were used: hemodynamic response function (HRF) models and finite impulse response (FIR) models. HRF models use stimulus onset timing and stimulus duration to create an idealized model of the predicted BOLD response to that stimulus via convolution with a sum of two gamma functions (Friston et al., 1998). Since the BOLD response measured by fMRI is a sluggish metabolic process that occurs following local increases in neural activity, and not a direct measure of neural activity itself, the HRF model provides an approximation of the shape and latency of the actual evoked response. FIR models estimate the changes in BOLD signal at specific peristimulus scans relative to all other scans. In an FIR model, a value (e.g. 1) is placed in the rows of the design matrix for which 14  BOLD signal amplitude is to be estimated, and another value (e.g. 0) is put in all other rows (resulting in “mini boxcar” functions). As opposed to the HRF model, the FIR model does not make assumptions about the shape of the actual evoked hemodynamic response, only about the time frame in which a hemodynamic response is expected. Multivariate Techniques Although all mass-univariate techniques are variants based the general linear model, there are many different methods for performing multivariate data analyses. These methods can be separated into exploratory and confirmatory techniques (Tabachnick and Fidell, 1989). Exploratory techniques are primarily used in situations where one wishes to examine the relationships between the variables in the dataset in an unconstrained manner. In contrast, confirmatory techniques are used to assess how well a particular theoretical model matches the obtained data. In fMRI research, the use of confirmatory multivariate techniques, like structural equation modeling (SEM) and dynamic causal modeling (DCM) require a theory about which brain regions are functionally relevant to the experimental task, and ideally, the structural connections of those regions (Penny et al., 2004). Although many brain regions have been implicated in the performance of self-other source monitoring, few of the neuroimaging experiments investigating this cognitive operation have used another form of source monitoring as a control condition, which makes it unclear whether these brain regions are uniquely involved in self-other source monitoring, or source monitoring more generally. Due to this uncertainty, only exploratory techniques have been used in the current project.  15  Principal component analysis  Principal component analysis (PCA) is used by two of the multivariate methods employed in this study (CPCA and PLS). A full description of PCA and its mathematical basis is presented in the Appendix of this report. PCA on an fMRI data set is not an ideal method to isolate functional networks of experimental interest, due to the fact that the components that are extracted by PCA are extracted in order of the amount of variance they explain, such that the first component explains the most variance, the second component explains the second most variance, and so forth (Daffertshofer et al., 2004). In the case of fMRI data, only 5% to 10% of the variance is attributable to task-induced changes in BOLD response (Biswal et al., 2007); therefore, it is likely that most of the components, particularly those extracted first and accounting for the majority of the variance, would be unrelated to experimental manipulations. Due to this, different methods have been developed to restrict the variance in the data set so that the PCA only extracts components that are sensitive to the experimental manipulations. In the following sections, two of these methods will be discussed. Constrained principal component analysis  Constrained Principal Component Analysis (CPCA) is a general method for structural analysis of multivariate data that combines regression analysis and principal component analysis into a unified framework (Takane and Shibayama, 1991). CPCA proceeds in 2 steps: first, the total variability in the criterion data (i.e. the dependent variables) is partitioned into variability related to the predictor data (i.e. the independent variables) and variability that is unrelated to the  16  predictor data via multivariate regression. In step two, PCA is performed on each of the two resulting matrices in order to detect possible underlying structures related (or unrelated) to the predictor variables. As two methods of structure extraction, regression and PCA are complementary as they allow the criterion data to be separated into both known and unknown structures. The regression analysis decomposes the criterion data on the basis of its relationship to known structure in the predictor data, whereas the PCA decomposes the criterion data into unknown structure on the basis of patterns of intercorrelations found within the criterion data itself. The full form of the CPCA model can be found below: (2) where Z is the criterion data, G is the matrix of constraints on the rows of Z, H is the matrix of constraints on the columns of Z, and E is the variance in Z that cannot be estimated by G, H or the interaction of G and H. In the full model, the matrices of to-be-estimated parameters are M, B, and C. The first term in the model (GMH) assesses the variance that can be explained by the interaction between G and H, the second term in the model (BH) assesses the variance that can be explained by H, but not G, and the third term in the model (GC) assesses the variance that can be explained by G, but not H. For the purposes of fMRI experiments, Z can be composed of each subject‟s normalized and (optionally) smoothed images whereby each column contains data from a single voxel and each row contains data from a single scan, with the rows proceeding temporally so that the first subject‟s first scan comprises row one and the last subject‟s last scan comprises the last row. So each row in the Z matrix is m voxels in size and each column is n subjects by s scans. The 17  columns of Z should be mean-centered and normalized to unit length prior to analysis. The G matrix can be any type of design matrix as long as it is full rank and has the same number of rows (i.e., n subjects by s scans) as the Z matrix to enable the matrix algebra to proceed. To date, the variations that have been used for the G matrix include various types of design matrices: canonical HRFs generated using SPM5 and FIR basis sets (Woodward et al., 2006; Metzak et al., 2010). The canonical HRF and FIR-based G matrices both produce images of brain systems that account for a large portion of the variability in the data, however, the HRF designs produce predictor weights that display how involved the network is in the each of the experimental conditions modeled in the design matrix, whereas the FIR G matrices produce predictor weights that display how involved the network is in the each of the experimental conditions modeled in the design matrix at each time point in the FIR time window. In either case, the rows of the G matrix have a length of k conditions by t timepoints, where an HRF G matrix has one timepoint per condition. The H matrix can be used to constrain results to hypothetical brain networks of interest; however it was not used in this project and is mentioned only for completeness. For instance, H matrices may be used to examine brain networks that are lateralized to one hemisphere or the other, brain networks localized dorsal or ventral portions of the cortex, or any other brain network involving areas that can be theoretically pre-specified. Row constraint matrices similar to the G matrix can also be used to explicitly remove sources of nuisance variance from the data. For instance, estimates of autocorrelation between scans (as was mentioned in the preceding section on SPM) or linear trends in the data that reflect signals of non-neuronal origin (Smith et al., 1999) can be in row constraints in order to remove these effects from the measured BOLD response.  18  Although it is possible to use the full model listed in (1) above in the analysis of fMRI data, to date only the CPCA model with row constraints has been used: (3) where n is the number of subjects, s is the number of scans, and m is the number of voxels. In this model, the total variance in the data is partitioned into two matrices; the matrix GC, which contains the variability that can be explained by the timing of stimulus presentation, and the matrix E which contains the variability that cannot be explained by the timing of stimulus presentation. The matrix G contains the timing information for the fMRI experiment, where each row of G represents a different scan. C is the matrix of condition specific regression weights and is obtained by regressing Z onto G via the following formula: (4) where n is the number of subjects, s is the number of scans, m is the number of voxels, k is the number of conditions, and t is the number of timepoints modeled in each condition. The condition-specific regression weights are often referred to (in conventional univariate fMRI analyses) as beta images. The matrix GC is then subjected to PCA in order to identify patterns of intercorrelated voxel activity that are predictable from the presentation of experimental stimuli via the following analysis: (5) where n is the number of subjects, s is the number of scans, m is the number of voxels, p is the number of components extracted, and the square brackets denote the products of singular value decomposition.  19  This decomposition yields: a) right singular vectors (V) which can be overlaid on a structural brain image to indicate patterns of functionally connected voxel activity related to the presentation of the experimental stimuli, b) the diagonal matrix of singular values (D), and c) the left singular vectors (U) which can be used to produce predictor weights (uppercase P): (6) where n is the number of subjects, s is the number of scans, (lowercase) p is the number of components extracted k is the number of conditions, and t is the number of timepoints modeled in each condition. As can be seen from the formula 6 above, the predictor weights (P) are the weights that would be applied to G to produce U, and indicate the importance of each column in the G matrix to the networks represented by each component. Partial least squares  Partial Least Squares (PLS) refers to both a general statistical technique, as well as the program that implements that technique for neuroimaging data. In its original form, PLS (the neuroimaging software) operates on the covariation between the obtained data and either the contrasts of experimental interest or a behavioural measure (McIntosh et al., 1996). However, a newer version of PLS has also been introduced that uses a mean centering approach which does not require contrasts to be specified and yields identical results (McIntosh and Lobaugh, 2004).  As originally formulated, PLS was intended for use with positron emission tomography (PET) data and operated on the covariance between the PET data and the contrasts of experimental interest. Since PET studies require a blocked design as well as the use of the subtraction method (e.g. where task specific activity is assessed via the subtraction of control 20  condition activity from experimental condition activity), the contrasts of experimental interest can be thought of as being akin to a simple design matrix. When PLS was adapted for use with event related fMRI (ER-fMRI) data, it continued with the same analytical method; that is, it decomposes the covariance matrix obtained from the data matrix, which will be further described in the following paragraph, and the contrasts of experimental interest (McIntosh and Lobaugh, 2004). However, in the case of an ER-fMRI design, it is much more difficult to see the contrasts of experimental interest as being akin to a simplified design matrix since they do not contain information about expected response shapes (e.g. HRF convolutions), or the timing of the different events. Instead the structure of the data matrix itself contains the timing information about the experiment, as will be explained below. All of the PLS variations require that the obtained BOLD data be rearranged into a configuration that is unique to this method. A very common way to arrange BOLD time series data (e.g. this method is implemented in SPM fixed effects analyses, as well as CPCA and ICA) is for each row of the data matrix to contain measurements for a single scan (from a single subject) and for each column in the matrix to contain data from a single voxel, with the rows proceeding temporally from first to last scan in the run. In contrast, PLS arranges the obtained BOLD data into matrix M, such that each row of the data matrix contains data from one observation in one condition, where observations are individual subjects in a study with multiple participants, and trials in a study involving a single individual. In all cases the observations are normalized within a trial by expressing each voxel and time point combination as percent signal change from the onset of the trial. Furthermore, in multisubject analyses, each observation is the mean of all trials in that condition for that particular subject. So the rows in the M matrix are comprised of n observations × k conditions. The columns of the matrix contain measures from 21  each voxel at each pre-specified time point. The number of time points from each observation included in the columns of the M matrix is called the lag-window (McIntosh and Lobaugh, 2004). A lag window is chosen such that a voxel‟s mean hemodynamic response in each condition may be visualized, as PLS does not impose a canonical HRF shape on the response. The columns in the M matrix are arranged so that the activity measured in voxel 1 at time point 1 goes in the first column; the activity measured in voxel 1 at time point 2 goes into the second column, and so forth until all the voxel values at each time point are placed in the columns. The number of columns in the M matrix is m voxels × t time points. Please see Figure 1 for a schematic of the process involved in generating a PLS M matrix from the more commonly encountered BOLD data matrix used by SPM and others. Another point to note is that measures of voxel activity that occur outside the lag windows of the observations are not included in the computations performed during the PLS analysis. Depending on how often the observations occur and the duration of the lag windows, this means that some of the data from the scanning session may be omitted from any calculations. In this way, the PLS analysis removes some of the sources of variance extraneous to the cognitive operations under investigation. PLS can be performed in three separate ways: (1) contrast PLS, (2) mean centering PLS and (3) behavioural PLS. Each of these three ways of performing a PLS analysis will be discussed, but the behavioural PLS will be presented only in the Appendix. Contrast based PLS is the original form of the analysis procedure, and involves preparing the M matrix mentioned above as well as a matrix of orthonormal contrasts C. k-1 contrasts may be set up and these contrasts are repeated for every observation so that matrix C is n × k rows and k-1 columns. Then a k-1 by m×t covariance matrix Y is generated:  22  (7) Where k is the number of conditions, m is the number of voxels, t is the number of time points, and n is the number of observations. The Y matrix does not contain any subject specific information as the multiplication of the data and contrast matrices creates condition specific averages. The transpose of the Y matrix is then subjected to PCA: (8)  where k is the number of conditions, m is the number of voxels, t is the number of time points, and the square brackets denote the products of singular value decomposition. The left singular vectors, U, are labeled the voxel saliences, the right singular vectors, V, are labeled the contrast saliences, and S is the diagonal matrix of singular values. The voxel saliences indicate the level of involvement of each voxel in the component and the brain activity at each time point in the lag window can be depicted, whereas the contrast saliences indicate how related each experimental task is to the brain activity detected in the component. The mean-centering approach to PLS involves preparing a matrix (T) which is composed of the column wise means, averaged over subjects, from each condition coded in the M matrix. Therefore the matrix contains k rows and m × t columns. The T matrix is mean-centered via the following formula: (9) where k is the number of conditions, m is the number of voxels, t is the number of time points, and  is a column of ones of length k. Then PCA is performed on the matrix Tdev:  23  (10) k is the number of conditions, m is the number of voxels, t is the number of time points, and the square brackets denote the products of singular value decomposition. The left singular vectors, U, are labeled the voxel saliences, the right singular vectors, V, are labeled the contrast saliences, and S is the diagonal matrix of singular values. As above, the voxel saliences indicate the level of involvement of each voxel in the component and the brain activity at each time point in the lag window can be depicted, whereas the contrast saliences indicate how related each experimental task is to the brain activity detected in the component. Aside from these differences in scaling, the results from the mean-centered and contrast based approaches produce identical analytic results. In the case of the mean-centering PLS, the contrast saliences, V, are also known as the design scores. These design scores are plotted in the Figures along with the functional brain images resulting from the PLS analyses. Note that for all the PLS analyses, the functional brain image selected was from the third time point in the lag window. Independent component analysis Independent component analysis (ICA) is a blind source separation technique that is used to separate signal mixtures into their constituent source signals (Stone, 2004). This problem can be illustrated by the following formula: (11) where s is the number of scans (or time information), m is the number of voxels, and p is the number of components,  is the matrix of obtained signal mixtures, C is the matrix of  independent components, and M is the mixing matrix that governs how the independent components are combined. 24  The signal mixtures can be any physically measured be speech signals, sonar or radar signals, stock prices, neuroimaging time series data, or any physically measured set of signals. ICA performs this analysis on the basis of 2 underlying assumptions. The first is that if signals originate from different physical processes, they should be statistically independent (Hyvarinen and Oja, 2000). The statistical independence requirement for the extracted source signals (or components) is quite stringent, as it goes beyond a lack of correlation or orthogonality between the components. Strictly speaking, the independence requirement states that information about the value of any particular component does not provide information about the value of any other components. Uncorrelated vectors are only partially independent as they may still contain higher order dependencies. For instance, non-linear functions of the variables in question may be dependent but not display correlations. The second factor that ICA uses in identifying source signals is their normality. According to the central limit theorem, a signal mixture that is the linear sum of source signals will tend to display a more normal or Gaussian histogram than the source signals themselves. Furthermore, as any mixture tends to have a Gaussian distribution, the source signals are assumed to have sparse non- Gaussian distributions. ICA decomposition uses an iterative self-organizing learning algorithm to generate an unmixing matrix that extracts components that demonstrate these two properties: statistical independence and a non-Gaussian distribution. This process can be characterized via the following formula: (12)  25  where s is the number of scans (or time information), m is the number of voxels, and p is the number of components, C is the matrix of independent components, X is the obtained mixing matrix, and W is the unmixing matrix generated via the ICA algorithm. If the data can be perfectly reconstructed via the unmixing matrix W, then W-1 = M. Thus, the algorithm that generates the unmixing matrix is at the heart of the ICA decomposition. Although many blind source separation algorithms have been developed (Bell and Sejnowski, 1995; Hyvarinen and Oja, 2000), empirical research has shown that, in general, the ICA algorithms converge on similar solutions (Giannakopoulos et al., 1999; Esposito et al., 2002). The first ICA algorithm that was adapted for use in fMRI analysis was the Infomax algorithm (Bell and Sejnowski, 1995). One technical problem that had to be overcome in adapting ICA for fMRI was that there must be at least as many mixtures as there are source signals. In fMRI, this is not the case, as the voxels can be thought of as the source signals whereas the time points (scans) can be thought of as the mixtures. This technical limitation was overcome by McKeown et al. (1998), by transposing the obtained BOLD data matrix so that the independent components were computed based on the correlations between the time points and assessed over the voxels.  26  Comparison Between fMRI Data Analysis Techniques The methods used to analyze the source monitoring data in this project are diverse, yet there exist several grounds for comparison. The most basic distinction that can be drawn is between the mass-univariate general linear model based analysis performed by SPM versus the multivariate analyses performed by ICA, PLS, and CPCA. Clearly, the mass-univariate approach is easily distinguished from the multivariate methods on the basis of the questions it is ideally suited to solve, as well as the type of statistical barriers it must overcome. Specifically, the massunivariate approach is ideally suited to identifying localized brain regions involved in performing discrete cognitive functions because it is optimal for use with the subtraction method, whereby the BOLD activity from a control condition is subtracted from the BOLD activity from a tightly matched experimental condition in order to identify only those brain areas with differing activation between the two conditions. Additionally, the fact that in a typical mass-univariate analysis, upwards of 50,000 individual GLMs are calculated requires a more sophisticated significance testing procedure than those implemented in the multivariate methods discussed herein. By contrast, the multivariate techniques are ideally suited to provide information about changes in activity foci or changes in BOLD response over time in functionally connected neural networks. Some of the multivariate techniques provide functional brain images for each time point in a trial (the voxel saliences in PLS) whereas others provide information about changes in BOLD response in each of the experimental conditions over time (CPCA with an FIR design matrix, and ICA). The multivariate methods may be compared on the basis of the format and contents of the matrices they employ, and on the basis of the mathematical operations performed on the matrices.  27  Matrices Used In terms of the matrices used in the analysis, the one matrix that is necessary for every method is the matrix of obtained BOLD data, which is also referred to as the data matrix. The BOLD data is usually retrieved from the MRI scanner as a series of individual image files, where each file contains BOLD activity data from a single scan for a single subject. The data in each file is represented as a row vector containing BOLD activity measurements for each voxel in the whole brain image. The vector data from these individual files is often concatenated to form a matrix with each voxel‟s BOLD activity information placed in a column and the scans arranged in a descending order from first to last placed in the rows. This format for the BOLD activation data is used by CPCA, ICA, and SPM; although they differ in that the former uses data from all subjects whereas the latter two use this format on a subject specific basis. As mentioned above, PLS acquires the BOLD activation value from each trial for each voxel at each point in the lagwindow for each participant, and arranges them into a row vector such that the value of the first voxel at the first time point in the lag-window is in the first column, the value of the first voxel at the second time point in the lag-window is in the second column, and so forth until all the voxel values at all time points in the lag-window have been included. Furthermore, the BOLD values for each subsequent time point are expressed as a percentage difference from the signal present at the first time point. This process is repeated for each condition and each subject so that the data matrix employed by PLS will generally be a subset of the original data matrix with only the data from time windows of interest included in the analysis. Thus, PLS incorporates information from the experimental design into the data matrix itself. The other matrix that is most frequently used in fMRI data analysis is the design matrix, which codes the timing information for the conditions of interest in the experiment. Unlike the 28  BOLD activation matrix, this matrix is not required for all the techniques discussed. For example, ICA does not require the use of a design matrix as it is a blind source separation technique, although a design matrix generated by SPM can be used in a post-hoc fashion to identify the ICA components with activity time courses that best match the synthetic hemodynamic responses of conditions of interest. The SPM fMRI package is able to implement a number of types of design matrices, including a synthetic HRF (Friston, 1994b), the partial derivatives of the synthetic HRF, and finite impulse response (FIR) models. The synthetic (or canonical) HRF aims to provide a reasonable approximation of the BOLD response to neural activity, and is modeled as the combination of two gamma functions: one modeling the peak, and the other modeling the undershoot (Frackowiak et al., 2004d). Since the BOLD response is known to differ between participants, and even between brain regions within participants (Handwerker et al., 2004), the partial derivatives of the canonical HRF can be included in the model to help account for variations in the latency and duration of the peak response. The FIR basis function approach is one of the simplest and least constrained approaches to deconvolution which provides an estimate of average BOLD response increases over baseline for a number of peristimulus time points (Henson et al., 2001; Serences, 2004). All of these types of design matrices may also be implemented in CPCA, as this method is indiscriminate with regards to the model that serves as the predictor variables. PLS can be thought of as using something akin to an FIR basis function as part of its methodology; the lagwindow option specified by the user reflects the number of peristimulus time points included in the analysis. However, the „FIR design matrix‟ used by PLS is embedded into the structure of the BOLD activation matrix by virtue of how it is setup. As was mentioned above, PLS uses the trial timing and lag-window information to generate a BOLD activation matrix that only contains data  29  from the time points and conditions of interest. PLS does not use separate design and activation matrices as it combines the features of both into a single matrix. Mathematical Operations The other way in which these statistical methods may vary is in terms of the mathematical operations that are performed on the matrices. In the mass-univariate approach implemented in SPM, a regression analysis is performed on each voxel in the brain, whereby a beta-value is generated that indicates the fit between the modelled activity in the design matrix and the obtained BOLD activity. Two of the multivariate methods used in this experiment, CPCA and PLS, are identical to each other in the sense that they both use PCA to reduce the number of variables from 50,000+ (e.g., the number of voxels) to a small number of components (usually < 10) composed of weighted combinations of the original variables that serve to explain the largest proportion of the variability in the dataset. Also, they both restrict the use of PCA to a subset of the variability in the data (for reasons mentioned in the PCA section of this report); however, they do not yield identical results because they use different methods to define this subset. PLS performs PCA on the covariance matrix obtained from cross multiplying the data from the peristimulus time points of the specified conditions of interest and the experimental contrasts. Importantly, what PLS does not do is decompose the covariance between the data matrix and a timing (design) matrix. Data from time points that are not captured by the lagwindows are ignored. CPCA, on the other hand, regresses the acquired BOLD data onto the design matrix model and then performs PCA only on the portions of variance that is predictable from the design matrix. This process can also be understood as performing PCA on the covariance between the data matrix and the timing (design matrix). The ICA analysis proceeds in a much different fashion, as it relies on an iterative self-organizing learning algorithm that 30  separates signal mixtures (e.g. fMRI signals) into their statistically independent source components. Since this process relies on a learning algorithm, identifies components without using any sort of design matrix, and requires the components to meet a more stringent statistical criteria than PCA (independence versus orthogonality), ICA will always produce results that differ from the methods mentioned above. Please see Table 1 for a summary of some of the salient differences between the techniques used in this report.  31  Hypotheses In this study, we hypothesized that: (1) relative to healthy controls, schizophrenia patients would make significantly more errors in the self-recalled condition, (2) both patients and controls would show increased mPFC activity in the source monitoring conditions relative to the task monitoring conditions as assessed by the differences in predictor weights (CPCA), beta weights (ICA), and design scores (PLS), and (3) patients and controls would utilize essentially the same neural network to perform the tasks but that the patients would show increased hemodynamic activity relative to the controls due to reduced information processing efficiency (Metzak et al., 2011). All of the statistical methods employed in the analysis of the source monitoring data set can be used to test these hypotheses.  