Human Brain Mapping Foundational Issues in Human Brain Mapping

edited by Stephen José Hanson and Martin Bunzl FOUNDATIONAL ISSUES IN Human Brain Mapping Foundational Issues in Human Brain Mapping Foundational Issues in Human Brain Mapping edited by Stephen Jose´ Hanson and Martin Bunzl A Bradford Book The MIT Press Cambridge, Massachusetts London, England ( 2010 Massachusetts Institute of Technology All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from the publisher. MIT Press books may be purchased at special quantity discounts for business or sales promotional use. For information, please email [email protected] or write to Special Sales Depart- ment, The MIT Press, 55 Hayward Street, Cambridge, MA 02142. This book was set in Stone Serif and Stone Sans on 3B2 by Asco Typesetters, Hong Kong. Printed and bound in the United States of America. Library of Congress Cataloging-in-Publication Data Foundational issues in human brain mapping / edited by Stephen Jose´ Hanson and Martin Bunzl. p. ; cm. ‘‘A Bradford Book.’’ Includes bibliographical references and index. ISBN 978-0-262-01402-1 (hardcover : alk. paper)—ISBN 978-0-262-51394-4 (pbk. : alk. paper) 1. Brain mapping. 2. Brain—Magnetic resonance imaging. I. Hanson, Stephen Jose´. II. Bunzl, Martin. [DNLM: 1. Brain Mapping. 2. Brain—physiology. 3. Data Interpretation, Statistical. 4. Magnetic Resonance Imaging. 5. Research Design. WL 335 F771 2010] QP385.F68 2010 612.8'2—dc22 2009036078 10 9 8 7 6 5 4 3 2 1 Contents Acknowledgments vii Introduction ix I Location and Representation 1 1 A Critique of Functional Localizers 3 Karl J. Friston, Pia Rotshtein, Joy J. Geng, Philipp Sterzer, and Rik N. Henson 2 Divide and Conquer: A Defense of Functional Localizers 25 Rebecca Saxe, Matthew Brett, and Nancy Kanwisher 3 Commentary on Divide and Conquer: A Defense of Functional Localizers 43 Karl J. Friston and Rik N. Henson 4 An Exchange about Localism 49 Martin Bunzl, Stephen Jose´ Hanson, and Russell A. Poldrack 5 Multivariate Pattern Analysis of fMRI Data: High-Dimensional Spaces for Neural and Cognitive Representations 55 James V. Haxby II Inference and New Data Structures 69 6 Begging the Question: The Nonindependence Error in fMRI Data Analysis 71 Edward Vul and Nancy Kanwisher 7 On the Proper Role of Nonindependent ROI Analysis: A Commentary on Vul and Kanwisher 93 Russell A. Poldrack and Jeanette A. Mumford 8 On the Advantages of Not Having to Rely on Multiple Comparison Corrections 97 Edward Vul and Nancy Kanwisher 9 Confirmation, Refutation, and the Evidence of fMRI 99 Christopher Mole and Colin Klein vi Contents 10 Words and Pictures in Reports of fMRI Research 113 Gilbert Harman 11 Discovering How Brains Do Things 115 Stephen Jose´ Hanson and Clark Glymour III Design and the Signal 133 12 Resting-State Brain Connectivity 135 Bharat Biswall 13 Subtraction and Beyond: The Logic of Experimental Designs for Neuroimaging 147 Russell A. Poldrack 14 Advancements in fMRI Methods: What Can They Inform about the Functional Organization of the Human Ventral Stream? 161 Kalanit Grill-Spector 15 Intersubject Variability in fMRI Data: Causes, Consequences, and Related Analysis Strategies 173 Jean-Baptiste Poline, Bertrand Thirion, Alexis Roche, and Se´bastien Meriaux IV The Underdetermination of Theory by Data 193 16 Neuroimaging and Inferential Distance: The Perils of Pictures 195 Adina L. Roskies 17 Brains and Minds: On the Usefulness of Localization Data to Cognitive Psychology 217 Richard Loosemore and Trevor Harley 18 Neuroimaging as a Tool for Functionally Decomposing Cognitive Processes 241 William Bechtel and Richard C. Richardson 19 What Is Functional Neuroimaging For? 263 Max Coltheart References 273 Contributors 309 Index 311 Acknowledgments We gratefully recognize the support of the Office of the Vice President for Academic Affairs at Rutgers and the McDonnell Foundation in providing funds for the meeting that led to this volume, the unflagging patience and enthusiasm of our editor Tom Stone, the valiant and the precise work of Helen Colby in wrangling our contributions into presentable shape. Introduction This is a watershed moment in the field of neuroimaging. At least three emerging trends have been slowly reframing the field for the last five years. Each one of these trends intersects with a foundational issue in human brain mapping, affecting the measurement or methodological or theoretical nature of the overall field. First are methodological problems that have arisen concerning nonindependence of samples—in effect the lack of cross-validation procedures applied in the field, especially in social neuro- science (SN). This has caused a systematic reevaluation of scores of SN studies reporting correlational