32  Methods Participants Participants in this experiment were 33 schizophrenia patients, and 32 healthy control subjects. The two groups were matched on age, handedness, gender, parental socio-economic status, and pre-morbid IQ (all ps > 0.05). However, 2 controls and 6 patients needed to be removed from the functional data analysis for excessive movement during the scanning session. Additionally, 9 controls and 6 patients were removed from the functional imaging analysis due to poor performance and/or errors in task administration. The final sample included for analysis was 21 schizophrenia patients (mean age = 29.57, S.D. = 9.35; 14 males, 7 females), and 21 control subjects (mean age = 27.81, S.D. = 5.83; 9 males, 12 females). In this sample, the groups remained matched on all demographic variables mentioned above, (all ps > 0.05). Please see Table 2 for the demographic information from each sample. Encoding Prior to functional scanning, participants were shown 30 words in each of four randomly presented conditions (120 words total): Self (or Internal), Other (or External), Association, and Reading. In the Self and Other conditions, a jumbled word puzzle was presented in conjunction with a clue about the meaning of the word, for example, “BERAZ” would appear on the screen with the clue “a striped grazing animal”. In the Self/Internal condition, participants were required to say the target word aloud once they had solved the puzzle. In the Other/External condition, a pre-recorded voice said the target word aloud as soon as the jumbled word and clue appeared on the screen. In both the Self and Other conditions, the participant advanced to the next trial by pressing a button on the keyboard. In the Association condition, the (correctly spelled) target word was presented in the center of the screen along with two other words in the lower left and 33  lower right corners. The participants were to indicate, via key press, which of the two words they felt was a closer semantic associate to the target word. For each target word, a strongly associated word and a weakly associated word were presented, with the relations selected based on the Edinburgh Associative Thesaurus (Kiss, Armstrong, Millroy, & Piper, 1973). When the participants made their selection, the target word and the selected associated remained on the screen for three seconds before the next trial began. In the Read condition, the target word was presented in the center of the screen along with the instructions “Please read silently”. Participants pressed a key on the keyboard to advance to the next trial. For the encoding run, the trials were self-paced, and the inter-trial interval was one, two, or three seconds (randomly chosen). For this study, the experimental condition of interest were the Self and Other conditions, referred to in this paper as Source Monitoring (SM), and the control conditions were the Read and Associate conditions, referred to in this paper as Task Monitoring (TM). In a previous study of source monitoring from our lab, it was found that words encoded in the internal condition were easier to recall than words encoded in the external condition (Woodward et al., 2007); therefore, we deliberately selected the association and reading as the TM condition since, through pilot testing, it was found that it was more difficult to recall words encoded in the read condition than the association condition). All words used in the encoding run were concrete nouns, and 2 versions of the encoding run were designed with source and task monitoring words reversed between conditions. See Figure 2 for examples of the encoding tasks. Recall After the completion of the encoding run, participants were taken to the MRI suite where they underwent a final MRI compatibility screening with the MRI technologist prior to 34  functional scanning. The recall run began approximately ten minutes after the encoding run. All 120 words presented during the encoding run were also presented during the recall run. The recall run lasted 15.5 minutes, and consisted of six alternating source monitoring and task monitoring blocks. Each SM block began with the following set of instructions printed on the screen: “Who solved it? You or Computer?” Then twenty words from the internal and external trials were presented sequentially in the center of the screen, and participants were asked to indicate, via keypress, whether “me” or “computer” had solved the puzzle. The words “me” and “computer” appeared on the bottom left and bottom right hand corners of the screen to remind participants of the response mapping. The side on which of “me” and “computer” appeared was alternated between participants. Each TM block began with the following set of instructions being printed on the screen: “What did you do? Read silently or associate?” Then twenty words from the associate and read trials were presented sequentially in the center of the screen, and participants were asked to indicate, via keypress, whether they had “associated” or “read” the target word. The words “associated” and “read” appeared on the bottom left and bottom right hand corners of the screen to remind participants of the response mapping. The side on which of “associated” and “read” appeared was alternated between participants. Each block took 120 seconds to complete, and each target word was presented for a maximum of 5 seconds. The target word disappeared from the screen when a response was made, and the screen remained blank until the allotted 5 seconds for the trial elapsed. Each trial was separated with a 1, 2, or 3 second ITI, and a 10-second blank trial was inserted between each block to avoid multicolinearity (Cairo, Liddle, Woodward, & Ngan, 2004). The word “Relax” was presented for the first 9 seconds of each blank trial followed by a crosshair for 1 second to warn the participant that the next block was beginning.  35  Image Processing Imaging was performed at the University of British Columbia's MRI Research Centre on a Phillips Achieva 3.0 Tesla MRI scanner with Quasar Dual Gradients (maximum gradient amplitude 80mT/m and a maximum slew rate of 200mT/m/s). The participant's head was firmly secured using a custom head holder. Functional images volumes were collected using a T2*weighted gradient echo spin pulse sequence (TR/TE=2000/30ms, flip angle 90º, 36 slices, 3mm thick, 1mm gap, sense factor 2, 80x80 matrix reconstructed at 128, FOV 240.0mm, measured voxel is 3mm x 3mm x 3mm, actual band width = 53.4 Hz per pixel) effectively covering the whole brain (145mm axial extent). Functional images were reconstructed offline, and the scan series was realigned and normalized using the method implemented in Statistical Parametric Mapping 5 (SPM5; http:/www.fil.ion.ucl.ac.uk/spm). Translation and rotation corrections did not exceed 3mm or 3º for any of the participants. Parameters for spatial normalization into the modified Talairach space used in SPM5 were determined using mean functional images constructed from the realigned images of each participant and scan series. Voxels were normalized to 3mm x 3mm x 3mm. The normalized functional images were smoothed with an 8mm full width at half maximum Gaussian filter. Neuroimaging Data Analysis Procedure For the neuroimaging data presented below, the components presented were selected on the basis of their relevance to the experimental conditions. For the CPCA and PLS components, this meant that we selected the first components extracted using the PCA analysis, as these components explained the most variance in the dataset. The ICA components were selected on the basis of the R-square value obtained from regressing each component‟s activity time course  36  onto the SPM5 design matrix. In all figures the PLS analyses depict the functional brain image from the 3rd time point in the lag window. For the multiple component section below, the number of components used for each analysis was determined by selecting components until there was a large drop in the amount of variance explained or in the R-square value obtained from the analysis. However, in order to achieve maximum correspondence between the methods, non-significant components were included in some cases. These cases are mentioned in the sections in which they arise. In order to assess the significance of the components and the differences between the conditions within each of the analyses presented below, numerous different ANOVAs were run on the predictor weights from the CPCA components that accounted for the most variance, and the beta weights from the ICA components with the highest R-square values. The setup for each of these ANOVAs will be described here in order to avoid repeating the same information throughout the results section. Note that the PLS analysis uses its own statistical procedure to assess the significance of the components, so the following procedures apply only to the results of the ICA and CPCA analyses. For the single subject, single condition analyses (both HRF and FIR), there were no statistical tests run. For the multiple subject, single condition analyses with an HRF design matrix, a one-sample t-test was employed to determine whether the mean beta or predictor weights differed significantly from zero. For the multiple subject, single condition analyses with an FIR design matrix, repeated measures ANOVAs were conducted on the predictor and beta weights, using a single factor of Time (e.g. the 7 peristimulus time points after the beta/predictor weight for the first time point was zeroed). For the multiple subject, multiple condition, single group HRF analyses, a 2 x 2 x 2 within-subjects ANOVAs with the factors of Difficulty (Easy (Internal & Associate) or Hard  37  (External & Read)), Condition Type (Source (Internal & External) or Task (Associate & Read)), and Performance (Non-recalled or Recalled) were run on the predictor and beta weights. For the multiple subject, multiple condition, both groups HRF analyses, a 2 x 2 x 2 x 2 mixed-model ANOVAs were run on the beta and predictor weights, with the within-subject factors of Difficulty (Easy (Internal & Associate) or Hard (External & Read)), Condition Type (Source (Internal & External) or Task (Associate & Read)), and Performance (Non-recalled or Recalled) and the between-subject factor of Group (Patient or Control). For the multiple subject, multiple condition, single group FIR analyses, a 2 x 2 x 2 x 7 x 2 within-subjects ANOVAs with the factors of Difficulty (Easy (Internal & Associate) or Hard (External & Read)), Condition Type (Source (Internal & External) or Task (Associate & Read)), and Performance (Non-recalled or Recalled) were run on the predictor and beta weights. For the multiple subject, multiple condition, both groups FIR analyses, a 2 x 2 x 2 x 7 x 2 mixed-model ANOVAs were run on the beta and predictor weights, with the within-subject factors of Difficulty (Easy (Internal & Associate) or Hard (External & Read)), Condition Type (Source (Internal & External) or Task (Associate & Read)), and Performance (Non-recalled or Recalled) and the between -subject factor of Group (Patient or Control). The results of all significance tests can be found in the Appendices.  38  Results Behavioural Results In terms of accuracy and response times, there were no significant differences between the patient and control groups in any of the four conditions (internal, external, read, & associate), or when the four conditions were combined to form the source monitoring and task monitoring conditions (all ps < 0.05). Please see Table 3 and Table 4 for summaries of accuracy and response times by condition and group. Neuroimaging Results – Single Component Analyses The neuroimaging results section has been arranged on the basis of the sample and design matrix used in the analysis. Each sample and analysis subsection contains the functional brain images for each of the techniques, as well as a discussion of the relevant findings. A list of the anatomical regions found to be activated or deactivated in these analyses can be found in the Appendix. Note that PLS was included in the FIR analyses sections below as the PLS program cannot accommodate HRF design matrices. Although PLS does not implement an FIR matrix per se, it appeared to be a better fit in the FIR sections due to the fact that it operates on the average activity from each peristimulus time point of interest. For the CPCA analyses, the design matrices used were either the HRF models taken directly from the SPM5 software package, or FIR models of the type described in the CPCA section above, using an identical peristimulus time window as the SPM5 generated version. For the ICA analyses, the SPM5 HRF or FIR design matrices were used as criterion variables for the regression of the activity time courses of the independent components derived from the ICA analysis in order to determine which components appeared to be most closely related to the experimental design and presentation of stimuli during the experiment. For the PLS analysis, the mean centering approach was used so no 39  explicit contrast matrix was used; therefore it was the column-wise mean centered data matrix that was decomposed, and not a covariance matrix. Also, as mentioned in the PLS section above, PLS does not use a design matrix per se, rather they reshape the data matrix to minimize the variance in the data matrix that is unrelated to the experimental tasks. The lag window for the PLS analysis was 8 time points, in order to match most closely the timing of the FIR design matrices used for the SPM5, ICA, and CPCA analyses. Single subject, single condition (any stimulus) HRF design matrix analysis As a first step that can be considered the most basic comparison, a very simple design matrix was used in which the presentation of any stimulus was coded as a single condition. As can be seen in Figure 3, the results of two SPM analyses are reported here. The two analyses were identical in all respects, except that the second analysis did not use the autoregressor and high pass filter functions described in the SPM5 section above. For this analysis, the autoregressor function was AR(1) and the high pass filter was set to 128 seconds. The primary difference between these two analyses is that the non-auto-regressed/filtered results show less widespread activations, however, the brain regions in which activity is found is generally conserved between these two SPM5 analyses. These analyses demonstrate substantial overlap between the 3 methods used in terms of the detected patterns of functional activity. The SPM5 and CPCA analyses were found to be identical if the AR(1) auto-regressor and high pass filters were not used. This correspondence between the results is due to the fact that, even though the precise way in which the regression analysis is performed in SPM5 and CPCA differs (univariate vs. multivariate, respectively), the beta weights obtained should correspond if the same model is used as a predictor variable. The GC matrix that is decomposed in the subsequent step of the CPCA analysis comprises only one 40  pattern of variability (e.g. it is a rank of one), therefore only one component could be extracted, and this component reflects the same pattern of variability found in the beta weights. Additionally, the component from the ICA analysis that regressed most highly on the SPM5 design matrix had an activity pattern that coincided with some of the regions found to be active in the SPM5 and CPCA analyses. Single subject, single condition (any stimulus) FIR design matrix analysis This analysis employed the exact same subject and timing information as the previous HRF analysis, however, in this case, an FIR model with a peristimulus time window of 8 scans was used instead of an HRF model. This allowed us to assess whether a less constrained model would reveal identical patterns of brain activity as well as how the results from each of the statistical methods may diverge when another type of design matrix is employed. As can be seen in Figure 4, the results of two SPM analyses are reported here. As above, the two analyses were identical in all respects, except that the second analysis did not use the autoregressor and high pass filter functions described in the SPM5 section above. Once again, the autoregressor function was AR(1) and the high pass filter was set to 128 seconds. Similar to the previous analysis using the same timing and sample in conjunction with an HRF model, the primary difference between these two SPM FIR analyses is that the non auto-regressed/high-pass filtered results show less widespread activations, however, in this analysis, the overall difference in activity levels between the two is much greater. The activations and deactivations found in the multivariate analyses are also sparser and more widely distributed than their respective HRF analyses conducted on the same sample. The CPCA analysis shared a few foci of activation with the auto-regressed and high pass filtered  41  SPM analysis, notably in the right inferior frontal gyrus, as well as in the inferior retrosplenial cortex. Unlike the previous analysis, the CPCA results were not identical to the SPM5 results due to the fact that the GC decomposed in the CPCA analysis had a rank of 8 (due to the 8 columns in the FIR design matrix); therefore, the variance was able to be explained by multiple components, only one of which is presented here. The CPCA analysis also shared common activation foci with the ICA analysis, including left parahippocampal gyrus (which was also present in the non-auto-regressed and unfiltered SPM results), but the ICA analysis also found activity in the right parahippocampal gyrus as well. Interestingly, the auto-regressed and high pass filtered SPM analysis, the CPCA analysis, and the ICA analysis all found areas of deactivation in the medial prefrontal cortex, although the exact location of these deactivations did not overlap between these analyses. The PLS analysis did not retrieve an activity pattern that concorded with those of the other methods, although it is similar to the other multivariate methods in terms of sparsity of activations and deactivations. It should also be noted that a single subject single condition experimental design is poorly suited for the significance testing method employed by PLS, as in PLS analyses, the assessment of significance takes place at the voxel level using bootstrapping, and at the component level using permutation tests. The permutation analysis reassigns the order of conditions for each subject, then the PLS analysis is re-run and the resulting singular values are compared against the originally obtained singular values, and expressed as a probability, whereas the bootstrap analysis re-calculates the standard error of the voxels found to be maximally involved in each LV by rerunning the PLS analysis on randomly selected subsets of the entire sample (McIntosh et al., 2004). Given that the statistical testing procedure involves switching the order of conditions and selecting a subset of the subjects (in a single subject analysis, individual trials are coded as individual subjects), the use of a single  42  subject single condition design matrix is, at the very least, a sub-optimal choice for PLS analysis, as in this case there were no conditions to be switched and the choice of a subset of trials entails a loss in power. 21 controls, single condition (any stimulus) HRF design matrix analysis This analysis used an HRF design matrix with a single column design matrix for each subject that coded for any stimulus presentation. For this and all subsequent analyses, the SPM5 results presented will be those produced using the AR(1) autoregressor and 128 second high pass filter, as this is the default setting in SPM5. For each of the multivariate methods employed, only one component is depicted. The functional brain images from these analyses can be found in Figure 5. These HRF analyses revealed broadly overlapping activations using all three statistical methods. Furthermore, the SPM5 and CPCA analyses found similar regions of activation and deactivation. The ICA component whose activity time course regressed most strongly onto the SPM5 HRF design matrices of the 21 control subjects revealed activations in inferior parietal regions, and left pre- and post-central gyri, which were also found in the SPM5 and CPCA analyses. However, the dorsal anterior cingulate cortex (dAcc)/supplementary motor area (SMA) activations found in this ICA component appear to be located more caudally than the dAcc/SMA activations found in the SPM5 and CPCA analyses. The fact that only part of the activations found in the SPM and CPCA analyses are present in the ICA image suggest that perhaps several neural systems are working in coordination to perform this task. It is worth mentioning that only a single ICA component was overlaid onto a structural brain image in this analysis; given the nature of ICA component extraction, it is unlikely that any single component would capture the full sum of the activations and deactivations found using a single component/significant voxels 43  using the other methods, due to the fact that the ICA criteria for extracting components is unrelated to the general linear model. 21 patients, single condition (any stimulus) HRF design matrix analysis This analysis used an HRF design matrix with a single column design matrix for each subject that coded for any stimulus presentation. The functional brain images from these analyses can be found in Figure 6. The three analyses performed using this sample and HRF model revealed common patterns of activation, although there were differences in the amount of activity detected by each method. Specifically, all three methods found activations in the dAcc/SMA, and both CPCA and ICA shared additional activation foci with the SPM5 analysis, although not with each other. The activations in the CPCA analysis were more concentrated in posterior regions whereas the activations from the ICA analysis were concentrated the frontal lobe. The ICA results also contained deactivations not found in any of the other analyses presented in this section. These deactivations were found in the right temporal lobe and right inferior frontal gyrus. This suggests that, during this experiment, those areas acted in concert with the SMA/dAcc but in the opposite direction (e.g. they decreased activation as the SMA/dAcc increased activation). 21 controls, single condition (any stimulus) FIR design matrix analysis This analysis, and all the single condition FIR analyses that follow, used an FIR basis set consisting of impulse or „mini-boxcar‟ functions placed in each of the 8 time points following a stimulus presentation. All stimuli were coded in the same set of columns in the design matrix regardless of stimulus type or memory retrieval performance. The functional brain images resulting from these analyses can be found in Figure 7.  44  Although the results of these FIR analyses of the data from the 21 controls show more variability than those previously examined, there is still general concordance between them. For instance dAcc/SMA was found with every statistical technique except ICA; however the ICA component with activations in this region also regressed very strongly on the SPM5 FIR design matrix used as a criterion variable in this analysis. The pattern of visual cortex activation is also highly conserved in all four of the analyses presented in this section, however, here we observed the SPM5 FIR analysis identifying a visual cortex deactivation, whereas increases in visual cortex activations were detected using the three multivariate methods. Other notable overlaps include the cerebellar activations detected in both the CPCA and ICA analyses, and the bilateral temporal lobe deactivation found in the SPM5 and PLS analyses. In general, this analysis suggests that using an FIR matrix, as opposed to HRF matrix, leads to greater variability between the results obtained even when a very simple model of the stimuli is used, possibly due to the lack of an assumed shape for the response or the increased number of predictors. 21 patients, single condition (any stimulus) FIR design matrix analysis This analysis included the 21 schizophrenia patients using an FIR design matrix that coded for a single condition, the presentation of any stimulus. The functional brain images resulting from these analyses can be found in Figure 8. This series of analyses showed reasonable concordance between the different statistical methods. For instance, dAcc activity is present in both the PLS and CPCA analyses, and can also be faintly seen in the SPM5 results. The ICA component that regressed mostly highly onto the SPM5 FIR design matrix was characterized by visual cortex activity that co-occurs with stimulus presentations. The ICA, CPCA, and PLS brain images show a pattern of visual cortex activity that is nearly identical. Another notable result for these analyses is that the widespread 45  deactivations found in the SPM5 results were not found using any of the multivariate methods. It should be noted that this was also the case for the HRF analysis of the same sample (see Figure 5), where the SPM5 analysis was the only one to detect widespread cortical deactivations, especially in the medial regions of the frontal lobes. However, this result is largely due to a single component being extracted using the multivariate methods. 21 controls, 8 conditions HRF design matrix analysis These analyses used design matrices where each stimulus was coded based on the manner in which it was encoded, and whether that stimulus‟ encoding condition was successfully identified during recollection. The functional brain images from these analyses can be found in Figure 9. These results of the CPCA and ICA analyses show correspondence with the SPM5 analysis, but not with each other, as the results from the CPCA analysis resembles the pattern of deactivations found in the SPM5 analysis whereas the results from ICA analysis closely resemble the pattern of activations found in the SPM5 analysis. The regions that correspond between the SPM5 and CPCA analyses include the mPFC, the superior and middle temporal gyri, and the precuneus/posterior cingulate, which are all core regions of the default mode network (Gusnard and Raichle, 2001). The predictor weights for the CPCA analysis indicate that the network was least deactivated during the Internal (Self) Recalled condition and was most deactivated in the two Read conditions. The statistical analysis of the predictor weights found a significant effect of difficulty which was probably driven by the difference between the Internal and Read conditions. This finding agrees with previous literature regarding the role of medial prefrontal cortex activity in self-referential thought (Simons et al., 2005a). Additionally, the decrease in accuracy and 46  increase in response time for the Read conditions relative to the other experimental conditions, although non-significant, may indicate increased task difficulty. If this hypothesis is correct, then the increased deactivation may be related to the relative increase in difficulty in this condition (McKiernan et al., 2003). The regions that overlap between the SPM5 and ICA analyses include dAcc/SMA, insula, and inferior frontal gyri. These regions comprise part of the task positive network, which is anti-correlated to the task-negative network (Fox et al., 2005). The significant increase in the beta weights in the Non-recalled relative to the Recalled conditions is consistent with previous work implicating increased dAcc activity in situations where responses are made under conditions of uncertainty (Woodward et al., 2008). Although it interesting to see that ICA and CPCA are sensitive to different aspects of the same signal (i.e. the one present in the SPM5 analysis), it will become clear later in this in this work that this is an artifact resulting from examining only the single strongest component from each of these analyses. 21 patients, 8 conditions HRF design matrix analysis This set of analyses used design matrices where each stimulus was coded based on the manner in which it was encoded, and whether that stimulus‟ encoding condition was successfully identified during recollection. The functional brain images from these analyses can be found in Figure 10. This set of analyses revealed the strongest divergence between the statistical methods seen thus far. Although all three analyses show activity in the dAcc/SMA, the overall patterns of activity are more differentiated than in the previous analyses reported herein. The SPM5 analysis appears to show similar patterns of activity as the previous analyses, with activations in dAcc/SMA, insula, left inferior frontal gyrus and motor cortex, and deactivations in the medial 47  prefrontal cortices. However, the CPCA analyses revealed a network comprised of posteriorly located brain regions, and the ICA analysis found more activity in frontal areas including extensive bilateral DLPFC. Interestingly, the CPCA predictor weights and ICA R-square values suggest that the neural networks identified in these analyses display differential activity based on experimental context. For instance, although it is only a trend, the CPCA component appears to be important for distinguishing between correctly identified stimuli in the easy and hard conditions, a distinction which appears to be driven by the difference between the self and other encoding conditions. The ICA component appears to have a pattern of activity that is fairly consistent over Difficulty or Condition Type, however, there was a main effect of Performance, such that the Non-recalled conditions showed higher activity than the Recalled conditions. 21 controls, 8 conditions FIR design matrix analysis These analyses used design matrices where each stimulus was coded based on the manner in which it was encoded, and whether that stimulus‟ encoding condition was successfully identified during recollection. The functional brain images from these analyses can be found in Figure 11. These analyses demonstrate a large amount of divergence between the statistical methods, although the SPM5 and PLS analyses both identify a similar network of deactivated brain areas. Specifically, the SPM5 and PLS analyses both reveal deactivations in the mPFC, precuneus/posterior cingulate, and middle and superior temporal gyri; all regions purported to be part of the default mode network (Raichle et al., 2001). The visual cortex activity found in the ICA analysis does not correspond well with the visual cortex activity found using any of the other statistical methods, although it does appear to be well conserved from the previous 21 control subjects ICA analysis using the single condition FIR model (see Figure 7). The CPCA 48  component contains dAcc/SMA activity that concords with the SPM5 analysis, although the activity in this region is much more extensive in the CPCA analysis. Also, the occipital lobe activity in the CPCA analysis does not agree with that found in any of the other methods using the same design matrix. Specifically, the occipital lobe activity found in the CPCA analysis is located more medially than the occipital lobe activity found using ICA. Looking at the indicators of neural involvement in each condition and component from the multivariate methods (predictor weights for CPCA, beta weights for ICA, and design scores for PLS), it appears as though there is little agreement, although this is likely due to the differences between the functional networks detected. For the CPCA analysis, the predictor weights suggest that the internal recalled and associate recalled conditions elicited the most activity in this network. The internal recalled condition appears to reach its positive peak activity level more quickly than the other conditions, whereas the read recalled condition is slowest to reach its positive peak activity level. Also worth noting, is that the peak activity of the external recall condition is considerably lower than the peak activity in any of the other recall conditions, suggesting that this neural network plays a different role in that condition relative to the others. The beta weights from the ICA analysis show that all the conditions reach a positive peak at approximately 4 seconds peristimulus time; furthermore, the peaks for all conditions, with the exception of the External Non-recalled condition, are virtually indistinguishable. The beta weights also contain a negative peak (or undershoot) in which there is a distinction in undershoot magnitude on the basis of whether the stimulus was successfully recalled or not. The PLS design scores suggest that this network is least deactivated in the external nonrecalled condition and most deactivated in the association and internal recalled conditions. This is notable since this component produces opposite responses to the external non-recalled and  49  internal recalled conditions, and these conditions are the two in which participants responded that they themselves had solved the jumbled word. This component does discriminate between correct and incorrect “self” responses, yet the participants are unaware that the same behavioural response (e.g. “self”) was produced in response to differential levels of brain activity in this component depending on the correctness of their response. 21 patients, 8 conditions FIR design matrix analysis The functional brain images from the 21 patients, 8 conditions FIR design matrix analyses can be found in Figure 12. Once again, this set of analyses demonstrates a divergence in the first component obtained from each of the statistical methods even when identical design matrices are used. Although the SPM5 analysis reveals a pattern of brain activity that is consistent with those obtained using SPM5 in the analyses described in the preceding sections, there are some notable differences. Specifically, the areas of activation found in this analysis are much less distributed than in previous analyses with fewer voxels and clusters achieving the threshold for statistical significance. Additionally, one of these areas of activation appears to be located in the lateral ventricles just inferior to the corpus callosum. The CPCA analysis identified a component consisting of deactivated regions located in the occipital parietal and temporal lobes. This network shows some correspondence with the SPM5 analysis but is missing the mPFC deactivations found in the SPM5 analysis. The predictor weights from this CPCA analysis indicate that this network was least deactivated in the Internal (Self) conditions, with the Internal Recalled conditions showing the highest activity peak. This finding agrees with the 21 control FIR analysis in the previous section, as the CPCA component in that analysis was also least deactivated in the Internal Recalled condition. The ICA analysis identified a component primarily involving occipital cortex 50  as having an activity time course that most closely matched that of the SPM5 FIR design matrix. This result is consistent with the results obtained previously when regressing ICA component activity time courses onto SPM5 FIR design matrices, although it does not agree well with the results obtained from the other statistical methods using the same design matrix and sample. The mean betas for each condition and time point also concord well with the results from the ICA 21 controls SPM FIR design matrix analysis presented above, in that the conditions reach a positive peak around 4 seconds peristimulus time and a negative peak (undershoot) around 10 seconds perstimulus time. The differences between the Recalled and Non-recalled conditions in the latter half of the trial are also very similar to the ones seen in the ICA 21 controls SPM FIR design matrix analysis. The PLS analysis identified a network of diverse regions of activation, including thalamus and brain stem, which overlaps very little with the patterns identified using the other techniques, although the mPFC, and middle and superior temporal gyri were also found to be active in this analysis. However, since the sign (e.g. positive or negative) of the component is arbitrary following PCA, it is likely that the regions depicted as activated in this brain image are, in fact, deactivated. However, there is no way to reverse the sign of the component and design scores in the PLS GUI. Assuming that they should be reversed, the design scores indicate that the deactivations are strongest in the Read Non-recalled condition and that the activations are most pronounced in the Internal Non-recalled condition. 21 controls and 21 patients, 8 conditions HRF design matrix analysis The functional brain images from the 21 controls and 21 patients, 8 conditions HRF design matrix analyses can be found in Figure 13. The results of the three analyses presented here are not identical, but they concord generally with one another. The CPCA and SPM analyses showed deactivations in many overlapping regions, and the occipital cortex activity in the ICA 51  component with the highest R-square values overlapped with the occipital activations found in the SPM results. The predictor weights from the CPCA analysis indicate that the patients and controls show similar patterns of activity in this network, however, it should be noted that this component is deactivated more strongly in controls when unsuccessfully remembering a stimulus as being previously presented in the „Read‟ condition. In fact, the controls show more deactivation during the „Read‟ condition encoded words, regardless of whether the correct encoding context is recalled or not. The beta weights from the ICA component show that the activity in this component was sensitive to Performance and Difficulty for both the patients and controls. In the patients, Performance and Difficulty each independently affected the pattern of the beta weights, with the Non-recalled stimuli eliciting larger beta weights than the Recalled stimuli and the Difficult conditions (External and Read) eliciting larger beta weights than the Easy conditions. In the controls, the effect of Performance was only seen in the Easy conditions, with the Non-recalled stimuli resulting in larger beta weights than the Recalled stimuli; in the Difficult conditions this relationship was not present perhaps due to larger betas in the Read Recalled condition than in the Read Non-recalled condition. 21 controls and 21 patients, 8 conditions FIR design matrix analysis The functional brain images from the 21 controls and 21 patients, 8 conditions FIR design matrix analyses can be found in Figure 14. This set of analyses once again demonstrates a divergence in the results obtained from each of the statistical methods even when identical design matrices are used. The pattern of brain activity identified in the SPM5 analysis is consistent with the SPM5 results obtained using other design matrices and samples. CPCA analysis identified a similar cluster of dAcc/SMA activity as the SPM5 analysis; however, some of the brain regions identified as being activated in the SPM5 analysis, including posterior 52  cingulate/precuneus and right middle frontal gyrus were found to have a more complex relationship in the CPCA analysis, in which the valence of the activity varied as a function of condition and time. The ICA analysis, revealed an occipital lobe component that is highly similar to components identified in the previous analyses using the same sample with an HRF design matrix, as well as some of the FIR design matrix analyses presented above that used different samples. The beta weight pattern found in the ICA component suggests that there is no effect of condition until the latter half of the trial, where the Recalled and Non-recalled betas diverge, with the controls showing a greater deactivation in the Non-recalled betas than the patients. The PLS results display a pattern of deactivations highly consistent with those found in the SPM5 analysis, as well as many of the previous analyses presented in this section. This pattern of deactivations, which can be found in medial prefrontal cortices, superior and middle temporal gyri, and posterior cingulate/precuneus regions, has been termed the “default mode” network (Raichle et al., 2001), and this network is thought to reduce in activation when a demanding cognitive task needs to be performed. Neuroimaging Results – Multiple Component Analyses Although many of the voxels and/or networks identified as being active in the previous set of analyses differ on the basis of the statistical technique, sample, or design matrix used; a part of this disparity can be attributable to the fact that only the first principal component, or the highest regressing component were displayed. In the following section, the results of the previous multivariate analyses are redisplayed with multiple components from each analysis superimposed upon a single structural brain image in order to demonstrate that the previously presented, seemingly incongruous results were, in fact, tracking the same signal in the data. For this set of analyses, only the results for the 8 condition design matrices will be displayed, as 53  these analyses produced the most divergent results between the multivariate methods in this study. The functional brain images from these analyses clearly show the correspondence between the three statistical methods in identifying the common neural signal emerging from the data. The statistics assessing the significance of the components as well as the significant differences between groups, conditions, and time points can be found in the Appendix. In each of the analyses depicted here (in Figure 15, Figure 16, Figure 17, Figure 18, Figure 19, and Figure 20), it can be seen that similar regions of activity and de-activity are found whether CPCA, ICA, or PLS are used to analyze the data. The results of the SPM5 analyses using the same design matrix and sample are shown with the multivariate results for the purpose of comparison. The components in Figures 15-20 have been colour-coded, such that the first component (e.g. explaining the most variance in the CPCA and PLS analyses, and regressing most strongly (i.e. highest R-square value) on the design matrix in ICA) extracted is depicted in red, the second component is depicted in blue, the third component is depicted in green and the fourth component is depicted in purple. 21 controls, 8 conditions HRF design matrix analysis The functional brain images from the 21 controls 8 conditions HRF design matrix analyses can be found in Figure 15. This multicomponent analysis begins to reveal some of the correspondence between the methods that was lost when only a single component was extracted. Notably, SPM5, CPCA, and ICA all show common foci of activity in many regions including dAcc/SMA, mPFC, insula, and occipital cortices, although the extraction order of the components in the CPCA and ICA analyses differed. The predictor weights from the CPCA component 1 indicate that the deactivations in that network were condition dependent, as both Read conditions were more strongly deactivated than any of the other experimental conditions; 54  however the Read Recalled condition was much less deactivated than the Read Non-recalled condition. Visual examination of the predictor weights suggests that the significant effect of Difficulty in the first CPCA component was driven by the deactivations in the Read conditions. The predictor weights from the second and third component extracted in the CPCA analysis are characterized by increased activity in the Recalled conditions relative to the Non-recalled conditions. Since the first ICA component shared foci of activity with the second (dAcc/SMA) and third (insula) CPCA components, it is notable that the beta weights associated with the Nonrecalled stimuli are greater than those of the Recalled stimuli, although this difference may be attributable to the influence of the non-overlapping brain areas. Another notable finding is that the ICA and CPCA components with negative beta/predictor weights (ICA component 2 and CPCA component 1) both find that the Internal Recalled condition shows the least deactivation. This suggests that the right temporo-parietal junction, the overlapping region included in these components, is less deactivated when evaluating self-generated information, which agrees with previous research indicating the importance of brain regions other than the mPFC in evaluating self generated information (Jardri et al., 2007). 21 controls, 8 conditions FIR design matrix analysis The functional brain images from the 21 controls 8 conditions FIR design matrix analyses can be found in Figure 16. In this set of analyses as well, it can be seen that there are broad correspondences in the neural networks detected by each of the statistical methods. For instance, component 4 from the CPCA analysis and component 2 from the PLS analysis are highly similar, and both resemble the pattern of deactivations found in the SPM5 analysis. The areas detected by these three methods appear to correspond to the default mode network (Raichle et al., 2001). 55  Numerous components from multiple methods share distinct foci of activity. For instance, there is considerable overlap in the occipital cortex activity (component 2 from the CPCA analysis and component 1 from the ICA analysis), as well as in the dAcc and insula (component 3 from the CPCA analysis, component 2 from the ICA analysis, and component 1 from the PLS analysis). However, although these methods identify the same signal in the brain, there are notable differences in how the activity in these components should be assessed as the data used to interpret the contribution of each of the conditions to the pattern of brain activity, (i.e., predictor weights (CPCA), beta weights (ICA), and design scores (PLS)) seem to offer different interpretations. For example, with regards to the dAcc and insula component, the CPCA predictor weights suggest that the Recalled stimuli elicit greater neural activity than Non-recalled stimuli, whereas the ICA beta weights suggest that the Non-recalled stimuli are a better fit to the component activity time course than the Recalled stimuli. The PLS design scores suggest that the network activity is strongest in the External Non-recalled condition. Additionally, even though the occipital cortex activity from the CPCA and ICA analyses appears to overlap very closely, the CPCA predictor weights suggest that the network activity peaks at 6 seconds peristimulus time, whereas the beta weights from ICA suggest that the activity in this network peaks at 4 seconds peristimulus time. These conflicting results can be partially explained by noting that the components are not identical, however, it is surprising to find large timing differences in basic sensory regions. 21 patients, 8 conditions HRF design matrix analysis The functional brain images from the 21 patients 8 conditions HRF design matrix analyses can be found in Figure 17. This analysis once again highlights the high degree of correspondence between the functional networks extracted by SPM5, CPCA, and ICA. It is 56  interesting to note that the brain regions in components 2 and 3 from the ICA analysis correspond roughly to components 1 and 3 in the CPCA analysis. The predictor weights from CPCA component 1 shows a clear effect of stimulus type which differs from both CPCA component 2 and 3 and ICA components 2 and 3, in that the latter suggest a similar response regardless of stimulus type, although they do show an effect of Recall. The conditions from CPCA component 1 that do not show the same level of deactivation as the other conditions/components are the Internal Recalled/Non-recalled, and External Non-recalled conditions, which once again suggests that this network is sensitive to differences in recalling self and other generated information. In general, it appears that the CPCA components have stronger predictor weights (e.g. farther from zero) in the Recalled conditions, whereas the ICA components have stronger beta weights (e.g. farther from zero) in the Non-recalled conditions. 21 patients, 8 conditions FIR design matrix analysis The functional brain images from the 21 patients 8 conditions FIR design matrix analyses can be found in Figure 18. Firstly, it appears that there is less correspondence in the patient sample than the control sample when using an FIR design matrix. Once again, the occipital lobe components (CPCA component 2 and ICA component 1) agree as to the location of the signal but disagree as to its time course, with the CPCA analysis identifying the activity peak as occurring one TR later than where the peak is found in the ICA analysis. Component 3 from the ICA analysis appears to differentiate the Recalled Source conditions from the Recalled Task conditions after 4 seconds peristimulus time; however, in component 3 from the CPCA analysis, which overlaps with the ICA component, we do not see this same clear differentiation between Source and Task. In the opposite situation, component 4 from the CPCA analysis appears to distinguish between Source and Task Recalled between 4 and 8 seconds peristimulus time, but 57  the ICA component (component 4) that overlaps with this CPCA component does not identify this same distinction. In the PLS analysis, the first component is characterized by sparse activations that do not agree with the other analyses, or any of the previous PLS analyses, and the second component identified a network that resembled the default mode network but this component did not account for enough variance during permutation tests to be deemed significant. 21 controls and 21 patients, 8 conditions HRF design matrix analysis The functional brain images from the 21 controls and 21 patients 8 conditions HRF design matrix analyses can be found in Figure 19. Although the CPCA and SPM analyses correspond well in terms of the functional activations/deactivations, the brain images produced from the ICA analysis appear sparse relative to the others. Visual inspection suggests that, although there are some differences between the patients and controls in terms of the predictor or beta weights, most of these differences are small in magnitude and only found in a single experimental condition. For the CPCA analysis, the predictor weights for component 4 appear to best distinguish between patients and controls, as the predictor weights for the control subjects are all positive in valence whereas for the patients, the predictor weights for the Source Recalled and Read Recalled conditions have negative valence. For the ICA analysis, the biggest difference between patients and controls appears to be found in components 2 and 4. In component 2, the beta weights for the patients are larger than those of the controls (e.g. farther from zero), whereas in component 4, the beta weights for the patients reveal smaller deactivations relative to the controls. Also, for component 1 from the CPCA analysis, there is once again a difference between the predictor weights for the Internal Recalled condition relative to the any other  58  condition, such that the Internal Recalled condition has positive predictor weights in both the patients and controls whereas the other conditions have negative beta weights. 21 controls and 21 patients, 8 conditions FIR design matrix analysis The functional brain images from the 21 controls and 21 patients 8 conditions FIR design matrix analyses can be found in Figure 20. This analysis once again shows that the brain networks involved in performing self-other source monitoring tasks do not display gross differences between schizophrenia patients and healthy controls. Although the neural networks extracted using each statistical method show some overlap, this is less pronounced in this analysis than in the previous multiple component analyses. The CPCA predictor weights from component 2 indicate that, once again, Non-recalled stimuli elicit smaller (closer to zero) predictor weight peaks than the Recalled stimuli. For the ICA components, the beta weights indicate a clearer (though non-significant) difference between patients and controls for components 1 and 2. For component 1, the beta weights for the controls are greater than those of the patients (farther from zero) both when the network is activated (e.g., between 0 to 6 seconds peristimulus time), and when it is deactivated (e.g., between 8 to 14 seconds peristimulus time). For component 2, the beta weights for the patients are greater than those of the controls throughout the course of the trial. The occipital cortex components for the CPCA and ICA analyses (CPCA component 1 and ICA component 2), disagree in terms of which group has the higher predictor/beta weights, with the CPCA analysis indicating that the patients have a higher peak and generally higher activity through the course of the trial, whereas the ICA analysis shows that the beta weights for the control subjects have a higher positive peak than the patient subjects. With the exception of component 1, the other ICA components were characterized by more condition specific 59  variability in the Non-recalled conditions than in the Recalled conditions. Within each ICA component, the Recalled conditions appear to show fewer condition specific differences. For PLS component 1, the design scores suggest that for the controls, the Association Non-recalled and External Non-recalled conditions are most active, whereas for the patients, it is the Associate Non-recalled and Internal Recalled conditions, which suggests that this component is important for distinguishing Self from Other generated stimuli. For PLS component 2, the design scores for the controls were highest in the Association Non-recalled and External Non-recalled conditions whereas for the patients they were highest in the Read Non-recalled condition, thus rendering this finding difficult to interpret.  60  Discussion The primary aim of this study was to assess the differences in brain activity between healthy controls and schizophrenia patients while performing the recollection portion of a source monitoring memory paradigm. The secondary aim was to identify whether the recollection of words from the various encoding conditions would result in different patterns of brain activity. Multiple statistical fMRI data analysis techniques were employed in the course of this investigation in order to determine whether the there were differences between the patients and controls (or experimental conditions) in terms of either focal brain activations or functionally connected networks. Therefore, a tertiary aim of this project was to determine whether the brain networks and activations identified by the various statistical techniques converged upon a common solution, or whether they produced widely discrepant results. With regard to the primary aim of this study (i.e., assessing differences between controls and patients), our hypothesis that patients would show higher activity in the functional networks related to performing the task was not supported as there were no main effects of Group in any of the analyses reported herein, suggesting that the differences in self-other source monitoring abilities in schizophrenia patients and healthy volunteers do not rely on wholly different brain networks for this cognitive function, nor do the activity levels in these networks reveal a significant and consistent group difference. The lack of significant differences between schizophrenia patients and healthy controls in this study mirrors the findings of a recent review of experiments examining corollary discharge during inner speech in schizophrenia (Allen et al., 2007), which found that there was only ambiguous evidence for a verbal self-monitoring deficit in schizophrenia, whereby only a subset of the studies examined managed to find any significant differences between the two groups. Potential limitations of this study that may have contributed  61  to the lack of significant differences between controls and patients will be discussed in a subsequent section. With regards to the secondary aim of this study (i.e. determining whether different experimental conditions elicited variable levels of activity or involved different functional networks), many differences between conditions were found in all of the analysis methods used in this study. This section will describe some of these differences as well as some general remarks about the notable findings from each of the methods. With regards to the CPCA results from the analyses that coded all 8 conditions into the design matrix, one notable finding is that there is considerable overlap between the networks extracted from each of the CPCA analyses performed in this investigation. In each of the multiple component CPCA solutions presented herein, one can find a component that resembles the task-positive network (i.e., dAcc/SMA, DLPFC, insula), and another that resembles the default mode network (i.e., mPFC, posterior cingulate/precuneus). In the CPCA analyses, we consistently find that the task-positive network component is characterized by higher activity in the Recalled as opposed to Non-recalled conditions. For the CPCA FIR analyses, the Internal Recalled condition also appears to have the lowest estimated hemodynamic responses of all the experimental conditions for the control subjects; however, for the patients, both the Internal and External Recalled conditions have lower estimated hemodynamic responses than the rest of the experimental conditions. Although these differences are non-significant, the consistency of these findings is of note. However, in the CPCA HRF analyses, the match between the obtained data and the synthetic HRF model does not seem to be much different in the Internal Recalled condition than the other Recalled conditions. In the default mode components, the Internal Recalled condition shows less deactivation relative to the other experimental conditions, which  62  supports the view that the mPFC is a critical region supporting successful reality monitoring (Simons et al., 2006), as the mPFC is a major hub in the default mode network. However, it cannot be ruled out that the other brain regions in this network also play an important role in reality monitoring, as the ability to evaluate the contributions of individual clusters to the effects seen in the overall network is still being developed. With regards to the ICA results, one remarkable feature is the consistency of the extracted components regardless of which sample was used. For instance, the occipital cortex component appears to be surprisingly well conserved between the patients and controls. In most of the ICA analyses as well, there was a component that included the dAcc and bilateral insula. However, although this component is similar to the task-positive components extracted in the CPCA analysis, the beta weights from the ICA analyses show that this component is a better match to the design matrices (e.g., higher beta weights) in the Non-recalled conditions than the Recalled conditions, which is a finding opposite to that suggested by the CPCA analysis. However, the increased beta weights for this component accords with previous research implicating dAcc activity in conditions where response inertia must be changed, such as when an incorrect response is made (Kiehl et al., 2000; van Veen and Carter, 2002; Woodward et al., 2008). Although this feature of the ICA beta weights is most easily visible in the HRF analyses, the second component from the ICA 21 patients FIR analysis is also characterized by much larger beta weights in the Non-recalled conditions. The fact that the ICA and CPCA components identify a similar network but ascribe different properties to it could be due to several factors. First, it could be due to the fact that ICA and CPCA use different mathematical operations to identify functional networks, or that ICA identifies networks prior to incorporating information about the experimental paradigm. Second, it could be that the firing of different subpopulations  63  of neurons are responsible for the two different signals being identified but either the spatial resolution of fMRI is too coarse, or these subpopulations could be fed by the same arteries thus making them virtually indistinguishable when looking at the resultant metabolic brain activity. Thirdly, it could be the case that the behaviour of some of the regions in this network (e.g., dAcc) is dependent on the other brain regions to which it is functionally connected. Another notable feature that emerges from the ICA HRF analyses is that there is less variance between the conditions than was previously seen in the CPCA HRF analyses with the various conditions showing a similar response profile (apart from the Recalled/Non-recalled distinction mentioned above). With regards to the PLS results from the analyses that modeled all 8 conditions, the components extracted from the controls and patients data set do not show a great deal of correspondence, in terms of the regions found to be active in the functional networks, or in the design scores that depict the involvement of each condition in these functional networks. However, there was some regional overlap between the two functional networks including mPFC, middle frontal gyrus, inferior parietal lobe, and cerebellum. In general, it can be said that the functional network identified in the control subjects appears to be comprised of regions often described as the default mode network (Fox et al., 2005). However, the pattern of activity in the patients‟ functional network appears to be much less well-characterized. The PLS analyses did not display widespread agreement between the patient and control analyses in terms of the design scores. For the controls, the conditions where the functional network was most active was in the associate condition (recalled and non-recalled) and internal recalled. The conditions where this network was most deactivated were external non-recalled and read non-recalled. For the patients, the conditions in which this network was most active was during the read non-recalled condition;  64  and this network was most deactivated in the associate non-recalled, external non-recalled and internal