results literally ‘‘too good to be true,’’ because they weren’t. This type of methodological problem, at first glance, may appear arbitrary and due to an oversight of some sort; however, it actually relates to a larger set of methodological issues that are at the very foundation of the neuroimaging enterprise itself. The overall scale of the typical neuroimaging dataset, the intrinsic noise level, and its intrinsic multivariate nature have all been an enormous challenge in the field. To compound these issues, statisticians early in the enterprise advised those develop- ing new tools (e.g., statistical parametric mapping or SPM) to take a conservative approach to statistical analysis, estimation, and design in neuroimaging. Unfortu- nately, this led researchers to apply the general linear model (GLM) to single voxels and to engage in Neyman-Pearson hypothesis testing. But for what hypothesis? Appar- ently, that each voxel was significantly different from its level in a ‘‘minimally different’’ condition, the so-called Donders paradigm. Unfortunately, this approach suffers from the lack of cross-validation, which promotes the painting of voxels as ‘‘active,’’ ‘‘engaged,’’ or otherwise ‘‘on’’ when they actually may be false alarms. More disturb- ingly, in that brain mapping’s sin qua non is to localize function, many voxels that should be mapped would simply not be mapped. The second trend is, in fact, a direct response to the foundational nature of the neuroimaging data itself. New methods have begun to appear in the human brain mapping field that directly assess the voxel covariation in that they classify conditional on the multivariate patterns of voxels. At a micro-level, these new methods attempt to exploit the obvious neural level of interactions that are contiguous within areas of the x Introduction brain. These methods are based on statistical learning theories that have been devel- oped and applied for more than a decade in the neural information processing field for many kinds of problems and applications. They are standardly cross-validated, non- linear, multivariate, and regularized. In effect, they can provide valid and reliable esti- mates of cortical diagnosticity and reveal maps of the stimulus waveform resulting in new visualizations of cortical structure and similarity. This trend is likely to cause sig- nificant changes in human brain mapping as new questions and answers often result from the second generation of scientific methods that supplant and revisit the foundational questions that ignited the field in the first place. The intersection between the first trend and this second one leads to new concerns about localization and modularity. Is there a ‘‘face area,’’ a ‘‘place area,’’ and so on? Or are these well known results in the brain mapping literature a result of a methodological artifact promoted by features of the standard methods? These newer multivariate classifier methods raise many important questions about the original framing of the brain mapping problem and the nature of the brain response at the spatial level functional magnetic resonance imaging (fMRI) affords us. Wherever these new methods lead us, the human brain mapping field is evolving and is beginning to search for new metaphors, measurement, and data structures. There is a third trend that follows in the footsteps of the second, which relies on new methods from computer science and machine learning, and focuses on a new data structure—the graph. Recent trends have focused on the relationship between regions of interest and their interactivity. The brain is, of course, composed of sets of distinct and overlapping networks somehow creating cognitive and perceptual processing. In the past ten years, interest has accumulated concerning the identification of various brain networks. One of the more commonly encountered networks is the so-called mirror system, which appears to have some functional relationship to the nature of perception action coupling. Other ‘‘social’’ networks seemingly have been identified, including ‘‘mentalizing’’ and ‘‘face recognition’’ as well as ‘‘self’’ or ‘‘intrinsic’’ networks. All this is well and good, except for the fact that data structures more complex than a single region of interest (ROI) or node require search, since even simple networks of four nodes or more have greater than 59,065 possible alternative graph hypotheses. This kind of localization is therefore not possible without some sophisticated search methods. Unfortunately, graph search must be predicated on node localization, which, as we just discussed, is under some revision in terms of methods and tools. Graph search also depends on time series extraction, which with the blood oxygen

Load more