non-recalled conditions. Taken together, the design score pattern and the overlap in active brain regions suggests that these regions are recruited to play different functional roles within each network, or that these areas serve more general cognitive functions that can be recruited to different functional networks. With regards to the tertiary aim of this project (i.e. whether the results from the various statistical methods agreed with one another, overall, this study revealed that the univariate and multivariate statistical methods employed converged on a common signal within the dataset. Although there were differences in the results within and between analytical methods, overwhelmingly, the same brain regions emerged as being active during the course of the experiment. This is most apparent when the examining the multiple component functional brain images, which reveal a highly similar pattern of brain activity regardless of statistical method employed. What was most interesting in this regard was the fact that although many of the same regions were detected by each of the statistical methods, the manner in which these regions were connected as networks was not consistent between them, nor were the estimates of activity for those networks, even in cases where the brain regions in different networks were highly conserved. These differences are likely to be at least partially due to differences in how these networks are identified. As has been mentioned earlier, ICA identifies brain networks prior to incorporating any information about stimulus timing or experimental design, whereas CPCA and PLS both utilize this information prior to extracting functional networks. This difference in the stage of processing in which the experimental design information is incorporated (see Table 1) is likely to be responsible for some of the differences found in this investigation, as the networks identified by ICA are not optimized to be related to the experimental tasks, and CPCA and PLS  65  are limited by the accuracy of the model used to characterize the expected pattern of evoked neural activity in the experiment. As was mentioned in the introduction, several brain regions have been implicated in the performance of source monitoring, including the prefrontal cortices, the medial temporal lobes, and the parietal lobes. This experiment found that all of these regions were active in a variety of conditions. In the CPCA, ICA, and PLS analyses, the middle and inferior frontal gyri, two regions located in the DLPFC, were frequently found to be part of a network including dAcc/SMA and bilateral insula. Using both CPCA and ICA, this network showed increased activity relative to baseline in both the HRF and FIR analyses although the CPCA and ICA networks that included these DLPFC regions disagreed as to the role of the network in different conditions, as described above. mPFC activity was also found in many of the analyses, regardless of statistical methodology employed, and was frequently found to be part of a network which included posterior cingulate/precuneus, suggesting that the analyses were identifying the default mode network. Contrary to our expectations, evidence for medial temporal lobe involvement was limited in this experiment. Although there was hippocampal and parahippocampal activity detected in some of the analyses (notably the PLS and CPCA 21 control analyses and the PLS 21 patient analysis), this activity was not found consistently in every analysis, nor did it appear to be part of a consistent network as the other regions found in the same component varied between analyses. Parietal lobe activity was found in almost every analysis, but the precise anatomical foci of these activations/deactivations varied from analysis to analysis, suggesting that different areas in the parietal lobe may be supporting different specialized functions throughout the course of the trial, which is not surprising given the sheer variety of cognitive tasks the parietal lobe is hypothesized to support, including visual attention  66  (Nachev and Husain, 2006), action planning and decision making (Andersen and Cui, 2009), spatial cognition (Sack, 2009), and self-motion perception (Britten, 2008). Another brain region that was frequently found to be active during the performance of the source monitoring task using multiple statistical methods was the cerebellum, which is of interest because it is a key node in the brain network thought to underlie cognitive dysmetria (Andreasen et al., 1998). The theory of cognitive dysmetria proposes that the diverse symptoms of schizophrenia arise due to impairments in the sequencing and coordination of sensorimotor and cognitive processes, and that this lack of coordination results from abnormalities in the circuit connected prefrontal cortices, thalamus, and the cerebellum. Although the statistical methods used in this report did not appear to differ drastically in terms of the brain networks identified, there were notable differences in how each method is implemented and used. Firstly, although the components that were extracted were very similar, they were not identical, thus making it difficult to appraise any differences in activity time courses or differences in the conditions, as it is reasonable to expect that even slightly different networks could have different properties with regard to the estimates of condition specific activity. Furthermore, although we had hypothesized several brain regions that were likely candidates for involvement in the performance of the experimental task, it is impossible to say which of the analyses came closest to identifying the real pattern of neural activity in the data. With these caveats aside, however, it is still informative to examine some of the notable differences in terms of how each of the methods may be used, and what sort of information they provide to the user. As SPM is the most widely used fMRI data analysis software package in the world, the SPM5 analyses provided the benchmark by which the other methods were assessed. However, it is not without its shortcomings. The SPM5 approach is very useful for identifying  67  discrete brain regions where differences in neural activity can be found, however, the choice of experimental condition to use as a contrast greatly influences the results obtained. Please see the Appendix for the SPM5 contrasts that yielded significant results. The CPCA program is relatively flexible with regards to the choice of design matrix as it allows any matrix that is invertible and matches the number of scans entered into the dataset. Also, the resemblance between the predictor weights and hemodynamic responses increases interpretability of the components and lends support to the overall method. However, in its current implementation, this program does not attempt to correct for correlations over time (e.g. rows of G); nor does it attempt to remove any slow fluctuations related to physiological processes (e.g. using a high pass filter), although it should be noted that the removal of these signals is currently being implemented. This means that there are potential confounding factors which should be separated from the experimental effects. Also, since CPCA uses PCA to decompose the data, poor choice of window size or experimental design can lead to intractable problems in identifying functionally relevant components. Furthermore, since this method can only produce a single image of the brain system accounting for a large proportion of the variance in the data for each component, differences between groups must be assessed using the predictor weights, which can lead to problems if, in fact, different groups utilize different brain systems to perform an identical task. However, variants of the CPCA analysis presented here are being produced that will allow the examination of differences in functional activation between groups. The ICA analysis provides an elegant and model free method of decomposing functional imaging data into its constituent components, yet the choice of which algorithm to use and how many components to extract remains unclear. Also, the determination of which components are functionally important is difficult without an a prori network of interest, or the use of a design  68  matrix coding the events of experimental interest used to identify the components that are most strongly involved in the experimental tasks. Additionally, given that the components extracted by ICA are identified without any knowledge of the experimental conditions, it is possible that it would identify highly coordinated and stable brain networks that have little to do with the tasks of experimental interest. Although the PLS software package is the only method that produces a separate brain image for each time point modeled in the window, it only produces a single value (i.e. design score) indicating the importance of the component to each of the conditions which makes it difficult to interpret how the changes in brain activity throughout a time window are related to the experimental conditions of interest. Also, it is inflexible with regards to choice of design matrices, as it is does not use a design matrix in the conventional sense. This limitation rendered it impossible to run anything comparable to the HRF analyses performed with the other methods. Additionally, since PLS averages all the trials of a given experimental condition for a particular subject and then expresses the BOLD changes in peristimulus time as percentage signal change from stimulus onset, it seems particularly vulnerable to difficulties arising from the use of inappropriate lag windows, or in experiments where short inter-trial intervals that do not allow the BOLD response to return to baseline before the presentation of a new stimulus, which was the case in the current study. This study contained several limitations that may have contributed to the inability to detect significant differences between the healthy controls and the schizophrenia patients. Firstly, the same list of 120 words was used for both the encoding and recall portions of the experiment. Although the use of 30 items per source exceeds the number of items per source found in many other studies, the lack of foils meant that guessing rates could not be corrected for in this  69  experiment. This inability to account for guessing strategies is known to present a problem in the interpretation of the results of source monitoring studies, as real differences between the groups can be disguised by strategic responding, for instance, ensuring that each condition receives an equivalent number of responses (Woodward et al., 2007). The inclusion of foils would have allowed for the quantification of false positives (i.e., cases where participants responded with a condition name to „new‟ stimuli), which would have allowed the use of analysis of covariance (ANCOVA) to remove variance related to these types of guesses from the results. Secondly, this lack of a proper means to control for guessing is compounded by the fact that the recall portion of the experiment involved making forced choice responses with no way to indicate differences in judgment confidence. This meant that it was impossible to distinguish between legitimate incorrect responses, guesses, and inadvertent button presses. Thirdly, the experimental design for this study used very short inter-trial intervals (ITI‟s) of 1, 2, or 3 seconds. The rapidity of this design meant that the hemodynamic response to any particular stimulus would be unlikely to return to baseline prior to the presentation of the next stimulus (Aguirre et al., 1998; Handwerker et al., 2004), which may have led to difficulties in uniquely attributing a hemodynamic response to a particular stimulus, especially for PLS. Although it should be noted that ten second „break‟ trials were included to allow the hemodynamic response to return to baseline, and reduce multicolinearity between the conditions. Fourthly, the data from approximately one third of the subjects who participated in this study were excluded from this analysis due to excessive head movements, problems with experimental administration, and/or a lack of incorrect trials in certain conditions. Although the first problem is often encountered when scanning clinical populations, the latter two were unique to this experiment. The lack of incorrect trials in particular conditions meant that there were empty vectors in the design matrices, which resulted  70  in an inability to perform the inversion necessary for the regression analysis. These subjects had to be removed from this analysis, which may have impacted the detection of significant differences between conditions and/or groups. The lack of incorrect responses for some subjects suggests that ceiling effects may have been a potential problem in this experiment, as previous studies have noted that increasing task difficulty in reality monitoring experiments leads to increases in externalization errors (Larøi et al., 2004), and a lack of difficulty in this experiment may have helped to disguise differences between controls and patients. One possible method to alleviate these ceiling effects would have been to increase the time between the Encoding and Recall portions of the experiment, thus allowing for the memory trace to be degraded over time or interfered with by the re-encoding of the target words in a naturalistic (i.e., non-experimental) setting. Additionally, the removal of such a large proportion of the subjects meant that there was a significant loss of power in the experiment, such that the lack of significant differences between the groups and conditions may be attributable to Type II error. Lastly, only the Recall portion of the experiment was scanned, which suggests the possibility that significant differences between patients and controls may have been detected had they been scanned during Encoding. However, since the cognitive operations that underlie Encoding and Recall are likely to be highly overlapping, it is unclear how efficacious scanning both portions of the experiment would have been. Lastly, this study was not designed to assess differences in accuracy between statistical methods of fMRI data analysis, therefore there is no principled way to determine which of the methods comes closest to capturing the true signal in the data. Clearly, any attempt to directly compare these methods would require either a data simulation experiment, in which an artificial  71  signal of known magnitude and anatomical location is inserted into a dataset, or an extremely simple sensory detection paradigm in which the brain response is well characterized.  72  Conclusion The goal of this study was to examine differences brain activity between healthy controls and schizophrenia patients while performing self-other source monitoring, as well to examine similarities and differences in how various data analysis methods were able to identify the signal in the data. Although there were no significant differences detected between the patients and controls in this study, there was considerable agreement between the statistical methods used to analyze the data in terms of the brain regions identified as being involved in the performance of the experimental tasks. However, the differences between the statistical methods revealed themselves in terms of how these regions were arranged into coordinated brain networks, as well as in the level of involvement of each of these networks in the different experimental conditions. From the many analyses performed in this project, it seems as though SPM is most useful when one wishes to find a differences in activity in discrete brain regions when closely matched experimental conditions are able to be employed. SPM would not be recommended for any type of trial that requires multiple cognitive stages (e.g. working memory tasks) as only a single pattern of activity is able to be detected, and it is difficult to determine how the pattern of activity relates to any of the dicrete cognitive stages. PLS is best used in situations where one would like to visualize how brain activity changes over the course of a trial. It is ideally suited for this type of analysis as it produces different brain images for each time point in the lag window. It should be noted that PLS seems particularly susceptible to analysis problems arising from experimental designs with short inter-trial intervals. ICA is best used in situations where there is no design matrix (resting state studies) as well as in situations where there is an a priori network of interest to investigate as these situations capitalize on ICA‟s ability to blindly identify brain networks. The use of regression in ICA to identify experimentally related networks was moderately  73  successful but there were still many components that did not appear to be functionally relevant that had high R-square values. CPCA is ideally used for multistage experiments where the hemodynamic response in each network needs to be estimated, rather than assumed to be of a particular shape. CPCA has been most successful in cases where the cognitive load can be varied on a trial to trial basis. Since CPCA is limited to producing a single brain image for each network, there is no evidence that it would be of use when the experimenter wishes to know how the brain network changes over the course of a trial, or when the experimenter wishes to investigate whether different groups utilize different brain networks to perform the same cognitive task. However, none of the methods employed in this study found group differences, suggesting that perhaps no group differences were to be found in this experiment. In that case, it is remains an open question as to whether CPCA would detect group differences if they were present in the data.  74  Table 1. Important differences between the statistical analysis methods used in this project.  Method  General Approach  Information about Task Design  Theory or Data Driven  Analysis of Task Irrelevant Variance  Multivariate Pattern Extraction  Combination of Variables  SPM5  Univariate  Yes, design matrix is used for GLM regression  Theory  No  None  Linear  CPCA  Multivariate  Yes, design matrix is used for GLM regression  Both  (1st step in analysis)  ICA  Multivariate  Yes, but only for post-hoc regression and correlation  Data  (2nd step in analysis)  PLS  Multivariate  Yes, structure of the data matrix takes into account the experimental design st  (1 step in analysis)  Both  Possible, in separate analyses  Yes, in same analysis  No  Principal Component Analysis  Linear  (2nd step in analysis) Independent Component Analysis  Non-Linear  (1st step in analysis)  Principal Component Analysis  Linear  nd  (2 step in analysis)  75  Table 2. Group Demographics (standard deviations in parentheses).  Age Gender Handedness  Controls  Patients  27.81 (5.83)  29.57 (9.35)  9 males; 12 females  14 males; 7 females  All right-handed  All right-handed  Quick = 102.33 (9.74)  Quick = 102.80 (12.45)  NAART = 22.43 (8.87)  NAART = 21.95 (9.28)  41.05 (18.96)  46.20 (20.70)  Premorbid IQ Parental SES  Note. SES = Socioeconomic Status; NAART = North American Adult Reading Test  76  Table 3. Mean accuracy in percent correct for each task type by group (standard deviations in parentheses). Controls  Patients  Association  80.63 (18.09)  73.33 (18.07)  Read  78.57 (13.93)  70.95 (21.99)  Self (Internal)  80.79 (13.16)  76.35 (14.33)  Other (External)  82.06 (14.55)  72.54 (17.12)  Task  79.60 (12.53)  72.14 (17.36)  Source  81.43 (10.64)  74.44 (12.77)  Overall  80.52 (11.00)  73.29 (13.63)  Note: Task = (Association + Read)/2; Source = (Self + Other)/2  77  Table 4. Mean response times in milliseconds for each task type by group (standard deviations in parentheses). Condition  Controls Non-recalled  Patients Recalled  Non-recalled  Recalled  Association  1309.22 (362.72)  949.18 (265.17)  1188.99 (368.29)  1016.81 (285.01)  Read  1395.62 (316.75)  1159.57 (222.76)  1322.46 (371.07)  1150.19 (336.73)  Self (Internal)  1204.71 (325.52)  836.89 (242.03)  1206.06 (299.81)  918.71 (243.85)  Other (External)  1168.97 (351.71)  1058.14 (239.54)  1214.94 (304.06)  1050.84 (285.00)  Task  1352.42 (276.45)  1054.38 (223.85)  1255.73 (318.74)  1083.50 (291.55)  Source  1186.84 (270.52)  947.52 (226.65)  1210.50 (277.13)  984.77 (253.22)  Overall  1269.63 (233.69)  1000.95 (217.63)  1233.11 (265.58)  1034.14 (259.97)  Note: Task = (Association + Read)/2; Source = (Self + Other)/2  78  Figure 1. Schematic of how the PLS BOLD data vector is obtained from a single subject‟s conventional BOLD data matrix.  79  Figure 2. Examples of encoding tasks  80  Figure 3. Functional brain images from the single subject, single condition (any stimulus) analysis using an HRF model SPM5 - AR(1) & High Pass Filter  SPM5 - No Autoregressor & No Filter  CPCA  ICA  81  Figure 4. Functional brain images from the single subject, single condition (any stimulus) analysis using an FIR model SPM5 - AR(1) & High Pass Filter  SPM5 - No Autoregressor & No Filter  CPCA  ICA  PLS  82  Figure 5. Functional brain images from the 21 controls, single condition (any stimulus) analysis using an HRF model SPM5 AR(1) & High Pass Filter  CPCA  ICA  83  Figure 6. Functional brain images from the 21 patients, single condition (any stimulus) analysis using an HRF model  SPM5 AR(1) & High Pass Filter  CPCA  ICA  84  Figure 7. Functional brain images from the 21 controls, single condition (any stimulus) analysis using an FIR model  SPM5 AR(1) & High Pass Filter  CPCA  ICA  PLS  85  Figure 8. Functional brain images from the 21 patients, single condition (any stimulus) analysis using an FIR model  SPM5 AR(1) & High Pass Filter  CPCA  ICA  PLS  86  Figure 9. Functional brain images from the 21 controls, 8 condition analysis using an HRF model SPM5 AR(1) & High Pass Filter  CPCA  ICA  87  Figure 10. Functional brain images from the 21 patients, 8 condition analysis using an HRF model SPM5  CPCA  ICA  88  Figure 11.Functional brain images from the 21 controls, 8 condition analysis using an FIR model SPM5  CPCA  ICA  89  PLS  90  Figure 12. Functional brain images from the 21 patients, 8 condition analysis using an FIR model SPM5  CPCA  ICA  91  PLS  92  Figure 13. Functional brain images from the 21 controls and 21 patients, 8 condition analysis using an HRF model SPM5  CPCA  ICA  93  Figure 14. Functional brain images from the 21 controls and 21 patients, 8 condition analysis using an FIR model SPM  CPCA  ICA  94  PLS  95  Figure 15. 21 control multicomponent solution with an HRF design matrix. Each method lists the number of components used to create the functional brain image. SPM5 – all greater than baseline  CPCA – 3 components  96  ICA – 4 components  97  Figure 16. 21 control multicomponent solution with an FIR design matrix. Each method lists the number of components used to create the functional brain image. SPM5 – all greater than baseline  CPCA – 4 components  98  ICA – 2 components  99  PLS – 2 components  100  Figure 17. 21 patient multicomponent solutions with an HRF design matrix. Each method lists the number of components used to create the functional brain image. SPM5 – all greater than baseline  CPCA – 3 components  101  ICA – 3 components  102  Figure 18. 21 patient multicomponent solutions with an FIR design matrix. Each method lists the number of components used to create the functional brain image. SPM5 – all greater than baseline  CPCA – 4 components  103  ICA – 4 components  104  PLS – 2 components  105  Figure 19. 21 control & 21 patient multicomponent solutions with an HRF design matrix. Each method lists the number of components used to create the functional brain image. SPM5 – all greater than baseline  CPCA – 4 components  106  ICA – 4 components  107  Figure 20. 21 control & 21 patient multicomponent solutions with an FIR design matrix. Each method lists the number of components used to create the functional brain image. SPM5 – all greater than baseline  CPCA – 4 components  108  ICA – 4 components  5  109  PLS – 2 components  110  References  Aguirre GK, Zarahn E, D'Esposito M (1998) The Variability of Human, BOLD Hemodynamic Responses. Neuroimage 8:360-369. Aleman A, Böcker KBE, Hijman R, de Haan EHF, Kahn RS (2003) Cognitive basis of hallucinations in schizophrenia: role of top-down information processing. Schizophrenia Research 64:175-185. Allen P, Aleman A, McGuire PK (2007) Inner speech models of auditory verbal hallucinations: Evidence from behavioural and neuroimaging studies. International Review of Psychiatry 19:407-415. Andersen RA, Cui H (2009) Intention, Action Planning, and Decision Making in Parietal-Frontal Circuits. Neuron 63:568-583. Andreasen NC, Paradiso S, O'Leary DS (1998) “Cognitive Dysmetria― as an Integrative Theory of Schizophrenia: A Dysfunction in Cortical-Subcortical-Cerebellar Circuitry? Schizophrenia Bulletin 24:203-218. Barnes J, Boubert L, Harris J, Lee A, David AS (2003) Reality monitoring and visual hallucinations in Parkinson's disease. Neuropsychologia 41:565-574. Bell AJ, Sejnowski TJ (1995) An information-maximization approach to blind separation and blind deconvolution. Neural Comput 7:1129-1159. Bianciardi M, Fukunaga M, van Gelderen P, Horovitz SG, de Zwart JA, Shmueli K, Duyn JH (2009) Sources of functional magnetic resonance imaging signal fluctuations in the human brain at rest: a 7 T study. Magnetic Resonance Imaging 27:1019-1029. Biswal BB, Kannurpatti SS, Rypma B (2007) Hemodynamic scaling of fMRI-BOLD signal: validation of low-frequency spectral amplitude as a scalability factor. Magnetic Resonance Imaging 25:1358-1369. Blakemore S-J, Wolpert DM, Frith CD (2000a) Why can't you tickle yourself? Neuroreport 11:R11-R16. Blakemore S-J, Smith J, Steel R, Johnstone EC, Frith CD (2000b) The perception of selfproduced sensory stimuli in patients with auditory hallucinations and passivity experiences: evidence for a breakdown in self-monitoring. Psychological Medicine 30:1131-1139. Brebion G, Amador X, David A, Malaspina D, Sharif Z, Gorman JM (2000) Positive symptomatology and source-monitoring failure in schizophrenia - an analysis of symptom-specific effects. Psychiatry Research 95:119-131. 111  Brébion G, Smith MJ, Gorman JM, Amador X (1996) Reality monitoring failure in schizophrenia: The role of selective attention. Schizophrenia Research 22:173-180. Britten KH (2008) Mechanisms of Self-Motion Perception. Annual Review of Neuroscience 31:389-410. Burianova H, McIntosh AR, Grady CL (2010) A common functional brain network for autobiographical, episodic, and semantic memory retrieval. Neuroimage 49:865-874. Calvert GA (2001) Crossmodal Processing in the Human Brain: Insights from Functional Neuroimaging Studies. Cereb Cortex 11:1110-1123. Christoff K, Ream JM, Geddes LPT, Gabrieli JDE (2003) Evaluating Self-Generated Information: Anterior Prefrontal Contributions to Human Cognition. Behavioral Neuroscience 117:1161-1168. Chuang K-H, Chen J-H (2001) IMPACT: Image-based physiological artifacts estimation and correction technique for functional MRI. Magnetic Resonance in Medicine 46:344-353. Cusack R (2005) The Intraparietal Sulcus and Perceptual Organization. Journal of Cognitive Neuroscience 17:641-651. Daffertshofer A, Lamoth CJC, Meijer OG, Beek PJ (2004) PCA in studying coordination and variability: a tutorial. Clinical Biomechanics 19:415-428. Decety J, Lamm C (2007) The Role of the Right Temporoparietal Junction in Social Interaction: How Low-Level Computational Processes Contribute to Meta-Cognition. The Neuroscientist 13:580-593. den Ouden HEM, Frith U, Frith C, Blakemore SJ (2005) Thinking about intentions. Neuroimage 28:787-796. Diana RA, Yonelinas AP, Ranganath C (2007) Imaging recollection and familiarity in the medial temporal lobe: a three-component model. Trends in Cognitive Sciences 11:379-386. Dickerson BC, Eichenbaum H (2009) The Episodic Memory System: Neurocircuitry and Disorders. Neuropsychopharmacology 35:86-104. Ditman T, Kuperberg GR (2005) A Source-Monitoring Account of Auditory Verbal Hallucinations in Patients with Schizophrenia. Harvard Review of Psychiatry 13:280 299. Dobbins IG, Foley H, Schacter DL, Wagner AD (2002) Executive Control during Episodic Retrieval: Multiple Prefrontal Processes Subserve Source Memory. 35:989-996.  112  Dobbins IG, Rice HJ, Wagner AD, Schacter DL (2003) Memory orientation and success: separable neurocognitive components underlying episodic recognition. Neuropsychologia 41:318-333. Dodson CS, Johnson MK (1993) Rate of false source attributions depends on how questions are asked. Am J Psychol 106:541-557. Donner TH, Kettermann A, Diesch E, Ostendorf F, Villringer A, Brandt SA (2002) Visual Feature and Conjunction Searches of Equal Difficulty Engage Only Partially Overlapping Frontoparietal Networks. Neuroimage 15:16-25. Duarte A, Henson RN, Knight RT, Emery T, Graham KS (2009) Orbito-frontal Cortex is Necessary for Temporal Context Memory. Journal of Cognitive Neuroscience 0:1-13. Eichenbaum H, Yonelinas AP, Ranganath C (2007) The medial temporal lobe and recognition memory. Annual Review of Neuroscience 30:123-152. Eliassen JC, Holland SK, Szaflarski JP (2008) Compensatory brain activation for recognition memory in patients with medication-resistant epilepsy. Epilepsy & Behavior 13:463-469. Esposito F, Formisano E, Seifritz E, Goebel R, Morrone R, Tedeschi G, Di Salle F (2002) Spatial independent component analysis of functional MRI time-series: to what extent do results depend on the algorithm used? Hum Brain Mapp 16:146-157. Finger S (1994) The Origins of Neuroscience: A History of Explorations into Brain Function. New York: Oxford University Press. Fox MD, Snyder AZ, Vincent JL, Corbetta M, Van Essen DC, Raichle ME (2005) The human brain is intrinsically organized into dynamic, anticorrelated functional networks. Proceedings of the National Academy of Sciences of the United States of America 102:9673-9678. Frackowiak RSJ, Friston KJ, Frith CD, Dolan RJ, Price CJ, Zeki S, Ashburner JT, Penny WD (2004a) Introduction. In: Human Brain Function (Second Edition), pp xv-xvi. Burlington: Academic Press. Frackowiak RSJ, Friston KJ, Frith CD, Dolan RJ, Price CJ, Zeki S, Ashburner JT, Penny WD, Richard SJF, Karl JF, Christopher DF, Raymond JD, Cathy JP, Semir Z, John TA, William DP (2004b) Introduction to Random Field Theory. In: Human Brain Function (Second Edition), pp 867-879. Burlington: Academic Press. Frackowiak RSJ, Friston KJ, Frith CD, Dolan RJ, Price CJ, Zeki S, Ashburner JT, Penny WD, Richard SJF, Karl JF, Christopher DF, Raymond JD, Cathy JP, Semir Z, John TA, William DP (2004c) Analysis of fMRI Time Series: Linear Time-Invariant Models, Event-Related fMRI, and Optimal Experimental Design. In: Human Brain Function (Second Edition), pp 793-822. Burlington: Academic Press. 113  Frackowiak RSJ, Friston KJ, Frith CD, Dolan RJ, Price CJ, Zeki S, Ashburner JT, Penny WD, Richard SJF, Karl JF, Christopher DF, Raymond JD, Cathy JP, Semir Z, John TA, William DP (2004d) Haemodynamic Modelling. In: Human Brain Function (Second Edition), pp 823-841. Burlington: Academic Press. Friston KJ, Frith CD, Liddle PF, Frackowiak RS (1991) Comparing functional (PET) images: the assessment of significant change. J Cereb Blood Flow Metab 11:690-699. Friston KJ, Fletcher P, Josephs O, Holmes A, Rugg MD, Turner R (1998) Event-Related fMRI: Characterizing Differential Responses. Neuroimage 7:30-40. Friston KJ, Holmes, A. P., Worsley, K. J., Poline, J.-P., Frith, C. D., Frackowiak, R. S. J. (1994a) Statistical parametric maps in functional imaging: A general linear approach. Human Brain Mapping 2:189-256. Friston KJ, Jezzard, P., & Turner, R. (1994b) Analysis of functional MRI time-series. Human Brain Mapping 1:153-171. Frith C (2005) The neural basis of hallucinations and delusions. Comptes Rendus Biologies 328:169-175. Frith C, Rees G, Friston K (1998) Psychosis and the experience of self. Brain systems underlying self-monitoring. Ann N Y Acad Sci 843:170-178. Frith CD, Done RJ (1989) Experiences of alien control in schizophrenia reflect a disorder in the central monitoring of action. Psychological Medicine 19:359-363. Frith CD, Blakemore S-J, Wolpert DM (2000) Explaining the symptoms of schizophrenia: Abnormalities in the awareness of action. Brain Research Reviews 31:357-363. Frith U, Frith CD (2003) Development and Neurophysiology of Mentalizing. Philosophical Transactions: Biological Sciences 358:459-473. Giannakopoulos X, Karhunen J, Oja E (1999) An experimental comparison of neural algorithms for independent component analysis and blind separation. Int J Neural Syst 9:99-114. Grabowski TJ, Damasio AR, Arthur WT, John CM (2000) Investigating Language with Functional Neuroimaging. In: Brain Mapping: The Systems, pp 425-461. San Diego: Academic Press. Gusnard DA, Raichle ME (2001) Searching for a baseline: Functional imaging and the resting human brain. Nature Reviews Neuroscience 2:685-694.  114  Handwerker DA, Ollinger JM, D'Esposito M (2004) Variation of BOLD hemodynamic responses across subjects and brain regions and their effects on statistical analyses. Neuroimage 21:1639-1651. Harvey PD (1985) Reality monitoring in mania and schizophrenia. The association of thought disorder and performance. J Nerv Ment Dis 173:67-73. Hasher L, Zacks RT (1979) Automatic and effortful processes in memory. Journal of Experimental Psychology: General 108:356-388. Hashtroudi S, Johnson MK, Chrosniak LD (1989) Aging and source monitoring. Psychology and Aging 4:106-112. Hekkanen ST, McEvoy C (2005) Source monitoring in eyewitness memory: implicit associations, suggestions, and episodic traces. Mem Cognit 33:759-769. Henkel LA, Johnson MK, De Leonardis DM (1998) Aging and source monitoring: cognitive processes and neuropsychological correlates. J Exp Psychol Gen 127:251-268. Henson R, Rugg MD, Friston KJ (2001) The choice of basis functions in event-related fMRI. Neuroimage 13:S149-S149. Hickok G, Poeppel D (2007) The cortical organization of speech processing. Nat Rev Neurosci 8:393-402. Hofer A, Weiss EM, Golaszewski SM, Siedentopf CM, Brinkhoff C, Kremser C, Felber S, Fleischhacker WW (2003) An fMRI Study of Episodic Encoding and Recognition of Words in Patients With Schizophrenia in Remission. Am J Psychiatry 160:911-918. Husain M, Nachev P (2007) Space and the parietal cortex. Trends in Cognitive Sciences 11:3036. Hyvarinen A, Oja E (2000) Independent component analysis: algorithms and applications. Neural Netw 13:411-430. Janowsky JS, Shimamura AP, Squire LR (1989) Source memory impairment in patients with frontal lobe lesions. Neuropsychologia 27:1043-1056. Jardri R, Pins D, Bubrovszky M, Despretz P, Pruvo J-P, Steinling M, Thomas P (2007) Self awareness and speech processing: An fMRI study. Neuroimage 35:1645-1653. Johnson MK, Raye CL (1981) Reality monitoring. Psychological Review 88:67-85. Johnson MK, Hashtroudi S, Lindsay DS (1993) Source monitoring. Psychol Bull 114:3-28.  115  Johnson MK, Kounios J, Reeder JA (1994) Time-course studies of reality monitoring and recognition. Journal of Experimental Psychology: Learning, Memory, and Cognition 20:1409-1419. Johnson MK, O'Connor M, Cantor J (1997) Confabulation, Memory Deficits, and Frontal Dysfunction. Brain and Cognition 34:189-206. Johnson MK, Foley MA, Suengas AG, Raye CL (1988) Phenomenal characteristics of memories for perceived and imagined autobiographical events. Journal of Experimental Psychology: General 117:371-376. Kiehl KA, Liddle PF, Hopfinger JB (2000) Error processing and the rostral anterior cingulate: an event-related fMRI study. Psychophysiology 37:216-223. Kircher TTJ, Brammer M, Bullmore E, Simmons A, Bartels M, David AS (2002) The neural correlates of intentional and incidental self processing. Neuropsychologia 40:683-692. Larøi F, Van der Linden M, Marczewski P (2004) The effects of emotional salience,cognitive effort and meta-cognitive beliefs on a reality monitoring task in hallucination-prone subjects. British Journal of Clinical Psychology 43:221-233. Lindsay DS (1990) Misleading suggestions can impair eyewitnesses' ability to remember event details. Journal of Experimental Psychology: Learning, Memory, and Cognition 16:10771083. Lindsay DS (2008) Source Monitoring. In: Cognitive Psychology of Memory (Roediger HL, ed), pp 325-348. Oxford: Elsevier. Lindsay DS, Johnson MK (1989) The eyewitness suggestibility effect and memory for source. Mem Cognit 17:349-358. Mather M, Knight M (2008) The emotional harbinger effect: Poor context memory for cues that previously predicted something arousing. Emotion 8:850-860. McGuire PK, Shah GMS (1993) Increased blood flow in Broca's area during auditory hallucinations in schizophrenia. Lancet 342:703. McIntosh AR, Lobaugh NJ (2004) Partial least squares analysis of neuroimaging data: applications and advances. Neuroimage 23 Suppl 1:S250-263. McIntosh AR, Chau WK, Protzner AB (2004) Spatiotemporal analysis of event-related fMRI data using partial least squares. Neuroimage 23:764-775. McIntosh AR, Bookstein FL, Haxby JV, Grady CL (1996) Spatial pattern analysis of functional brain images using partial least squares. Neuroimage 3:143-157.  116  McKeown MJ, Makeig S, Brown GG, Jung TP, Kindermann SS, Bell AJ, Sejnowski TJ (1998) Analysis of fMRI data by blind separation into independent spatial components. Hum Brain Mapp 6:160-188. McKiernan KA, Kaufman JN, Kucera-Thompson J, Binder JR (2003) A Parametric Manipulation of Factors Affecting Task-induced Deactivation in Functional Neuroimaging. Journal of Cognitive Neuroscience 15:394-408. McNally RJ, Kohlbeck PA (1993) Reality monitoring in obsessive-compulsive disorder. Behaviour Research and Therapy 31:249-253. Metzak P, Feredoes E, Takane Y, Wang L, Weinstein S, Cairo T, Ngan ETC, Woodward TS (2010) Constrained principal component analysis reveals functionally connected loaddependent networks involved in multiple stages of working memory. Human Brain Mapping:doi:10.1002/hbm.21072. Metzak PD, Riley JD, Wang L, Whitman JC, Ngan ETC, Woodward TS (2011) Decreased Efficiency of Task-Positive and Task-Negative Networks During Working Memory in Schizophrenia. Schizophrenia Bulletin:doi:10.1093/schbul/sbq1154. Miall RC, Wolpert DM (1996) Forward Models for Physiological Motor Control. Neural Networks 9:1265-1279. Mitchell KJ, Johnson MK (2009) Source monitoring 15 years later: what have we learned from fMRI about the neural mechanisms of source memory? Psychol Bull 135:638-677. Mitchell KJ, Johnson MK, Raye CL, Greene EJ (2004) Prefrontal cortex activity associated with source monitoring in a working memory task. J Cogn Neurosci 16:921-934. Morgane PJ, Galler JR, Mokler DJ (2005) A review of systems and networks of the limbic forebrain/limbic midbrain. Progress in Neurobiology 75:143-160. Nachev P, Husain M (2006) Disorders of Visual Attention and the Posterior Parietal Cortex. Cortex 42:766-773. Nolde SF, Johnson MK, Raye CL (1998) The role of prefrontal cortex during tests of episodic memory. Trends in Cognitive Sciences 2:399-406. Ogawa S, Tank DW, Menon R, Ellermann JM, Kim S-G, Merkle H, Ugurbil K (1992) Intrinsic Signal Changes Accompanying Sensory Stimulation: Functional Brain Mapping with Magnetic Resonance Imaging. Proceedings of the National Academy of Sciences of the United States of America 89:5951-5955. Penny WD, Stephan KE, Mechelli A, Friston KJ (2004) Modelling functional integration: a comparison of structural equation and dynamic causal models. Neuroimage 23:S264S274. 117  Petrides M, Pandya DN (2002) Comparative cytoarchitectonic analysis of the human and the macaque ventrolateral prefrontal cortex and corticocortical connection patterns in the monkey. The European Journal Of Neuroscience 16:291-310. Platel H, Baron J-C, Desgranges B, Bernard F, Eustache F (2003) Semantic and episodic memory of music are subserved by distinct neural networks. Neuroimage 20:244-256. Polyn SM, Natu VS, Cohen JD, Norman KA (2005) Category-Specific Cortical Activity Precedes Retrieval During Memory Search. Science 310:1963-1966. Poole DA, Lindsay DS (2002) Reducing Child Witnesses' False Reports of Misinformation from Parents. Journal of Experimental Child Psychology 81:117-140. Posner MI, Petersen SE, Fox PT, Raichle ME (1988) Localization of Cognitive Operations in the Human Brain. Science 240:1627-1631. Raichle ME, MacLeod AM, Snyder AZ, Powers WJ, Gusnard DA, Shulman GL (2001) A default mode of brain function. Proceedings of the National Academy of Sciences of the United States of America 98:676-682. Ross RS, Slotnick SD (2008) The Hippocampus is Preferentially Associated with Memory for Spatial Context. Journal of Cognitive Neuroscience 20:432-446. Sack AT (2009) Parietal cortex and spatial cognition. Behavioural Brain Research 202:153-161. Saxe R, Xiao DK, Kovacs G, Perrett DI, Kanwisher N (2004) A region of right posterior superior temporal sulcus responds to observed intentional actions. Neuropsychologia 42:14351446. Scarmeas N, Zarahn E, Anderson KE, Hilton J, Flynn J, Van Heertum RL, Sackeim HA, Stern Y (2003) Cognitive reserve modulates functional brain responses during memory tasks: a PET study in healthy young and elderly subjects. Neuroimage 19:1215-1227. Schmidt D, Krause BJ, Mottaghy FM, Halsband U, Herzog H, Tellmann L, Müller-Gärtner HW (2002) Brain systems engaged in encoding and retrieval of word-pair associates independent of their imagery content or presentation modalities. Neuropsychologia 40:457-470. Scoville WB, Milner B (1957) Loss Of Recent Memory After Bilateral Hippocampal Lesions. J Neurol Neurosurg Psychiatry 20:11-21. Serences JT (2004) A comparison of methods for characterizing the event-related BOLD timeseries in rapid fMRI. Neuroimage 21:1690-1700.  118  Simons JS, Owen AM, Fletcher PC, Burgess PW (2005a) Anterior prefrontal cortex and the recollection of contextual information. Neuropsychologia 43:1774-1783. Simons JS, Gilbert SJ, Owen AM, Fletcher PC, Burgess PW (2005b) Distinct Roles for Lateral and Medial Anterior Prefrontal Cortex in Contextual Recollection. J Neurophysiol 94:813-820. Simons JS, Davis SW, Gilbert SJ, Frith CD, Burgess PW (2006) Discriminating imagined from perceived information engages brain areas implicated in schizophrenia. Neuroimage 32:696-703. Smith AM, Lewis BK, Ruttimann UE, Ye FQ, Sinnwell TM, Yang Y, Duyn JH, Frank JA (1999) Investigation of Low Frequency Drift in fMRI Signal. Neuroimage 9:526-533. Smith SM (2001) Overview of fMRI analysis. In: Functional MRI: An Introduction to Methods (Jezzard P, Matthews, P. M., & Smith, S. M., ed), pp 216-227. Oxford: Oxford University Press. Squire LR, Zola-Morgan S (1991) The medial temporal lobe memory system. Science 253:13801386. Squire LR, Stark CEL, Clark RE (2004) The Medial Temporal Lobe. Annual Review of Neuroscience 27:279-306. Stephane M, Barton S, Boutros NN (2001) Auditory verbal hallucinations and dysfunction of the neural substrates of speech. Schizophrenia Research 50:61-78. Stone JV (2004) Independent component analysis : a tutorial introduction. Cambridge, Mass.: MIT Press. Tabachnick BG, Fidell LS (1989) Using multivariate statistics. New York: Harper & Row. Takane Y, Shibayama T (1991) Principal component analysis with external information on both subjects and variables. Psychometrika 56:97-120. Troyer AK, Winocur G, Craik FIM, Moscovitch M (1999) Source memory and divided attention: Reciprocal costs to primary and secondary tasks. Neuropsychology 13:467-474. Turner R, Howseman A, Rees GE, Josephs O, Friston K (1998) Functional magnetic resonance imaging of the human brain: data acquisition and analysis. Experimental Brain Research 123:5-12. Uncapher MR, Otten LJ, Rugg MD (2006) Episodic Encoding Is More than the Sum of Its Parts: An fMRI Investigation of Multifeatural Contextual Encoding. Neuron 52:547-556.  119  van Veen V, Carter CS (2002) The anterior cingulate as a conflict monitor: fMRI and ERP studies. Physiology & Behavior 77:477-482. Versmissen D, Janssen I, Johns L, McGuire PK, Drukker M, Campo JA, Myin-Germeys I, Van Os J, Krabbendam L (2007) Verbal self-monitoring in psychosis: a non-replication. Psychological Medicine 37:569-576. Vinogradov S, Luks TL, Schulman BJ, Simpson GV (2008) Deficit in a Neural Correlate of Reality Monitoring in Schizophrenia Patients. Cereb Cortex 18:2532-2539. Wagner AD, Shannon BJ, Kahn I, Buckner RL (2005) Parietal lobe contributions to episodic memory retrieval. Trends in Cognitive Sciences 9:445-453. Wheeler ME, Petersen SE, Buckner RL (2000) Memory's echo: Vivid remembering reactivates sensory-specific cortex. Proceedings of the National Academy of Sciences of the United States of America 97:11125-11129. Whitehead C, Marchant JL, Craik D, Frith CD (2009) Neural correlates of observing pretend play in which one object is represented as another. Social Cognitive and Affective Neuroscience 4:369-378. Woodward TS, Menon M, Whitman JC (2007) Source monitoring biases and auditory hallucinations. Cogn Neuropsychiatry 12:477-494. Woodward TS, Metzak PD, Meier B, Holroyd CB (2008) Anterior cingulate cortex signals the requirement to break inertia when switching tasks: A study of the bivalency effect. Neuroimage 40:1311-1318. Woodward TS, Cairo TA, Ruff CC, Takane Y, Hunter MA, Ngan ETC (2006) Functional connectivity reveals load dependent neural systems underlying encoding and maintenance in verbal working memory. Neuroscience 139:317-325. Worsley K (2001) Statistical analysis of activation images. In: Functional MRI: An Introduction to Methods (Jezzard P, Matthews, P. M., & Smith, S. M., ed), pp 251-270. Oxford: Oxford University Press. Worsley KJ, Evans AC, Marrett S, Neelin P (1992) A three-dimensional statistical analysis for CBF activation studies in human brain. J Cereb Blood Flow Metab 12:900-918. Zaragoza MS, Lane SM (1994) Source misattributions and the suggestibility of eyewitness memory. Journal of Experimental Psychology: Learning, Memory, and Cognition 20:934-945.  120  Appendices  Principal Component Analysis (Mathematics and Description) Principal component analysis refers to a multivariate technique that seeks to identify whether, for a given dataset, there exists a linear combination of variables that best expresses the original dataset in a more concise manner. (Given that singular value decomposition (SVD) and PCA are highly similar, the terns will be used interchangeably). PCA is used a variable reduction technique, where latent variables (or principal components) are generated that re-express a dataset using a smaller number of variables while conserving the bulk of the variability (Daffertshofer et al., 2004). The new latent variables identified via PCA are comprised of a linear sum of the weighted original variables, and are obtained by extracting eigenvectors that are uncorrelated with each other and account for a portion of the variability in the data set. The eigenvectors are extracted in descending order by the amount of variance they explain, and the number of eigenvectors in the data is equal to the rank of the matrix GC, which is the minimum row or column subscript of G. However, the majority of the total variance in the dataset can usually explained by the first few components thereby reducing the total number of variables in the dataset. In addition to the relation between component extraction order and variance explained (described in the PCA section in the main text), PCA also requires the following assumptions to be met in order to identify the underlying structure in large datasets (Daffertshofer et al., 2004): (1) the underlying structure in the dataset must conform to a linear structure, and (2) the principal components extracted must be orthogonal.  121  Given an m × n matrix X, PCA proceeds by creating the n × n covariance matrix X’X. From this covariance matrix, PCA finds the set of orthonormal eigenvectors and corresponding eigenvalues that solve the following equation: (13) where vi is the ith eigenvector, and λi is the ith eigenvalue. However when SVD rather than PCA is used, it is singular values and not eigenvalues that are extracted. The formula for obtaining the singular values from the eigenvalues is: (14)  where  is the ith scalar singular value, and  indicates congruence. Furthermore, the  component scores mentioned in the section on CPCA can be obtained via the following formula: (15)  Where  is the ith component score vector, and  obtained from this decomposition,  and  indicates congruence. The sets of vectors  have an orthonormal basis. Specifically, the  algebraic model for PCA incorporating all of these terms is: (16) When this formula is rearranged and placed into matrix notation, it takes the following form: (17) where  is the matrix notation for  .  122  Activated and Deactivated Brain Regions – Single Component Analyses In addition to the brain images included in this report, this section contains a listing of all the brain regions found to be activated or deactivated from each of the analysis presented herein, in the same order in which they appear in the text. Single subject, single condition (any stimulus) HRF design matrix analysis The functional brain images from these analyses can be found in Figure 3. SPM5 Both SPM5 analyses (with and without auto-regressor and high pass filter) revealed activations in the dorsal anterior cingulate (dAcc)/supplementary motor area (SMA), primary visual cortices, and superior temporal gyri, including the primary auditory cortices. Deactivations were detected in the medial prefrontal cortices, posterior cingulate/precuneus, cerebellum, lateral parietal cortices including angular gyri, and in the cerebellum including Crus II and the 8th and 9th lobule. CPCA As can be seen from examining Figure 3, the CPCA analysis revealed results that were identical to those obtained from the SPM5 analysis that did not use the AR(1) autoregressor or high pass filter. This analysis revealed activations in the dAcc/SMA, primary visual cortices, and superior temporal gyri, including the primary auditory cortices. Deactivations were detected in the medial prefrontal cortices, posterior cingulate/precuneus, cerebellum, lateral parietal cortices including angular gyri and in the cerebellum including Crus II and the 8th and 9th lobule.  123  ICA The component whose activity time course regressed most highly on the SPM5 HRF design matrix used in this analysis revealed activations in dAcc/SMA, bilateral activations in inferior frontal gyri (pars opercularis), primary motor cortices, and superior temporal gyri including the primary auditory cortices. Deactivations were found primarily bilaterally in the parahippocampal gyri, inferior temporal gyri including the fusiform area, retrosplenial cortices, and in the cerebellum including the 9th lobule. Single subject, single condition (any stimulus) FIR design matrix analysis The functional brain images from these analyses can be found in Figure 4. SPM5 The SPM5 analysis that used a high pass filter and an AR(1) autoregressor found activations in the left middle frontal gyrus and retrosplenial cortex, and deactivations in the mPFC,left inferior frontal gyrus (pars opercularis), right middle temporal cortex, and bilateral inferior occipital gyri. The SPM5 analysis that did not use a high pass filter and an AR(1) autoregressor found a more extensive pattern of activations and deactivations including bilateral activations in the superior frontal gyri, middle frontal gyri, supramarginal gyri, superior temporal gyri, as well as left hippocampal gyrus, and deactivations in the mPFC, inferior frontal gyri (pars opercularis and orbitalis) insula, superior parietal lobule, posterior cingulate/precuneus, retrosplenial cortex, superior, middle and inferior occipital gyri, and Crus II and lobule 6 of the cerebellum. CPCA This analysis revealed activations in the right inferor frontal gyrus (pars opercularis), right temporal pole, left hippocampal gyrus, retrosplenial cortex, and middle and inferior occipital  124  gyri. Deactivations were found bilaterally in the middle and superior frontal gyri, as well as in the posterior cingulate/precuneus. ICA The ICA component revealed bilateral activations in the superior frontal gyri, precentral gyrihippocampal gyri, middle temporal gyri, and middle occipital gyri, and Crus II and lobule 6 of the cerebellum as well as left middle and inferior frontal gyri (pars orbitalis), right supramarginal gyrus and right superior occipital gyrus. Deactivations were found in caudal mPFC, temporal poles, and posterior cingulate/precuneus. PLS The PLS analysis involved used a lag-window of 8 scans, and analyzed the pattern of brain activity associated with the presentation of any stimulus during this time frame. It should be noted that design matrices using a synthetic HRF are not usable in the current implementation of the PLS software, therefore this analysis utilized the FIR-style model implemented in the PLS software package. Also, according to the User‟s Guide, PLS analyses are unstable when groups are made up of less than 3 subjects. These caveats notwithstanding, the PLS analysis revealed significant activations bilaterally in the orbito-frontal cortices, in the posterior cingulate, visual cortex, and in the right superior parietal lobe. Significant deactivations were found in the medial temporal lobe, cerebellum, and brainstem. 21 controls, single condition (any stimulus) HRF design matrix analysis The functional brain images from these analyses can be found in Figure 5. SPM5 This SPM5 analysis revealed activations in the dAcc/SMA, bilaterally in the inferior frontal gyri (pars triangularis), insula, retrosplenial cortex, middle and inferior occipital lobes,  125  and Lobule 6 of the cerebellum, as well in the left middle frontal gyrus, left pre- and post-central gyri, right caudate, left inferior parietal lobule, left thalamus, right calcarine sulcus, and left fusiform gyrus. Deactivations were found bilaterally in the medial prefrontal cortices, superior and middle temporal gyri, supramarginal gyri, and precuneus/posterior cingulate, as well as in the right superior frontal gyrus and right lingual gyrus. CPCA The CPCA analysis revealed bilateral activations in the dAcc/SMA, inferior frontal gyri (pars triangularis), insula, retrosplenial cortex, middle and inferior occipital lobes, and Lobule 6 of the cerebellum, as well in the left middle frontal gyrus, left pre- and post-central gyri, left inferior parietal lobule, and left fusiform gyrus. Deactivations were found bilaterally in the medial prefrontal cortices, superior and middle temporal gyri, supramarginal gyri, and precuneus/posterior cingulate, as well as in the right superior frontal gyrus and right lingual gyrus. ICA The ICA analysis revealed bilateral activations in dAcc/SMA, post-central gyrus and inferior parietal lobule including supramarginal gyri, as well as in the left superior frontal gyrus, left pre-motor area, and right lobule 6 of the cerebellum. 21 patients, single condition (any stimulus) HRF design matrix analysis The functional brain images from these analyses can be found in Figure 6. SPM5 This analysis revealed significant bilateral activations in the dAcc/SMA, insula, inferior frontal gyri (pars triangularis), pre-central gyri, pallidum, angular gyri, hippocampi/posterior  126  parahippocampal gyri, middle and inferior occipital gyri, calcarine sulcus, and lobule 4, 5, and 6 and Crus 1 and 2 of the cerebellum, as well as left thalamus and left post-central gyri. Significant deactivations were found in medial prefrontal cortices, superior and middle frontal gyri, middle cingulum, posterior cingulate/precuneus, retrosplenial cortex, superior, middle, and inferior temporal gyri, as well as right orbitofrontal gyrus and right Heschl‟s gyrus. CPCA This CPCA analysis revealed bilateral activations in the dAcc/SMA, superior parietal lobule, calcarine sulcus, middle and inferior occipital gyri, lingual gyri, and lobule 6 and Crus 1 of the cerebellum, as well as in left angular gyrus, and left pre- and post-central gyri. ICA The ICA component that regressed most strongly on the SPM5 HRF design matrices for the 21 patients revealed significant bilateral activations in the dAcc/SMA, inferior frontal gyri (pars triangularis), right insula, and right Crus 1 of the cerebellum. 21 controls, single condition (any stimulus) FIR design matrix analysis The functional brain images from these analyses can be found in Figure 7. SPM5 This SPM5 analysis revealed significant activations in the dAcc/SMA, left middle frontal gyrus extending into left orbitofrontal gyrus, left inferior frontal gyrus (pars triangularis), left angular gyrus, and right Crus 1 of the cerebellum. Significant deactivations were found bilaterally in the medial prefrontal cortices, middle frontal gyri, superior and middle temporal gyri, posterior cingulate/precuneus, fusiform gyri, calcarine sulci, lingual gyri, and occipital cortices, as well as right inferior frontal gyrus (pars triangularis).  127  CPCA This CPCA analysis revealed bilateral activations in dAcc/SMA, fusiform gyri, lingual gyri, calcarine sulci, middle and inferior occipital gyri, and lobule 6 of the cerebellum, as well as left pre- and post-central gyri extending into the superior parietal lobule. ICA The ICA component that regressed most strongly on the SPM5 FIR design matrices for the 21 controls revealed bilateral activations in the superior, middle and inferior occipital gyri, as well as lingual gyri, fusiform gyri, calcarine sulci, and lobule 6 and Crus 1 of the cerebellum. PLS This PLS analysis revealed significant bilateral activations in the dAcc/SMA, inferior frontal gyri (pars triangularis), superior parietal lobule, middle and inferior occipital gyri, fusiform gyri, and lobule 6 of the cerebellum, as well as left pre- and post-central gyri, left insula, left caudate, left hippocampus and parahippocampal gyrus, left thalamus, and right angular gyrus. Significant bilateral deactivations were found in the superior and middle temporal gyri, precuneus and precuneus, as well as right superior frontal gyrus, and right Crus 1 of the cerebellum. 21 patients, single condition (any stimulus) FIR design matrix analysis The functional brain images from these analyses can be found in Figure 8. SPM5 The SPM5 FIR analysis revealed significant activations in the dAcc/SMA, left pre- and post-central gyri, right thalamus, right parahippocampal gyrus, right fusiform gyrus, and Crus 1 and 2 and lobule 6 of the cerebellum. Significant bilateral deactivations were found in the medial  128  prefrontal cortices, middle frontal gyri, middle and posterior cingulate/precuneus, middle and superior temporal gyri, lingual gyri, and occipital gyri, as well as right inferior frontal gyrus (pars triangularis and orbitalis), right superior frontal gyrus, right pre- and post-central gyri, right supramarginal gyrus, and right calcarine sulcus. CPCA This CPCA analysis revealed bilateral activations in the dAcc/SMA, superior parietal lobules, angular gyri, calcarine sulci, fusiform gyri, lingual gyri, occipital gyri, and Crus 1 and lobule 6 of the cerebellum, as well left pre- and post-central gyri. ICA The ICA component that regressed most strongly on the SPM5 FIR design matrices for the 21 patients revealed bilateral activations in the superior, middle and inferior occipital gyri, as well as lingual gyri, fusiform gyri, calcarine sulci, and lobule 6 and crus 1 of the cerebellum. PLS This PLS analysis revealed significant bilateral activations in the dAcc/SMA, insula, putamen, thalamus, superior parietal lobules, fusiform gyri, lingual gyri, calcarine sulci, superior, middle and inferior occipital gyri, and Crus 1 and lobule 6 of the cerebellum, as well as left inferior frontal gyrus (pars triangularis), and left pre- and post-central gyri. Significant bilateral deactivations were found in the medial prefrontal cortices, middle frontal gyri, superior and middle temporal gyri, posterior cingulate/precuneus, cuneus, and Crus 2 of the cerebellum, as well as right inferior frontal gyrus (pars orbitalis), right post-central gyrus, right supramarginal gyrus, and right angular gyrus.  129  21 controls, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 9. SPM5 This SPM5 analysis modeled each of the 4 conditions (association, other, self, or reading) and 2 outcomes (recalled or non-recalled) separately, however the contrast employed here was all 8 stimuli versus baseline. This analysis revealed significant bilateral activations in dAcc/SMA, insula, inferior frontal gyrus (pars triangularis), pallidum, angular gyri, calcarine sulci, fusiform gyri, middle and inferior occipital gyri, and crus 1 and lobule 6 of the cerebellum, as well as left pre- and post-central gyri, and superior parietal lobule. Significant bilateral deactivations were in the medial prefrontal cortices, superior and middle temporal gyri, supramarginal gyri, and precuneus/posterior cingulate, as well as in the right superior frontal gyrus. CPCA The CPCA analysis revealed bilateral deactivations in the medial prefrontal cortices, middle frontal gyri, thalamus, hippocampi, superior and middle temporal gyri, posterior cingulate/precuneus, retrosplenial cortex, calcarine sulci, as well as right pre-and post-central gyri. The graph below the functional brain image from this analysis depicts the mean predictor weights for each condition. The highest values in the internal recalled condition indicates that these regions are least deactivated in this condition whereas the lowest values in the read nonrecalled condition indicate that these regions are most deactivated in this condition. ICA The ICA component that regressed most highly on the 8condition SPM5 HRF design matrix revealed bilateral activations in dAcc/SMA, and insula, extending into the inferior frontal 130  gyri (pars orbitalis). The regression values for each of the conditions has been plotted on a graph undersneath the functional brain image. The highest R-square values were in the internal nonrecalled and associate non-recalled conditions. This indicates that activity in this network fit the SPM5 HRF model most closely in these conditions. 21 patients, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 10. SPM5 This SPM5 analysis modeled each of the 4 conditions (association, other, self, or reading) and 2 outcomes (recalled or non-recalled) separately, however the contrast employed here was all 8 stimuli versus baseline. This analysis revealed significant bilateral activations in dAcc/SMA, insula, pallidum, fusiform gyri, middle and inferior occipital gyri, and crus 1 and 2 and lobule 6 of the cerebellum, as well as left inferior frontal gyrus (pars triangularis), left preand post-central gyri, left thalamus, and left superior parietal lobule. Significant bilateral deactivations were found in medial prefrontal cortices, middle/posterior cingulate, superior, middle, and inferior temporal gyri, as well as right middle and inferior frontal gyri (pars triangularis and opercularis), right Heschl‟s gyrus, and right supramarginal gyrus. CPCA This CPCA analysis revealed bilateral activations in dAcc/SMA, inferior frontal gyri (pars opercularis), middle cingulate, posterior cingulate/precuneus, retrosplenial cortex, superior parietal lobules, cuneus, superior and middle occipital gyri, and crus 1 and 2 and lobule 6 of the cerebellum, as well as left calcarine sulcus and left lingual gyrus. The predictor weights reveal that this component was most deactivated in the other recalled and read recalled conditions and most activated in the self recalled condition. 131  ICA The ICA component that regressed most highly on the SPM5 HRF design matrix displayed bilateral activations in the dAcc/SMA, middle and inferior frontal gyri (pars triangularis and opercularis), and insula, as well as left inferior frontal gyrus (pars orbitalis), left pre-central gyrus, left superior parietal lobule, left angular gyrus, and left supramarginal gyrus. The R-square values reveal that the activity of this component matches most closely the SPM5 design matrix canonical HRF response pattern for the read recall and associate recall conditions. 21 controls, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 11. SPM5 This SPM analysis revealed significant bilateral activations in dAcc/SMA, as well as left middle frontal gyrus, left orbitofrontal gyrus, left inferior frontal gyrus (pars triangularis), and left angular gyrus. Significant bilateral deactivations were found in medial prefrontal cortices, middle and posterior cingulate/precuneus, superior and middle tempotal gyri, parahippocampal gyri, calcarine sulci, fusiform gyri, superior and middle occipital gyri, as well as right middle frontal gyrus. CPCA This CPCA analysis revealed bilateral activations in dAcc/SMA, insula, middle cingulate, precuneus, superior parietal lobules, superior temporal gyri, lingual gyri, calcarine sulci, cuneus, superior and middle occipital gyri, and crus 1,2 and lobule 6 of the cerebellum, as well as left pre- and post-central gyri. The predictor weights indicate that this network is less active during the other recalled condition, and that the hemodynamic response associated with the self recalled condition is quicker than that of the remaining recalled conditions. 132  ICA The component that regressed most strongly onto the SPM5 FIR design matrix revealed bilateral activations in the superior, middle and inferior occipital gyri, fusiform gyri, lingual gyri, calcarine sulci, as well as crus 1 and lobule 6 of the cerebellum. The R-square values plotted by condition and time point indicate that the visual cortex activity matched the SPM5 FIR design matrix more closely in the recalled than non-recalled conditions, and that the read recalled and other recalled conditions peaked earlier than the association recalled and self recalled conditions. PLS This PLS analysis revealed bilateral deactivations in the medial prefrontal cortices, middle frontal gyri, superior and middle temporal gyri, superior parietal lobules, angular gyri, supramarginal gyri, posterior cingulate/precuneus, cuneus, crus 1 and 2 and lobule 6 of the cerebellum. The design scores indicate that this network is least deactivated active during association recalled, association non-recalled, and self recalled conditions, and most deactivated during the other non-recalled condition. 21 patients, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 12. SPM5 This SPM5 analysis revealed significant bilateral activations in dAcc/SMA, as well as left pre- and post-central gyri, and right crus 2 of the cerebellum. Significant bilateral deactivations were found in medial prefrontal cortices, middle cingulate, temporal lobes, posterior cingulate/precuneus, superior parietal lobule, as well as right middle and inferior frontal gyri (pars triangularis), left caudate, right superior frontal gyrus, right post-central gyrus, right  133  angular gyrus, right fusiform gyrus, right calcarine sulcus, right superior and middle occipital gyri, and left crus 2 of the cerebellum. CPCA This CPCA analysis revealed bilateral deactivations in pre-central gyrus, posterior cingulate/precuneus, superior parietal lobules, angular gyri, calcarine sulci, lingual gyri, cuneus, superior, middle and inferior occipital gyri, and crus 1, 2 and lobule 6 of the cerebellum. The predictor weights indicate that this network is more deactivated in the self recalled condition than in the other recalled condition, and that there is little difference between recalled and nonrecalled in this network. ICA The ICA component that regressed most highly on the 21 patients SPM5 FIR model revealed bilateral activations in the superior, middle and inferior occipital gyri, fusiform gyri, lingual gyri, calcarine sulci, as well as crus 1 and lobule 6 of the cerebellum. The R-square values plotted by condition and time point indicate that the visual cortex activity matched the SPM5 FIR design matrix more closely in the recalled than non-recalled conditions, and that the read recalled and other recalled conditions peaked earlier than the association recalled and self recalled conditions, which mirrors the results from the controls in the previous analysis (see Figure 11). PLS The PLS analysis revealed significant bilateral activations in medial prefrontal cortices, middle frontal gyri, insula, thalamus, temporal cortices, superior parietal lobules, middle and superior temporal gyri, supramarginal gyri, angular gyri, fusiform gyri, middle and inferior  134  occipital gyri, and crus 1, 2 and lobule 6 of the cerebellum as well as right hippocampus and brain stem. Significant bilateral deactivations were found in middle and posterior cingulate/precuneus, and retrosplenial cortex. The design scores indicate that the regions with significant activations were most active in the read non-recalled condition whereas the regions with significant deactivations were most deactivated in the self recalled condition. 21 controls and 21 patients, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 13. SPM5 The SPM5 analysis revealed significant bilateral activations in the dAcc/SMA, insula, pallidum, angular gyri, middle and inferior occipital gyri, fusiform gyri, and crus 1, 2 and lobule 6 of the cerebellum, as well as left inferior frontal gyrus (pars triangularis), left pre- and post central gyri, left thalamus, and left superior occipital gyrus. Significant bilateral deactivations were found in the medial prefrontal cortices, supramarginal gyri, superior and middle temporal gyri, posterior cingulate/precuneus, as well as right middle frontal gyrus and right inferior frontal gyrus (pars orbitalis). CPCA The CPCA component that was extracted from this analysis displayed bilateral deactivations in medial prefrontal cortices, middle cingulate, thalamus, superior and middle temporal gyri, posterior cingulate/precuneus, retrosplenial cortex, cuneus, calcarine sulci, and crus 2 of the cerebellum, as well as right middle frontal gyrus, and right pre-central gyrus. The predictor weights suggest that these brain regions were least deactivated during the internal recalled condition in both controls and patients. Also, for the patients only, it appears as though  135  the activation or deactivation in each condition is greater when the encoding context was recalled as opposed to non-recalled. ICA The ICA component that regressed most strongly onto the SPM5 HRF design matrices of the patients and controls revealed activations in the superior, middle and inferior occipital gyri, fusiform gyri, lingual gyri, calcarine sulci, as well as crus 1 and lobule 6 of the cerebellum. The R-square values plotted by condition indicate that the activity time course of the occipital cortex component matched the SPM5 HRF design matrix for both the controls and patients most closely when items encoded during the read condition were correctly recalled. The condition in which the R-square value was lowest was in the internal recalled condition for both the controls and the patients. 21 controls and 21 patients, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 14. SPM The SPM5 analysis revealed significant bilateral activations in the dAcc/SMA, insula, and calcarine sulci, as well as left middle and inferior frontal gyrus (pars triangularis and orbitalis), left superior parietal lobule, left supramarginal gyrus, left angular gyrus, and right crus 1 and crus 2 of the cerebellum. Significant bilateral deactivations were found in medial prefrontal cortices, superior, middle, and inferior temporal gyri, fusiform gyri, and posterior cingulate/precuneus, as well as right middle frontal gyrus, right superior frontal gyrus, right preand post-central gyri, and right middle and inferior occipital gyri.  136  CPCA The CPCA analysis revealed bilateral activations in dAcc/SMA, middle cingulate, preand post-central gyri, superior parietal lobules, precuneus/posterior cingulate, cuneus, calcarine sulci, lingual gyri, occipital cortices, and crus 1,2 and lobule 6 of the cerebellum, as well as right middle frontal gyrus and left superior temporal gyrus. The predictor weights associated with this component indicate that successfully recollection of the stimulus context led to increased activity relative to unsuccessfully recalled words. Also, in both the patients and controls, the internal recalled words resulted in the highest mean predictor weight peak, whereas the external recalled words resulted in the lowest mean predictor weight peak. Additionally, in the controls, the read recalled activation peak was delayed related to the peaks of the other recalled conditions. This delayed peak is not found in the patients. ICA The ICA component that regressed most strongly onto the SPM5 FIR design matrices of the patients and controls revealed activations in the superior, middle and inferior occipital gyri, fusiform gyri, lingual gyri, calcarine sulci, as well as crus 1 and lobule 6 of the cerebellum. The R-square values plotted by condition indicate that the activity time course of the occipital cortex component matched the SPM5 FIR design matrix for both the controls and patients more closely in the recalled than non-recalled conditions. For both the patients and controls the highest Rsquare values are found in the read recalled condition. Additionally, for both the patients and controls, it appears that the read recalled and external recalled conditions peak earlier than the associate recalled and internal recalled conditions.  137  PLS The PLS analysis identified a component comprised of significant bilateral deactivations in medial prefrontal cortices, superior middle and inferior frontal gyri (pars triangularis), superior and middle temporal gyri, parahippocampal gyri, pre- and post-central gyri, posterior cingulate/precuneus, superior parietal lobules, angular gyri, supramarginal gyri, middle and inferior occipital gyri, lingual gyri, and crus 1 and 2 of the cerebellum. The design scores indicate that this component is most deactivated in the internal recalled condition in both patients and controls, and in the associate non-recalled and recalled condition in controls only. This component is least deactivated in the external miss condition in controls and the associate nonrecalled condition in patients. Activated and Deactivated Brain Regions – Multiple Component Analyses This section lists the brain regions where activations and/or deactivations were found in the analyses where multiple components were extracted. These analyses will be examined separately according to the sample and design matrix used. 21 controls, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 15. CPCA The first CPCA component extracted from the analysis of this sample and design matrix displayed bilateral deactivations in medial prefrontal cortices, ventral anterior cingulate, middle cingulate, thalamus, superior and middle temporal gyri, supramarginal gyri, posterior cingulate/precuneus, retrosplenial cortex, cuneus, calcarine sulci,lingual gyri, and crus 2 of the cerebellum, as well as right inferior frontal gyrus (pars opercularis), and right pre-central gyrus. The second CPCA component extracted displayed bilateral activations in dAcc/SMA, angular 138  gyri, supramarginal gyri, calcarine sulci, lingual gyri, occipital cortices, and crus 1,2 and lobule 6 of the cerebellum, as well as left inferior frontal gyrus (pars triangularis), left middle frontal gyrus, left precentral gyrus, and left inferior postcentral gyrus. The third CPCA component displayed bilateral activations in dAcc/SMA, inferior frontal gyri (pars triangularis), as well as right insula, pars opercularis of the left inferior gyrus, and right fusiform gyrus. ICA The ICA component that was found to have the highest R-square value was characterized by bilateral activations in the dAcc/SMA, insula, inferior frontal gyri (pars orbitalis), and temporal poles. The ICA component that was found to have the second highest R-square value was characterized by activations in the right temporo-parietal junction, a region including the posterior portion of the middle and superior temporal gyris, as well as the inferior portion of the angular and supramarginal gyri. The ICA component that was found to have the third highest Rsquare value was characterized by activations in the superior, middle and inferior occipital gyri, fusiform gyri, lingual gyri, calcarine sulci, as well as crus 1 and lobule 6 of the cerebellum. The ICA component that was found to have the fourth highest R-square value displayed bilateral activations in the middle cingulate gyrus, superior frontal gyri, pre-and post-cenral gyri, angular gyri, supramarginal gyri, as well as right lobule 6 of the cerebellum. 21 controls, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 16. CPCA The component from this CPCA analysis that accounted for the most variance was characterized by bilateral activations in dAcc/SMA, angular gyri, pre- and post-central gyri, superior parietal lobule, precuneus, cuneus, calcarine sulci, lingual gyri, occipital cortices, and 139  crus 1,2 and lobule 6 of the cerebellum, as well as left inferior frontal gyrus (pars triangularis), left middle frontal gyrus, left precentral gyrus, and left inferior postcentral gyrus. The CPCA component that accounted for the 2nd largest amount of variance was characterized by bilateral activations in the dAcc/SMA, angular gyri, superior, middle, and inferior occipital cortices, Crus 1 and lobule 6 of the cerebellum, as well as left inferior frontal gyrus (pars triangularis), pre- and post-central gyri, and left supramarginal gyrus. The CPCA component that accounted for the 3rd most variance was characterized by bilateral activations in the dAcc/SMA, middle frontal gyri, inferior frontal gyri (pars triangularis), insula, as well as the left pars orbitalis region of the inferior frontal gyrus and left angular gyrus. The CPCA component that accounted for the 4th most variance was characterized by bilateral deactivations in medial prefrontal cortices, posterior cingulate/precuneus, and retrosplenial cortices, as well as right angular gyrus. ICA The component that was found to have the highest R-square value was characterized by bilateral activations in the superior, middle and inferior occipital gyri, fusiform gyri, lingual gyri, calcarine sulci, as well as crus 1 and lobule 6 of the cerebellum. The component that was found to have the second highest R-square value was characterized by bilateral activations in the dAcc/SMA, insula, inferior frontal gyri (pars orbitalis), and temporal poles. PLS The component that accounted for the most variance in the PLS analysis was characterized by bilateral deactivations in the medial prefrontal cortices, middle frontal gyri, inferior frontal gyri (pars triangularis & orbitalis), pre- and post-central gyri, angular gyri, supramarginal gri, superior, middle, and inferior temporal gyri, precuneus, cuneus, superior and middle occipital gyri, and Crus 2 of the cerebellum. The component that accounted for the 140  second most variance in the PLS analysis was characterized by bilateral activations in the dAcc/SMA, inferior frontal gyri (pars triangularis and opercularis), middle frontal gyri, insula, pre-central gyri, thalamus, supramarginal gyri, and lingual gyri, as well as right post-central gyrus, left middle occipital gyrus, and right lobule 9 of the cerebellum. 21 patients, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 17. CPCA The CPCA component that accounted for the most variance in this analysis revealed a network of regions whose behaviour depended upon the experimental condition. The regions in this network included the medial prefrontal cortices, middle cingulate gyri, middle frontal gyri, thalamus, post-central gyri, supramarginal gyri, superior temporal gyri, precuneus, cuneus, lingual gyri, calcarine sulci and Crus 1 and 2 of the cerebellum. The CPCA component that accounted for the second most variance in this analysis displayed bilateral activations in dAcc/SMA, inferior frontal gyri (pars triangularis), middle and inferior occipital gyri, and Crus 1 and lobule 6 of the cerebellum, as well as left the pars operclaris region of the left inferior frontal gyrus, left pre-central gyrus, and left supramarginal and angular gyri. The CPCA component that accounted for the third most variance in this analysis displayed bilateral deactivations in the medial prefrontal cortices, superior, middle, and inferior temporal gyri, supramarginal gyri, and middle occipital gyri. ICA The component that was found to have the highest R-square value was characterized by bilateral activations in the dAcc/SMA, inferior frontal gyri (pars opercularis and triangularis),  141  middle frontal gyri, as well as left superior temporal pole, right insula, and left supramarginal gyrus. The component that was found to have the second highest R-square value was characterized by bilateral deactivations in the medial prefrontal cortices. The component that was found to have the third highest R-square value was characterized by bilateral deactivations in the retrosplenial cortices. 21 patients, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 18. CPCA The CPCA component that accounted for the most variance in this analysis revealed a network of regions that were predominantly deactivated during the experiment. These regions included the dAcc/SMA the middle cingulate, pre-central gyri, superior parietal lobule, precuneus, cuneus, superior and middle occipital gyri, calcarine sulci, lingual gyri, and Crus 1 and 2 and lobule 6 of the cerebellum. The CPCA component that accounted for the second most variance in this analysis revealed bilateral activations in dAcc/SMA, inferior frontal gyri (pars triangularis and opercularis), middle frontal gyri, angular gyri, superior, middle, and inferior occipital gyri, fusiform gyri, calcarine sulci, lingual gyri, and Crus 1 and 2 and lobule 6 of the cerebellum, as well as left pre-central gyrus, left supramarginal gyrus, and left middle temporal gyrus. The CPCA component that accounted for the third most variance in this analysis revealed bilateral deactivations in medial prefrontal cortices, middle and superior frontal gyri, anterior temporal poles, and middle temporal gyri, as well as right inferior frontal gyrus (pars triangularis and orbitalis), right pre-central gyrus, right superior temporal gyrus and right supramarginal gyrus. The CPCA component that accounted for the fourth most variance in this analysis revealed bilateral activations in dAcc/SMA, middle and inferior (pars triangularis, opercularis  142  and orbitalis) frontal gyri, insula, as well as left superior frontal gyrus and and left pre-central gyrus. ICA The ICA component that was found to have the highest R-square value revealed bilateral activations in the superior, middle and inferior occipital gyri, fusiform gyri, lingual gyri, calcarine sulci, as well as crus 1 and lobule 6 of the cerebellum. The ICA component that was found to have the second highest R-square value was characterized by activation in the left insula, and left inferior frontal gyrus (pars orbitalis and triangularis). The ICA component that was found to have the third highest R-square value was characterized by bilateral activations in medial prefrontal cortices, superior frontal gyri, and middle frontal gyri. The ICA component that was found to have the fourth highest R-square value revealed bilateral activations in dAcc/SMA. PLS The PLS component that accounted for the most variance using this sample and design matrix revealed a network of regions whose behaviour (activity or deactivity) depended on the experimental condition. This network included the following bilateral regions: middle frontal gyri, inferior frontal gyri (pars triangularis), pre-central gyri, inferior temporal gyri, middle cingulate gyri, posterior cingulate/precuneus, lingual gyri, and Crus 1 and lobule 6 of the cerebellum; as well as the following lateralized regions: right superior frontal gyrus, left insula, left middle temporal gyrus, left angular gyrus, and left supramarginal gyrus. The PLS component that accounted for the most variance using this sample and design matrix revealed a network of regions whose behaviour (activity or deactivity) depended on the experimental condition. This network included the following bilateral regions: medial prefrontal cortices, inferior frontal gyri 143  (pars opercularis, triangularis, and orbitalis), pre- and post-central gyri, inferior, middle, and superior temporal gyri, posterior cingulate, angular gyri, and Crus 1 and 2 and lobule 6 of the cerebellum; as well as the following lateralized regions: right superior frontal gyrus, right supramarginal gyrus, and left inferior occipital gyrus. 21 controls and 21 patients, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 19. CPCA The CPCA component that accounted for the most variance in this analysis revealed a network of regions that were deactivated in all the experimental conditions except the internal recalled condition. These regions included bilateral medial prefrontal cortices, middle frontal gyri, thalamus, superior and middle temporal gyri, posterior cingulate/precuneus, calcarine sulci, lingual gyri, as well as right post-central gyrus, right supramarginal gyrus, left fusiform gyrus, and left Crus 1 and 2 and lobule 6 of the cerebellum. The CPCA component that explained the second most variance was characterized by bilateral activations in the dAcc/SMA, supramarginal gyri, angular gyri, middle and inferior occipital gyri, fusiform gyri, and Crus 1 of the cerebellum, as well as left inferior frontal gyrus (pars triangularis), left middle frontal gyrus, left pre- and post-central gyri. The CPCA component that explained the third most variance was characterized by bilateral deactivations in the medial prefrontal cortices, superior, middle, and inferior temporal gyri, posterior cingulate/precuneus, and supramarginal gyri, as well as right middle and superior frontal gyri, and right post-central gyrus. The CPCA component that explained the fourth most variance was a network of regions that activated or deactivated on the basis of the experimental condition. This network included bilateral dAcc/SMA, middle frontal gyri, inferior  144  frontal gyri (pars orbitalis and triangularis), pre-central gyri, thalamus, superior temporal poles, as well as right pallidum and left parahippocampal gyrus. ICA The ICA component that was found to have the highest R-square value revealed bilateral activations in the superior, middle and inferior occipital gyri, fusiform gyri, lingual gyri, calcarine sulci, as well as Crus 1 and lobule 6 of the cerebellum. The ICA component that was found to have the second highest R-square value revealed bilateral activations in middle cingulate, posterior superior frontal gyrus, post-central gyrus, supramarginal gyri, as well as left pre-central gyrus, left superior parietal lobule, and right lobule 6 of the cerebellum. The ICA component that was found to have the third highest R-square value revealed bilateral activations in inferior frontal gyri (pars triangularis and operculus) which extended into pre-central gyri. The ICA component that was found to have the fourth highest R-square value revealed activation in the right temporo-parietal junction; specifically, the right posterior superior temporal gyrus, right supramarginal gyrus and right angular gyrus. 21 controls and 21 patients, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 20. CPCA The CPCA component that explained the most variance was characterized by a network of regions that showed a variable pattern of activity or deactivity over the course of a trial, depending on which experimental condition is being examined. This network of regions included the SMA, middle cingulate, posterior cingulate/precuneus, pre- and post-central gyri, supramarginal gyri, superior parietal lobule, cuneus, middle occipital gyri, calcarine sulci, lingual gyri, fusiform gyri, and Crus 1 and 2 of the cerebellum, as well as left superior temporal gyrus. 145  The CPCA component accounting for the second most variance was characterized by bilateral activations in dAcc/SMA, angular gyrus, superior, middle and inferior occipital gyri, fusiform gyri, lingual gyri, calcarine sulci, and Crus 1 and lobule 6 of the cerebellum, as well as left inferior frontal gyrus (pars triangularis), left pre- and post-central gyrus, left supramarginal gyrus, and left superior parietal lobule. The CPCA component accounting for the third most variance was characterized by bilateral activations in dAcc/SMA, middle frontal gyri, inferior frontal gyri (pars opercularis, triangularis, and orbitalis), insula, as well as left pre-central gyrus and left angular gyrus. The CPCA component accounting for the third most variance was characterized by bilateral deactivations in medial prefrontal cortices, superior and middle frontal gyri, middle cingulate, posterior cingulate/precuneus, as well as right superior and middle temporal gyri and right angular gyrus. ICA The ICA component that was found to have the highest R-square revealed bilateral activations in the superior, middle and inferior occipital gyri, fusiform gyri, lingual gyri, calcarine sulci, as well as Crus 1 and lobule 6 of the cerebellum. The ICA component that was found to have the second highest R-square value revealed bilateral activations in middle cingulate, posterior superior frontal gyrus, post-central gyrus, supramarginal gyri, as well as left pre-central gyrus, left superior parietal lobule, and right lobule 6 of the cerebellum. The ICA component that was found to have the third highest R-square value revealed bilateral activations in inferior frontal gyri (pars triangularis and operculus) which extended into pre-central gyri. The ICA component that was found to have the fourth highest R-square value was characterized by bilateral deactivation in the medial prefrontal cortices, which extended into the orbital section of the middle frontal gyri and pars orbitalis of the inferior frontal gyri. 146  PLS The PLS component that accounted for the most variance using this sample and design matrix revealed bilateral deactivations in medial prefrontal cortices, middle frontal gyri, inferior frontal gyri (pars triangularis), caudate, pre-central gyri, superior, middle, and inferior temporal gyri, superior parietal lobule, supramarginal gyri, angular gyri, middle and inferior occipital gyri and Crus 1 and 2, and lobule 6 of the cerebellum. The PLS component that accounted for the second most variance using this sample and design matrix revealed bilateral activations in dAcc/SMA, insula, inferior frontal gyri (pars triangularis and operculus), pre- and post-central gyri, and lingual gyri, as well as left inferior frontal gyrus (pars orbitalis). Statistics for the Single Component Analyses 21 controls, single condition (any stimulus) HRF design matrix analysis The functional brain images from these analyses can be found in Figure 5. For the CPCA analysis, the mean of the predictor weights was found to differ significantly from zero, t(20) = 12.05, p < .001. For the ICA results, a one-sample t-test was conducted on the beta weights obtained from the component with the highest R-square value and it was found that the beta weights of this component differed significantly from zero, t(20) = 4.54, p < .001. 21 patients, single condition (any stimulus) HRF design matrix analysis The functional brain images from these analyses can be found in Figure 6. For the CPCA analysis, the mean of the predictor weights was found to differ significantly from zero, t(20) = 3.77, p < .005. For the ICA results, a one-sample t-test was conducted on the beta weights obtained from the component with the highest R-square value and  147  it was found that the mean of the betas for this component differed significantly from zero, t(20) = 2.22, p < .05. 21 controls, single condition (any stimulus) FIR design matrix analysis The functional brain images from these analyses can be found in Figure 7. The results of the CPCA ANOVA, F(6,120) = 4.44, p < .05, indicate that this pattern of results was unlikely to have occurred by chance. For the ICA results, it was found that there was a significant effect of time, F(7,160) = 2.42, p < .05. For the PLS analysis, only one LV could be extracted due to the fact that a single condition was modeled. Therefore, the significance of the LV could not be determined via permutation as there were no condition orders to permute. 21 patients, single condition (any stimulus) FIR design matrix analysis The functional brain images from these analyses can be found in Figure 8. The results of the CPCA significance test, F(6,120) = 18.39, p < .001, indicate that this pattern of results was highly unlikely to have occurred by chance. For the ICA results, and it was found that there was a significant effect of time, F(7,160) = 2.47, p < .05. For the PLS analysis, only one LV could be extracted due to the fact that a single condition was modeled. Therefore, the significance of the LV could not be determined via permutation as there were no condition orders to permute. 21 controls, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 9. For the CPCA analysis, a significant effect of Difficulty was found, F(1,20) = 10.03, p < .01, which indicates that the estimated hemodynamic response was significantly more deactivated in the Difficult conditions relatively to the Easy conditions.  148  The ICA analysis revealed a significant main effect of Performance, F(1,20) = 17.49, p < .001, which was likely due to the increased beta weights in the Non-recalled as opposed to Recalled conditions for this component. 21 patients, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 10. For the CPCA analysis, there was a trend towards an interaction between Difficulty and Performance, F(1,20) = 4.22, p < .10. Follow-up analyses revealed that, in the Non-recalled conditions, there were no significant main effects or interactions, whereas in the Recalled condition, there was a significant effect of Difficulty, F(1,20) = 4.22, p < .10, such that the Difficult conditions elicited a larger estimated hemodynamic response that the Easy conditions. For the ICA analysis, a significant main effect of Performance was detected, F(1,20) = 6.04, p < .05, which was likely due to the increased beta weights in the Non-recalled conditions relative to the Recalled conditions for this component. 21 controls, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 11. For the CPCA analysis, a significant main effect of Difficulty was found, F(1,20) = 4.39, p < .05, which reflects the increased deactivation found in the more difficult trials. Additionally, this analysis discovered a significant interaction between Performance and Time, F(6,120) = 4.73, p < .05. Follow-up tests revealed that the interaction was due to differences between the activity time courses of the Non-recalled and Recalled conditions between 4 seconds and 6 seconds peristimulus time, F(1,20) = 23.17, p < .001, between 8 seconds and 10 seconds peristimulus time, F(1,20) = 17.20, p < .001, and between 10 seconds and 12 seconds peristimulus time, F(1,20) = 4.44, p < .05.  149  For the ICA results, a significant main effect of Difficulty was found, F(1,20) = 4.39, p < .05, which reflects the increased deactivation found in the more difficult trials. Additionally, this analysis discovered a significant interaction between Performance and Time, F(6,120) = 4.73, p < .05. Follow-up tests revealed that the interaction was due to differences between the activity time courses of the Non-recalled and Recalled conditions between 4 seconds and 6 seconds peristimulus time, F(1,20) = 23.17, p < .001, between 8 seconds and 10 seconds peristimulus time, F(1,20) = 17.20, p < .001, and between 10 seconds and 12 seconds peristimulus time, F(1,20) = 4.44, p < .05. For the PLS analysis, the first LV extracted accounted for 32.50% of the cross block covariance, and had a p-value of p > .001. 21 patients, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 12. For the CPCA analysis, a significant interaction between Performance and Time was detected, F(6,120) = 5.90, p < .01. Follow-up tests indicated that this interaction was due to a significant main effect of time when only the Recalled conditions were entered into the ANOVA, F(6,120) = 5.08, p < .01, in conjunction with a lack of a significant main effect of time when only the Non-recalled conditions were analyzed. For the ICA results, a significant interaction between Performance and Time, F(6,120) = 6.48, p < .01. Follow-up tests showed that this interaction was due to significant differences between the Recalled and Non-recalled conditions between 2 to 4 seconds, F(1,20) = 5.57, p < .05, 6 to 8 seconds, F(1,20) = 8.39, p < .01, and 10 to 12 seconds, F(1,20) = 7.28, p < .05. For the PLS analysis, the first LV extracted accounted for 30.12% of the cross block covariance, and had a p-value of p < .01.  150  21 controls and 21 patients, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 13. For the CPCA analysis, a significant interaction between Difficulty and Performance was found, F(1,40) = 9.29, p < .005. This interaction resulted from the fact that for the Easy conditions, there was a significant main effect of Performance, F(1,41) = 6.12, p < .05, but for the Hard condition, there were no significant effects, indicating that the neural response to Recalled and Non-recalled stimuli did not differ. For the ICA results, a significant 3-way interaction between Difficulty, Performance, and Group was detected, F(1,40) = 10.80, p < .005. Follow-up tests indicated that this interaction was due to a significant interaction between Difficulty and Performance in the control subjects, F(1,20) = 7.84, p < .05, but significant main effects of Difficulty, F(1,20) = 7.84, p < .05, and Performance, F(1,20) = 7.84, p < .05, in the patient subjects. The interaction between Difficulty and Performance in the controls was examined further, and it was found that it resulted from a significant effect of Performance in the Easy conditions, F(1,20) = 16.02, p < .005, and a non-significant effect of Performance in the Difficult conditions. 21 controls and 21 patients, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 14. For the CPCA analysis, a significant main effect of Difficulty was found, F(1,40) = 5.68, p < .05, with the more difficult conditions being characterized by an increase in activation in this component in the more difficult conditions. This analysis also found a significant 4-way interaction between Condition Type, Performance, Time and Group, F(6,240) = 3.58, p < .05. Separate examinations of both the control group and the patient group revealed that the patients had a marginally significant interaction between Condition Type, Performance, and Time after  151  Greenhouse-Geisser correction, F(2.89,57.84) = 2.56, p < .10, whereas the controls did not show a significant interaction between this set of variables, instead they showed only an interaction between Performance and Time, F(6,120) = 4.99, p < .01. This interaction in the patient group was broken further by separately examining Non-recalled and Recalled responses. This analysis revealed that Time was significant for Recalled responses, F(6,120) = 3.92, p < . 05 but not for Non-recalled responses, F(6,120) = 1.40, p > .25. There were no significant interactions, suggesting that the 4 way interaction may be due to marginal effects. For the ICA results, a significant 3 way interaction between Difficulty, Performance, and Time was found, F(7,280) = 4.40, p < .01. In order to explore this interaction further, follow up test were undertaken in which the easy and difficult conditions were analyzed separately. In the Easy conditions, a significant interaction was found between Performance and Time, F(7,287) = 7.39, p < .001. It was found that this interaction resulted from differences in the Non-recalled and Recalled conditions between 4 and 6 seconds peristimulus time, F(1,41) = 6.39, p < .05 , and between 12 and 14 seconds peristimulus time, F(1,41) = 17.01, p < .001. An analysis of the Difficult conditions revealed a significant interaction between Condition Type and Performance, F(1,41) = 5.39, p < .05. This interaction resulted from a greater difference between recalled and non-recalled words from the external condition than recalled and non-recalled words from the read condition, F(1,41) = 5.38, p < .05. The first component extracted from the PLS accounted for 20.20% of the cross block covariance, which corresponds to p < .0001.  152  Statistics for the Multiple Component Analyses For all of these analyses, the first component described for each statistical method will be identical to the component and analysis described in the relevant single component section of this work. 21 controls, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 15. For the CPCA analysis, 3 components were used. For component 1, a significant effect of Difficulty was found, F(1,20) = 10.03, p < .01, which indicates that the estimated hemodynamic response was significantly more deactivated in the Difficult conditions relatively to the Easy conditions. For component 2, a significant interaction was found between Difficulty and Performance, F(1,20) = 7.41, p < .05. Follow-up tests conducted on the Recalled and Non-called conditions separately revealed that this interaction was due to a significant main effect of Difficulty for the Recalled conditions, F(1,20) = 14.87, p < .005, such that the Hard conditions elicited more activation than the Easy conditions. There were no significant effects in the Nonrecalled conditions. Component 3 resembled component 2, in that a significant interaction was found between Difficulty and Performance, F(1,20) = 6.08, p < .05. Follow-up tests conducted separately on the Recalled and Non-recalled conditions suggested that this interaction was due to the fact that, for the Recalled conditions, there was a significant main effect of Difficulty, F(1,20) = 4.74, p < .05, whereby the Hard conditions elicited more activation than the Easy conditions. There were no significant main effects for the Non-recalled conditions. For the ICA analysis, 4 components were extracted. For component 1, a significant main effect of Performance was detected, F(1,20) = 17.49, p < .001, which was likely due to the increased beta weights in the non-recalled conditions for this component. The ANOVA  153  conducted on component 2 revealed a significant main effect of Performance, F(1,20) = 5.78, p < .05, as well as a significant main effect of Condition Type, F(1,20) = 8.83, p < .01. These findings were likely due to the greater deactivation in Non-recalled than Recalled conditions, and greater deactivation in the Task Condition type relative to the Source Condition Type. For component 3, the ANOVA revealed a significant interaction between Difficulty and Performance, F(1,20) = 16.35, p < .005. Follow-up tests showed that in the Easy conditions, there was a significant main effect of Performance, F(1,20) = 15.49, p < .005, whereas in the Hard conditions there were no significant main effects or interactions. For the ICA component with the fourth highest R-square value, the ANOVA revealed a significant interaction between Condition Type and Performance, F(1,20) = 5.04, p < .05. Further analysis showed that this interaction was due to a significant main effect of Condition Type when Non-recalled responses were examined, F(1,20) = 4.42, p < .05, such that the Source Condition Type stimuli were associated with higher beta values. No significant main effects or interactions were found when only the Recalled responses were examined. 21 controls, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 16. For the CPCA analysis, 4 components were extracted. For component 1, significant main effect of Difficulty was found, F(1,20) = 4.39, p < .05, which reflects the increased deactivation found in the more difficult trials. Additionally, this analysis discovered a significant interaction between Performance and Time, F(6,120) = 4.73, p < .05. Follow-up tests revealed that the interaction was due to differences between the activity time courses of the Non-recalled and Recalled conditions between 4 seconds and 6 seconds peristimulus time, F(1,20) = 23.17, p < .001, between 8 seconds and 10 seconds peristimulus time, F(1,20) = 17.20, p < .001, and between 10 154  seconds and 12 seconds peristimulus time, F(1,20) = 4.44, p < .05. The analysis of the second CPCA component revealed a significant three way interaction between Difficulty, Performance, and Time, F(6,20) = 4.27, p < .01. Follow-up tests revealed that this interaction was due to a significant interaction between Difficulty and Time in the Recalled condition, but no significant interaction in the Non-recalled condition. Furthermore, the interaction between Difficulty and Time in the Recalled condition was found to be caused by significant differences between the Easy and Hard conditions between 4 to 6 seconds, F(1,20) = 6.49, p < .05, 6 to 8 seconds, F(1,20) = 14.48, p < .005, 8 to 10 seconds, F(1,20) = 8.29, p < .01, and 10 to 12 seconds, F(1,20) = 6.56, p < .05. The analysis of the third CPCA component found 2 significant interactions; the first was between Difficulty and Condition Type, F(1,20) = 5.82, p < .05, and the second was between Performance and Time, F(6,120) = 23.96, p < .001. The interaction between Difficulty and Condition Type arose due to the fact that the Internal condition (which is the Easy Source condition) shows less deactivation than the other three encoding conditions. The interaction between Performance and Time was caused by significant differences between Recalled and Non-recalled between 4 to 6 seconds, F(1,20) = 7.82, p < .05, 6 to 8 seconds, F(1,20) = 7.82, p < .05, 8 to 10 seconds, F(1,20) = 40.61, p < .001, and 10 to 12 seconds, F(1,20) = 8.75, p < .01. The analysis of the fourth CPCA component revealed 2 significant interactions; the first was between Performance and Time, F(6,120) = 4.09, p < .05, and the second was between Difficulty and Condition Type, F(1,20) = 5.74, p < .05. Follow-up analyses revealed that the interaction between Performance and Time was due to significant differences between the activity profiles of the Recalled and Non-recalled responses between 2 to 4 seconds, F(1,20) = 6.76, p < .05, 8 to 10 seconds, F(1,20) = 8.74, p < .01, and 12 to 14 seconds peristimulus time, F(1,20) = 10.82, p < .01. The significant interaction between Difficulty and Condition Type was due to a significant  155  difference between the Condition Type in the Easy Difficulty condition, F(1,20) = 5.05, p < .05, but no significant differences between Condition Type in the Hard Difficulty condition. For the ICA analysis, 2 components were extracted. The ANOVA conducted on the beta weights of the ICA component with the highest R-square value revealed a significant interaction between Performance and Time, F(6,120) = 18.38, p < . 001, such that the betas for the Recalled and Non-recalled stimuli differed significantly between 6 to 8 seconds, F(1,20) = 16.28, p < . 001, and 8 to 10 seconds peristimulus time, F(1,20) = 12.71, p < . 005. The second ICA component extracted revealed a significant interaction between Difficulty and Time, F(6,120) = 3.00, p < . 05, and a significant interaction between Performance and Time, F(6,120) = 9.59, p < . 001. Follow-up analyses revealed that the interaction between Difficulty and Time was due to significant differences between the activity profiles of the Easy and Hard responses between 12 to 14 seconds, F(1,20) = 4.48, p < .05, The interaction between Performance and Time was due to significant differences between the activity profiles of the Recalled and Non-recalled responses between 2 to 4 seconds, F(1,20) = 10.53, p < .005, 4 to 6 seconds, F(1,20) = 7.81, p < .05, 8 to 10 seconds, F(1,20) = 10.08, p < .01, and 10 to 12 seconds peristimulus time, F(1,20) = 8.81, p < .01. For the PLS analysis, the first LV extracted accounted for 32.50% of the cross block covariance, and had a p-value of p > .001. The second LV accounted for 25.40% of the cross block covariance, and had a p-value of p > .005. 21 patients, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 17. For the CPCA analysis, 3 components were extracted. A repeated measures ANOVA conducted on the predictor weights from the first component found a trend towards an interaction between Difficulty and Performance, F(1,20) = 4.22, p < .10. Follow-up analyses revealed that, in 156  the Non-recalled conditions, there were no significant main effects or interactions, whereas in the Recalled condition, there was a significant effect of Difficulty, F(1,20) = 4.22, p < .10, such that the Difficult conditions elicited a larger estimated hemodynamic response that the Easy conditions. For the second CPCA component extracted in this analysis, a repeated measures ANOVA found a significant interaction between Difficulty and Performance, F(1,20) = 27.37, p < .001. Follow-up tests showed that this interaction was due to a non-significant difference between Recalled and Non-recalled in the Easy conditions, but a significant difference between Recalled and Non-recalled in the Difficult conditions, F(1,20) = 5.14, p < .05. The statistical analysis of the third CPCA component revealed a significant main effect of Performance, F(1,20) = 23.33, p < .001, such that the Recalled conditions elicited greater deactivation that the Nonrecalled conditions. For the ICA analysis, 3 components were extracted. For the component with the highest R-square value, a significant main effect of Performance was detected, F(1,20) = 6.04, p < .05, which was likely due to the increased beta weights in the Non-recalled conditions relative to the Recalled conditions for this component. For the component with the second highest R-square value, there were no significant main effects or interactions. For the component with the third highest R-square value, there were also no significant main effects or interactions, although there was a trend with regards to performance, F(1,20) = 3.08, p < .10, most likely due to increased beta weights in the Non-recalled relative to Recalled conditions. 21 patients, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 18. For the CPCA analysis, 4 components were extracted. A repeated measures ANOVA of the predictor weights from the first component extracted from the CPCA analysis revealed a  157  significant interaction between Performance and Time, F(6,120) = 5.90, p < .01. Follow-up tests indicated that this interaction was due to a significant main effect of time when only the Recalled conditions were entered into the ANOVA, F(6,120) = 5.08, p < .01, in conjunction with a lack of a significant main effect of time when only the Non-recalled conditions were analyzed. The repeated measures ANOVA conducted on the second CPCA component from this analysis revealed a significant three-way interaction between Difficulty, Performance, and Time, F(6,120) = 3.13, p < .05. This three-way interaction was found to be due to differences between the two-way interactions between Difficulty and Time when Recalled and Non-recalled conditions were examined separately. In the Recalled conditions, a significant interaction between Difficulty and Time was found, F(6,120) = 3.58, p < .05,whereas in the Non-recalled conditions, there was only a significant main effect of Time, F(6,120) = 32.45, p < .001. With regards to the significant interaction between Difficulty and Time in the Recalled conditions, further analysis found that the difference between Easy and Hard conditions was significant between 2 to 4 seconds, F(6,120) = 4.90, p < .05, and 4 to 6 seconds, F(6,120) = 4.45, p < .05 , indicating that the deactivations in this network were stronger during these time periods in the Hard conditions. For the third component extracted from this CPCA analysis, 2 different interactions were found. The first was between Performance and Time, F(6,120) = 13.28, p < .001, and the second was between Performance and Difficulty, F(1,20) = 5.49, p < .05. The interaction between Performance and Time was found to be caused by differences between Recalled and Non-recalled conditions between 0 to 2 seconds, F(1,20) = 5.87, p < .05, 2 to 4 seconds, F(1,20) = 17.59, p < .001, 4 to 6 seconds, F(1,20) = 8.62, p < .01, and 6 to 8 seconds, F(1,20) = 41.35, p < .001. The interaction between Difficulty and Performance was due to the fact that, in the Easy conditions, Nonrecalled stimuli elicited greater activation than Recalled stimuli but in the Hard condition,  158  Recalled stimuli elicited more activation than Non-recalled stimuli. For the fourth component extracted in this CPCA analysis, a significant interacton was found between Performance and Time, F(6,120) = 9.72, p < .001. Follow-up tests demonstrated that this interaction was due to differences between the predictor weights for the Recalled and Non-recalled conditions between 0 to 2 seconds, F(1,20) = 16.87, p < .005, 4 to 6 seconds, F(1,20) = 17.37, p < .001, and 6 to 8 seconds, F(1,20) = 4.63, p < .05. For the ICA analysis, the 4 components with the highest R-square regression values were extracted. The ANOVA performed on the component with the highest R-square value found a significant interaction between Performance and Time, F(6,120) = 6.48, p < .01. Follow-up tests showed that this interaction was due to significant differences between the Recalled and Nonrecalled conditions between 2 to 4 seconds, F(1,20) = 5.57, p < .05, 6 to 8 seconds, F(1,20) = 8.39, p < .01, and 10 to 12 seconds, F(1,20) = 7.28, p < .05. For the component found to have the second highest R-square value, no significant main effects or interactions were detected. For the component with the third highest R-square value, the ANOVA found a significant main effect of Time, F(6,120) = 15.31, p < .001, as well as a significant interaction between Condition, Time, and Performance, F(1,20) = 4.56, p < .05. The interaction between Condition Type and Performance was found to be due to a significant main effect of Condition Type in the Recalled beta weights, F(1,20) = 10.15, p < .01, but a non-significant effect of Condition Type in the Non-recalled beta weights. This significant effect of Condition Type in the Recalled beta weights is likely due to decreased deactivation/increased activation in the Source Condition Type stimuli relative to the Task Condition Type stimuli throughout the course of the trial. With regard to the component with the fourth highest R-square value, this analysis found no significant main effects or  159  interactions, although there was a trend towards significance in the interaction between Condition Type, Performance, and Time, F(6,120) = 2.22, p < .10. For the PLS analysis, the first LV extracted accounted for 30.12% of the cross block covariance, which corresponds to p < .01. The seconds LV extracted accounted for 20.85% of the cross-block covariance, which corresponds to p < .15. Although this 2nd LV was not found to be significant, it was included for purposes of comparison to the other multivariate methods. 21 controls and 21 patients, 8 conditions HRF design matrix analysis The functional brain images from these analyses can be found in Figure 19. For the CPCA analysis, 4 components were extracted. The ANOVA conducted on the first component extracted from this CPCA analysis found a significant interaction between Difficulty and Performance, F(1,40) = 9.29, p < .005. This interaction resulted from the fact that for the Easy conditions, there was a significant main effect of Performance, F(1,40) = 6.12, p < .05, but for the Hard condition, there were no significant effects, indicating that the neural response to Recalled and Non-recalled stimuli did not differ. The analysis of the second extracted CPCA component revealed a significant interaction between Difficulty and Performance, F(1,40) = 15.21, p < .001. This interaction appears to be driven by the fact that the predictor weights in the Recalled conditions increased from the Easy to the Hard difficulty condition, whereas the predictor weights in the Non-recalled conditions decreased from the Easy to the Hard condition. The third component extracted from the CPCA analysis revealed a significant four-way interaction between Difficulty, Condition Type, Performance, and Group, F(1,40) = 5.77, p < .05. Follow-up tests found that this was due to a significant three-way interaction between Difficulty, Condition Type, and Performance in the controls subjects, F(1,20) = 5.02, p < .05, but only a significant main effect of Performance in the patient group, F(1,20) = 160  34.94, p < .001.The three-way interaction in the control subjects was further analyzed, and was found to result from a significant interaction between Difficulty and Condition Type in the Nonrecalled conditions, F(1,20) = 7.84, p < .05,but no significant main effects or interactions in the Recalled conditions. For the significant interaction between Difficulty and Condition Type in the Non-recalled conditions in the control subjects, the marginal means indicate that in the Source Condition Type, the predictor weights were greater for the Easy condition (Internal) than the Difficult condition (External); whereas in the Task Condition Type, the predictor weights were greater for the Difficult condition (Reading) than the Easy condition (Association). For the fourth component extracted from this CPCA analysis, no significant main effects or interactions were found. For the ICA analysis, 4 components were extracted. The ANOVA conducted on the beta weights from the ICA component with the highest R-square value indicated that there was a significant 3-way interaction between Difficulty, Performance, and Group, F(1,40) = 10.80, p < .005. Follow-up tests indicated that this interaction was due to a significant interaction between Difficulty and Performance in the control subjects, F(1,20) = 17.46, p < .001, but significant main effects of Difficulty, F(1,20) = 5.80, p < .05, and Performance, F(1,20) = 7.90, p < .05, in the patient subjects. The interaction between Difficulty and Performance in the controls was examined further, and it was found that it resulted from a significant effect of Performance in the Easy conditions, F(1,20) = 16.02, p < .005, and a non-significant effect of Performance in the Difficult conditions. With regards to the ICA component with the second highest R-square value, the ANOVA revealed a significant interaction between Condition Type and Group, F(1,40) = 4.36, p < .05, such that the controls exhibited a significant effect of Condition Type, F(1,20) = 5.07, p < .05, whereas the patients did not. In the controls, the betas from the Source Condition Type were  161  greater than the betas for the Task Condition Type. For the component with the third highest Rsquare value, the ANOVA found 2 significant interactions, the first was between Difficulty and Performance, F(1,40) = 7.94, p < .01 , and the second was between Condition Type and Group, F(1,40) = 5.35, p < .05. The interaction between Difficulty and Performance was found to be due to a significant main effect of Performance in the Easy conditions, F(1,40) = 19.07, p < .001, which was not found in the Hard conditions. The interaction between Condition Type and Group was due to a significant main effect of Condition Type in the patient sample, F(1,20) = 7.10, p < .05, that was not found in the control sample. In the patients, the Source Condition Type betas were larger than the Task Condition Type betas. For the component with the fourth highest Rsquare value, the ANOVA revealed a significant main effect of Difficulty, F(1,40) = 4.28, p < .05 , whereby the betas from the Difficult conditions were more negative than the betas from the Easy conditions. Also, this analysis found a significant interaction between Performance and Group, F(1,40) = 5.70, p < .05, which, when analyzed further, was discovered to result from a significant effect of Performance being detected only in the control group, F(1,20) = 9.70, p < .01. In the controls, the betas from the Recalled conditions were less negative than the betas from the Nonrecalled conditions. 21 controls and 21 patients, 8 conditions FIR design matrix analysis The functional brain images from these analyses can be found in Figure 20. For the CPCA analysis, 4 components were extracted. This analysis revealed a significant main effect of Difficulty, F(1,40) = 5.68, p < .05, with the more difficult conditions being characterized by an increase in activation in this component in the more difficult conditions. This analysis also found a significant 4-way interaction between Condition Type, Performance, Time and Group, F(6,240) = 3.58, p < .05. Separate examinations of both the control group and the  162  patient group revealed that the patients had a marginally significant interaction between Condition Type, Performance, and Time after Greenhouse-Geisser correction, F(2.89,57.84) = 2.56, p < .10, whereas the controls did not show a significant interaction between this set of variables, instead they showed only an interaction between Performance and Time, F(6,120) = 4.99, p < .01. This interaction in the patient group was broken further by separately examining Non-recalled and Recalled responses. This analysis revealed that Time was significant for Recalled responses, F(6,120) = 3.92, p < . 05 but not for Non-recalled responses, F(6,120) = 1.40, p > .25. There were no significant interactions, suggesting that the 4 way interaction may be due to marginal effects. For the second component extracted from the CPCA analysis, a significant three-way interaction was found between Difficulty, Performance and Time, F(6,240) = 7.11, p < . 001. This interaction was due to the fact that in the Non-recalled conditions, there was only a main effect of Time, F(6,240) = 91.80, p < . 001, whereas in the Recalled condition there was a significant interaction between Difficulty and Time, F(6,240) = 11.45, p < . 001. This interaction between Difficulty and Time was driven by differences between the predictor weights between 4 seconds to 6 seconds, F(1,40) = 14.00, p < . 005, 6 seconds to 8 seconds, F(1,40) = 15.75, p < . 001, 8 seconds to 10 seconds, F(1,40) = 6.82, p < . 05, 10 seconds to 12 seconds, F(1,40) = 9.88, p < . 005, and 12 seconds to 14 seconds F(1,40) = 6.48, p < . 05, whereby the Difficult conditions had higher predictor weights than the Easy conditions. With regards to the third component extracted from this CPCA analysis, the ANOVA found 2 significant interactions, the first involved Difficulty, Condition Type, Time, and Group, F(6,240) = 2.77, p < . 05, and the second involved Performance and Time, F(6,240) = 6.81, p < . 001. With regards to the first interaction mentioned, further analysis found that in the Source Condition Type, there was a significant three-way interaction between Difficulty, Time, and Group, F(6,240) = 2.85, p < . 05, whereas in the Task Condition  163  Type, there was only a significant main effect of Time, F(6,240) = 15.35, p < . 001. When the significant interaction between Difficulty, Time, and Group was analyzed further, it was found that, for the control subjects, there was a significant interaction between Difficulty and Time, F(6,120) = 4.51, p < . 01, whereas for the patient subjects, there were no significant main effects or interactions. The interaction between Performance and Time in controls subjects was caused by significant differences between the predictor weights in the Recalled and Non-recalled conditions between 8 seconds to 10 seconds peristimulus time, F(1,20) = 5.80, p < . 05. With regards to the second significant interaction found in this analysis between Performance and Time, follow-up tests revealed that this was due to significant differences between the predictor weights for Recalled and Non-recalled conditions between 6 to 8 seconds, F(1,40) = 6.06, p < . 05, and 12 to 14 seconds, F(1,40) = 6.63, p < . 05. For the fourth component extracted from this CPCA analysis, a significant interaction between Performance and Time was found, F(6,240) = 9.27, p < . 001. Follow-up tests revealed that this was due to significant differences between the predictor weights for Recalled and Non-recalled conditions between 4 to 6 seconds, F(1,40) = 6.31, p < . 05, 8 to 10 seconds, F(1,40) = 25.51, p < . 001, and 10 to 12 seconds, F(1,40) = 4.54, p < . 05. For the ICA analysis, 4 components were extracted. For the component with the highest R-square value, the ANOVA conducted on the beta weights revealed a significant interaction between Performance and Time, F(6,240) = 18.38, p < . 001, such that the betas for the Recalled and Non-recalled stimuli differed significantly between 6 to 8 seconds, F(1,40) = 16.28, p < . 001, and 8 to 10 seconds peristimulus time, F(1,40) = 12.71, p < . 005. For the ICA component with the second highest R-square values, the results of the ANOVA indicated a significant main effect of Performance, F(1, 40) = 4.94, p < .05, such that the betas from the Non-recalled conditions were higher than the betas for the Recalled conditions. This analysis also revealed a significant  164  interaction between Difficulty and Time, F(6, 240) = 3.69, p < .05, which resulted from a significant difference between the betas for the Easy conditions and the Hard conditions between 4 to 6 seconds peristimulus time, F(1, 40) = 8.19, p < .01. The ICA component with the third highest R-square value was found to show a significant 5-way interaction between Difficulty, Condition Type, Performance, Time, and Group, F(6, 240) = 3.55, p < .01. This was found to result from a significant 3-way interaction in the controls between Difficulty, Performance, and Time, F(6, 120) = 4.02, p < .01, and a significant 3-way interaction in the patients between Difficulty, Condition Type, and Time, F(6, 120) = 3.01, p < .05. In the controls, the 3-way interaction between Difficulty, Performance, and Time was found to be due to a significant interaction between Performance and Time in the Easy conditions, F(6, 120) = 9.02, p < .001, whereas in the Hard conditions, there was only a significant main effect of Time, F(6, 120) = 16.46, p < .001. The betas for the Recalled and Non-recalled stimuli were found to differ between 6 to 8 seconds, F(1, 20) = 6.87, p < .05, and 8 to 10 seconds peristimulus time, F(1, 20) = 6.04, p < .05. For the significant 3way interaction in the patients between Difficulty, Condition Type, and Time, this was found to be caused by the fact that for the Source Condition Type there was a main effect of Time, F(6, 120) = 14.72, p < .001, whereas in the Task Condition Type, there was a significant interaction between Difficulty and Time, F(6, 120) = 3.51, p < .05, such that the Easy and Hard conditions differed significantly between 10 to 12 seconds peristimulus time, F(1, 20) = 6.03, p < .05. For the ICA component with the fourth highest R-square value, the ANOVA revealed a significant main effect of Time, F(6, 240) = 25.71, p < .001. For the PLS analysis, the first LV extracted accounted for 20.20 % of the cross block covariance, which corresponds to p < .001. The seconds LV extracted accounted for 16.35% of the cross-block covariance, which corresponds to p < .001. 165  Mathematics Underlying Behavioural PLS The behavioural PLS analysis involves generating a correlation matrix between the M matrix (brain activity data matrix) and a matrix (or vector) of behavioural measures A. The A matrix is n × k rows and contains one column for each of the c behavioural measures of interest. The columns of the A and M matrices are mean-centered for each of the k conditions (the following equation is performed separately for each of the k conditions): (18) Where c is number of conditions of behavioural interest, k is the number of conditions, m is the number of voxels, t is the number of time points, n is the number of observations, and Rk is the correlation of obtained BOLD response across subjects with the behavioural measures of interest for condition k. The Rk matrices (one for each condition of experimental interest) are stacked into one large R matrix, which is then subjected to PCA: (19) where k is the number of conditions, m is the number of voxels, t is the number of time points, and the square brackets denote the products of singular value decomposition. The left singular vectors, U, are labeled the voxel saliences, the right singular vectors, V, are labeled the contrast saliences, and S is the diagonal matrix of singular values. Similar to the mean centering PLS approach discussed earlier, the weights in V indicate task-related differences, however in this case, they reflect the correlation between brain activity and the behavioural measure of interest in each of the experimental conditions as related to each of the extracted components.  166  The Correlations between Behavioural Performance and Brain Activity Although the analyses presented in this report have not been fruitful in detecting a significant difference between the groups in either the location or intensity of neural activity, this investigation into the neural correlates of source monitoring in schizophrenia also sought to examine whether the individual differences in neural activity were correlated with individual differences in recall accuracy. Please see Figure 21 for the functional brain images, the graphical depictions of the correlations, and, where possible, the estimated activity time courses. This analysis is possible given that, for each of the multivariate methods employed in this study, there are individual subject values reflecting the intensity of each component‟s activation for that subject. However, there are notable differences in the way each of the methods performs this correlational analyses. The CPCA analysis involved examining the component loadings for each subject and component, and correlating the component loading peak activation (or deactivation) with that‟s subject‟s accuracy in correctly recalling the encoding condition for each stimulus. The ICA analysis involved extracting the activity time courses for the components that regress most highly onto the SPM5 design matrix. Then, for each component and each subject, a mean activity profile for each trial is produced and its peak activation (or deactivation) is correlated with that subject‟s accuracy. The PLS analysis proceeds by generating an M matrix (as discussed previously in the section outlining the computational processes involved in a PLS analysis). Then, the M matrix is pre-multiplied by the transpose of the vector or matrix of behavioural measures (called A), and the resulting covariance matrix, R, is decomposed using SVD. The formulae outlining this process can be found in equations 17 and 18. For each of these analyses, two components were correlated with behavioural accuracy.  167  The first component extracted from the CPCA analysis was characterized by bilateral activations in the dAcc/SMA, pre-central gyri, fusiform gyri, inferior, middle, and superior occipital gyri, crus I and lobule 6 of the cerebellum, and left post-central and supramarginal gyri. The correlation between this component‟s activity peak and individual accuracy was nonsignificant in the controls, r(21) = -.03, p > .50, and in the patients (although it was strong enough to constitute a trend), r(21) = .40, p < .10. The difference between the two correlations was also non-significant, z = -1.36 ,p > .15. Component 2 from the CPCA analysis was characterized by bilateral activations in the dAcc/SMA, inferior and middle occipital gyrti, as well as left precentral, post-central and supramarginal gyri. Deactivations were found bilaterally in mPFC and posterior cingulate, as well as in the right superior and middle temporal gyri,a nd right supramarginal gyrus. The correlation between this component‟s activity peak and individual accuracy was non-significant in the controls, r(21) = -.31, p > .15, but was significant for the patients, r(21) = -.86, p < .001. The difference between the two was also significant, z = 2.92, p < .005. The two highest regressing components from the previous ICA analysis were used for the correlational analysis. The first component was characterized by bilateral activations in the inferior, middle, and superior occipital cortices, as well as crus I and lobule 6 of the cerebellum. The correlations between peak activity in this component and performance accuracy was nonsignificant in the controls, r(21) = .31, p > .15, but was significant in the patients, r(21) = .48, p < .05. The difference between the two was non-significant, z = -0.61, p > .50. The second component was characterized by bilateral activations in dAcc, pre-central and post-central gyri, although the activations in the latter two regions are predominantly left-sided. The correlations between peak activity in this component and performance accuracy was non-significant in the  168  controls, r(21) = -.11, p > .50, and patients, r(21) = .28, p > .15. The difference between the two was non-significant, z = -1.19, p > .15. The two components resulting from the PLS analysis were characterized by widespread foci of activation. The first component revealed bilateral activations in the mPFC, dAcc/SMA, inferior temporal poles, hippocampi, superior and middle temporal gyri, thalamus, pre-central gyrus, post-central gyrus, supramarginal gyri, and crus II and lobule 8 of the cerebellum. The correlations between peak activity in this component and performance accuracy was nonsignificant in the controls, r(21) = .31, p > .15, but was significant in the patients, r(21) = .82, p < .001. The difference between the two was significant, z = -2.51, p < .01. The second component revealed bilateral activations in the dAcc/SMA, pre-central gyrus, post-central gyrus, middle and superior tempotal gyri, thalamaus, supramarginal gyri, middle and inferior occipital gyri, crus I and lobule 6 of the cerebellum, and brainstem.. The correlations between peak activity in this component and performance accuracy was significant in the controls, r(21) = .90, p < .001, and also in the patients, r(21) = .51, p < .05. The difference between the two was significant, z = 2.73, p < .01. Discussion of correlations between brain and behaviour The final analysis that was undertaken in this study was to correlate the components extracted from each of the multivariate methods with a behavioural measure, in this case, response accuracy. For the PLS analysis, this process is implemented within the program itself (see Equations 15 & 16 for a mathematical explanation of this analysis), however, for the CPCA and ICA analyses, each participant‟s component time course activity peaks was correlated with their overall accuracy in the experiment. The results of these analyses indicate that, with the exception of the 2nd PLS component, patients have stronger correlations between neural activity  169  and performance, perhaps reflecting the greater need for coordination in order to perform the task successfully. Additionally, for the CPCA and ICA analyses, the activity time courses of the components reveal no differences between the mean patient and control activation profiles, yet differences are found in the correlations between brain activity and behaviour. For the CPCA and PLS analyses, significant differences were found between the correlations for the patients and controls. It is most likely expected that at least the first component from the ICA analysis would not find significant differences between patients and controls given that the visual cortices are the areas of activity, and it would be unexpected if differences were found in regions that are nearly exclusively devoted to sensory processing. Analyses Specific to SPM5 All of the SPM5 results presented until this point have examined the contrast of any stimuli versus baseline in order to provide grounds for comparison with the multivariate methods under consideration in this work. However, SPM5 is more frequently to examine regional differences in condition-specific brain activity using the subtraction method (Posner et al., 1988). In a subtraction method paradigm, an experimental task is paired with a control task such that the control task and the experimental task are as closely matched as is possible (e.g. requiring that both tasks use the same stimuli as input and same motor responses as output). In the source monitoring experiment being described here, there were two separate choices for control and experimental tasks. The first experimental/control condition dyad is the contrast of the Self/Internal condition versus the Other/External condition, and the second experimental/control condition dyad is the contrast of the source monitoring (Self and Other ) condition versus the task monitoring (Association and Reading) condition. The SPM5 contrast based results are presented in Figure 22, Figure 23, Figure 24, Figure 25, and Figure 26, and are discussed below. 170  21 controls, 8 conditions HRF design matrix contrasts This section displays the significant clusters of activation found when contrasting the conditions of experimental interest in the 21 controls using an HRF design matrix. The functional brain images from these analyses can be found in Figure 21. Source recalled greater than task recalled This contrast revealed significant activations in the right middle and superior temporal gyri, as well as in the right angular gyrus. Task recalled greater than task non-recalled This analysis revealed significant activations in the left middle and superior temporal gyri, left angular gyrus, as well as bilateral activations in the posterior cingulate/precuneus. Internal recalled greater than external recalled This analysis revealed significant bilateral activations in the medial frontal gyri and medial prefrontal cortices. Internal recalled greater than internal non-recalled This analysis revealed significant bilateral activations in the medial frontal gyri and medial prefrontal cortices. 21 controls, 8 conditions FIR design matrix contrasts This section displays the significant clusters of activation found when contrasting the conditions of experimental interest in the 21 controls using an FIR design matrix modeling 8 post-stimulus time points. The functional brain images from these analyses can be found in Figure 23.  171  Self recalled greater than other recalled This contrast revealed significant activations in the medial prefrontal cortices and medial frontal gyri. 21 patients, 8 conditions SPM5 HRF design matrix contrasts This section displays the significant clusters of activation found when contrasting the conditions of experimental interest in the 21 patients using an HRF design matrix. The functional brain images from these analyses can be found in Figure 24. Source recalled greater than source non-recalled This contrast revealed significant bilateral activations in the precuneus. Self recalled greater than other recalled This contrast revealed significant bilateral activations in the posterior cingulate cortex, as well as in the left hippocampus and left middle temporal gyrus. Self recalled greater than self non-recalled This contrast revealed significant bilateral activations in the medial prefrontal cortices, subgenual region and posterior cingulate cortex. 21 patients 8 conditions FIR design matrix contrasts This section displays the significant clusters of activation found when contrasting the conditions of experimental interest in the 21 subjects using an FIR design matrix modeling 8 post-stimulus time points. The functional brain images from these analyses can be found in Figure 25.  172  Source recalled greater than task recalled This contrast revealed significant activations in the cuneus and posterior cingulate/precuneus. Self recalled greater than other recalled This contrast revealed significant activations in the right middle temporal gyrus that extend posteriorly into the right middle occipital gyrus. 21 controls and 21 patients, 8 conditions FIR design matrix contrasts This section displays the significant clusters of activation found when contrasting the conditions of experimental interest in the 21 subjects using an FIR design matrix modeling 8 post-stimulus time points. The functional brain images from this analysis can be found in Figure 26. Controls greater than patients, source recalled greater than source non-recalled This contrast revealed a significant cluster of activation in the left superior and middle temporal temporal gyri including Heschl‟s gyrus and planum temporale. Discussion of Analyses Specific to SPM5 As can be seen from this set of SPM analyses, the majority of the significant activations found using these SPM5 contrasts can be localized to the medial frontal/prefrontal cortices, the superior temporal cortices and the inferior parietal cortices. Not surprisingly, these are the three regions mentioned earlier as being the neuroanatomical areas most commonly found to be active in the source monitoring literature. In this set of SPM analyses, only a single statistically significant difference in regional or network brain activity was found between the schizophrenia and control groups. This finding 173  was the from the SPM5 FIR analysis examining the contrast of controls greater than patients and source recalled greater than source non-recalled. This cluster of significant activity was located in the left middle and superior temporal gyri, corresponding roughly to Brodmann‟s area 22. Since this region is known to be critically involved in speech perception (Hickok and Poeppel, 2007), this finding is in accordance with the contextual reinstatement theory of recollection, which states that when subjects attempt to remember an episode, their neural activity will resemble the pattern of activity that was present during encoding (Polyn et al., 2005). In this case, the source conditions involved hearing a word, spoken by either oneself or the computer, so the reinstantiation hypothesis would predict that speech perception regions would be involved in recalling those episodes. Decreased activation in the left superior and middle temporal cortices in schizophrenia patients during episodic recall has also been found in the previous studies (Hofer et al., 2003). The lack of any other significant differences between schizophrenia patients and healthy controls in this study mirrors the findings of a recent review of experiments examining corollary discharge during inner speech in schizophrenia (Allen et al., 2007), which found that there was only ambiguous evidence for a verbal self-monitoring deficit in schizophrenia, whereby only a subset of the studies examined managed to find any significant difference between the two groups. Potential limitations of this study that perhaps contributed to the lack of significant differences between controls and patients will be discussed in a subsequent section. An examination of the neural activity resulting from recall of words in different encoding conditions revealed many significant differences. These shall be discussed in turn, beginning with the SPM5 HRF results. In the control subjects, significant differences were found between the source monitoring recalled greater than task monitoring recalled contrast, the task monitoring  174  recalled greater than task monitoring non-recalled contrast, the self recalled greater than other recalled contrast, and the self recalled greater than self non-recalled contrast. The neural activity differences between the source monitoring recalled and task monitoring recalled conditions were found in posterior right superior temporal gyrus in a region also known as the temporo-parietal junction due to its close proximity to the junction between temporal, parietal, and occipital cortices. Previous studies have found that the right temporoparietal junction is involved in distinguishing between one‟s own thoughts and actions and those of another person (Saxe et al., 2004), as well as social cognition and theory of mind (Decety and Lamm, 2007), which fits well with the results obtained from this contrast, as the source monitoring conditions, but not the task monitoring conditions, required the participants to differentiate between actions performed by themselves or another (i.e. the computer). The contrast of task monitoring recalled greater than task monitoring non-recalled revealed significant activations in the left temporo-parietal junction and the bilateral precuneus. These findings agree with previous literature that has implicated the precuneus in both successful episodic memory retrieval (Schmidt et al., 2002; Platel et al., 2003), and making self-referential judgments (Kircher et al., 2002; den Ouden et al., 2005); and the left temporo-parietal junction in the processing of words (Hickok and Poeppel, 2007), as well as memory retrieval (Burianova et al., 2010). However, it is interesting to note that these two regions form part of the default mode network (Raichle et al., 2001), which implies that these regions typically show deactivations during the performance of effortful cognitive tasks. Furthermore, the degree of deactivation in these regions is usually thought to correlate positively with the difficulty of the task (McKiernan et al., 2003). If this logic is applied to the results of the current analysis, the healthy controls put in more effort on the task monitoring trials in which they responded incorrectly.  175  The self recalled greater than other recalled contrast revealed significant bilateral activations in the medial prefrontal cortices. Activations in this region (largely overlapping with Brodmann‟s area 10) are frequently found in studies examining reality monitoring (Simons et al., 2005a), differentiating between one‟s self and other persons (Christoff et al., 2003; Simons et al., 2005b), and theory of mind and mentalizing (Frith and Frith, 2003; Whitehead et al., 2009), and, thus the current results conform well to the previous literature outlining the functional roles this region is thought to play. The contrast of self recalled greater than self non-recalled led significant bilateral activations in the medial prefrontal cortices. As mentioned in the previous paragraph, this finding fits well with the hypothesized functional role of this region, as one would expect that the words recalled in the self condition would elicit more activity in the brain region responsible for differentiating self and other than the non-recalled words. The analysis of the 21 controls‟ data using an SPM FIR design matrix revealed one significant contrast, self recalled greater than other recalled. This significant difference in activity was located in the medial prefrontal cortices. This same region was identified when the same contrast was examined using the SPM HRF design matrix, indicating that this result is robust with regards to the statistical model selected as a regressor in the design matrix. The analysis of the 21 patients‟ data using an SPM HRF design matrix revealed three significant contrasts: source recalled greater than source non-recalled, self recalled greater than other recalled, and self recalled greater than self non-recalled. Using the contrast of source recalled greater than source non-recalled, a cluster of significant activity was found bilaterally in the precuneus, although this activation was much  176  more extensive in the right hemisphere. As mentioned in the preceding section, this region is thought to be involved in successful episodic recall, as well as making self-referential judgments. The contrast of self recalled greater than other recalled revealed significant bilateral activations in the posterior cingulate cortex, as well as in the left hippocampus and left middle temporal gyrus. Activations in the left hippocampus and left middle temporal gyrus are commonly found in studies of source monitoring as the hippocampus is thought to be a neural site for the binding of item and context memory (Dickerson and Eichenbaum, 2009), and the posterior cingulate cortex is thought to be important during recollection, especially when making „remember‟ responses, i.e. when contextual information is retrieved along with item information (Wagner et al., 2005). The contrast of self recalled greater than self non-recalled revealed significant bilateral activations in the medial prefrontal cortices, subgenual region, and posterior cingulate cortex. The medial prefrontal cortices were also found to be active in the 21 control subjects when the same contrast was examined, which suggests that this region performs a similar task in both groups. The subgenual region has been argued to play a role in varying cognitive processes including stress responses, addiction, and memory, although the precise function that it supports is unknown (Morgane et al., 2005). The role of the posterior cingulate cortex in memory has been discussed in the previous paragraph, but this region is thought to play an important role in recollection, especially when context information is recalled along with item information. The analysis of the 21 patients‟ data using an SPM FIR design matrix revealed two significant contrasts: source recalled greater than task recalled and self recalled greater than other recalled.  177  The contrast of source recalled greater than task recalled revealed activations in the cuneus and precuneus. As argued above, the precuneus activations could be reflective of the role of this region in making self-referential judgments and episodic memory recall. Cuneus activity has not generally been associated with episodic memory or source monitoring functions, although other groups have found activity in this region during recognition memory tasks in older adults (Scarmeas et al., 2003) and epilepsy patients (Eliassen et al., 2008), which suggests that this region may play a compensatory role in individuals with compromised neural integrity. The contrast of self recalled greater than other recalled revealed significant activations in the right middle temporal gyrus that extend posteriorly into the right middle occipital gyrus. This region is almost identical to the region found to be active in the 21 control SPM5 HRF analysis of the contrast between source recalled greater than task recalled. Therefore, it is likely that this region was involved in distinguishing between one‟s own thoughts and actions and those of another person. Notably, the majority of differences found between conditions in the SPM5 analyses in both patients and controls in the internal/external conditions. Only a single significant contrast in any of these analyses involved the source monitoring conditions exclusively. Furthermore, the only SPM5 contrast in which a significant difference between patients and controls was found was in the contrast of internal recalled greater than internal non-recalled. When coupled with the finding that the task monitoring greater than source monitoring contrast did not reveal a significant cluster in any of the analyses, this implies that the lack of differences between patients and controls did not result from a failure to find a difference between the two conditions of experimental interest (self/internal and other/external). Rather, this failure to find a difference seems to have resulted from the same neural network being utilized to perform the task by both  178  groups. Furthermore, the contrasts of self/internal recalled greater than self/internal non-recalled that identified significant clusters of activity were found in both patients and controls only when using an HRF model, and both sets of results point to the mPFC as being critically involved in successful recollection (see Figure 22, Figure 23, Figure 24) of self-generated responses; although it should be noted that, in the schizophrenia patients, significant clusters of activity were also found bilaterally in the posterior cingulate and in the left subgenual area, perhaps indicating the need for greater neural coordination in order to perform the task successfully. In the contrasts of self recalled greater than other recalled, significant clusters of activation were found in both the patients and controls, using both SPM5 HRF and FIR models, which implies that the differences that exist between these two conditions at the neural level are robust with regards to how these events are modeled. In the controls, for both FIR and HRF models, the mPFC is the only region found to be active, although the extent of this region is greater when using an HRF model. It is interesting to note that, in controls, virtually the same region of mPFC is involved both in distinguishing self recalled from other recalled, and in differentiating self recalled from self non-recalled. However, in the patients, different regions are found to be active depending on whether the design matrix uses an HRF or an FIR model. When the HRF model was used to analyze the data, the activations were found in the posterior cingulate whereas when the FIR model was used, the activations were found in the right temporo-parietal junction. Since the data used in the two analyses is exactly the same, this disagreement between the two analyses implies that the differential results are obtained solely as a consequence of how the activity in these regions fits the model. This difference is surprising given that the results from the controls for this contrast are robust with regards to model selection.  179  The contrast of source monitoring recalled greater than task monitoring recalled revealed significant clusters of activation in the controls HRF and patients FIR analyses. The clusters identified as being significantly more active in these analyses are located in different brain regions (right temporo-parietal junction and bilateral precuneus, respectively), yet they both are located in default mode regions (Raichle et al., 2001), and have putatively been argued to support source memory recollection (Schmidt et al., 2002; Platel et al., 2003).  180  Figure 21. Functional brain images and activity time courses from the 21 control 21 patient correlation with behavioural accuracy analysis. CPCA  181  ICA  182  PLS  183  Figure 22. Functional brain images of significant contrasts from the SPM5 - 21 control, 8 condition analysis using an HRF design matrix.  Source recalled greater than task recalled.  Task recalled greater than task non-recalled.  Self recalled greater than other recalled.  Self recalled greater rather self non-recalled.  184  Figure 23. Functional brain images of significant contrasts from the SPM5 - 21 control, 8 condition analysis using an FIR design matrix.  Self recalled greater than other recalled.  185  Figure 24. Functional brain images of significant contrasts from the SPM5 – 21 patient, 8 condition analysis using an HRF design matrix.  Source recalled greater than source non-recalled.  Self recalled greater than other recalled.  Self recalled greater than self non-recalled.  186  Figure 25. Functional brain images of significant contrasts from the SPM5 – 21 patient, 8 condition analysis using an FIR design matrix.  Source recalled greater than task recalled.  Self recalled greater than other recalled.  187  Figure 26. Functional brain images of significant contrasts from the SPM5 – 21 controls and 21 patients, 8 condition FIR analysis.  Controls greater than patients, source recalled greater than source non-recalled.  188  

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0071664/manifest

Comment

Related Items