Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Lesion–Symptom Mapping and Neuropsychological Dissociation

Abstract. This fundamental research monograph investigates how patterns of brain injury and behavioral dissociation constrain explanations of human psychological function. It integrates the history of lesion research, single-case inference, double dissociations, voxel-based lesion–symptom mapping, multivariate prediction, imaging registration, anatomical covariance, and white-matter disconnection. Its method is a source-attributed critical narrative synthesis of primary methodological experiments, clinical lesion studies, computational accounts, and network research. The principal finding is that lesions offer valuable perturbational evidence when psychological constructs and anatomical exposures are well measured, while spatially correlated injury, lesion volume, unequal coverage, task impurity, recovery, and predictive-model instability limit unique localization. Comparative evidence supports differentiated yet distributed neural implementation. The concluding synthesis identifies priorities in causal identification, reproducible imaging, cross-cohort validation, longitudinal recovery, and the responsible use of lesion models in clinical research.

Keywords: lesion–symptom mapping; voxel-based lesion–symptom mapping; neuropsychological dissociation; double dissociation; cognitive neuropsychology; single-case inference; multivariate lesion mapping; lesion volume; vascular confounding; lesion normalization; structural disconnection; connectomics; brain–behavior relationships; neural necessity; stroke; causal inference; predictive validation

The Research Object and Its Causal Question

Lesion–symptom mapping examines how spatially characterized brain injury relates to measured changes in cognition, perception, language, affect, and action. Its primary scientific question concerns which neural structures and connections make a functional contribution to performance under specified conditions. The empirical units are people, lesions, measurements, and the statistical relations among them. A lesion may disrupt local tissue, distant physiological activity, connecting pathways, or an entire coordinated system. Accordingly, evidence that damage predicts impairment requires an explanatory account of what was interrupted and how the corresponding psychological construct was measured. The contrast between activation and perturbation establishes the distinctive contribution of lesion studies while retaining their inferential limits. Bates et al. (2003) Rorden et al. (2007) Karnath et al. (2018)

Neuropsychological dissociation is a patterned difference between performances that constrains hypotheses about cognitive architecture. A person may show serious impairment on one test and comparatively preserved performance on another; a second patient may exhibit the reverse profile. Such observations require evaluation against the distribution, reliability, and difficulty of the tests and against the possibility of shared processing resources. The inference may be about functional separability even before precise anatomical location is known. Lesion mapping and dissociation thus provide complementary evidence: one analyzes relations between injured anatomy and behavior across individuals, while the other examines relations between cognitive capacities within and across individuals. Crawford & Garthwaite (2005b) Plaut (1995) Nickels et al. (2011)

The scope of this monograph includes single-case logic, lesion ascertainment, mass-univariate and multivariate inference, vascular and spatial biases, disconnection, prediction, reproducibility, and selected applications to language, attention, movement, and recovery. It treats acute stroke as an important empirical setting rather than a universal model of all lesions. Tumors, traumatic injury, epilepsy surgery, and neurodegeneration generate different processes of injury and adaptation. The study's aim is a critical narrative synthesis of identifiable primary experiments, methodological evaluations, and reviews; it does not present newly collected patient data or a registered systematic review. Mah et al. (2014) Sperber (2020) Grefkes & Fink (2014)

From Classical Localization to Quantitative Lesion Research

The nineteenth-century localization tradition made neurological impairment a means of investigating the organization of mental operations. Clinical observation, postmortem anatomy, and comparisons across patients motivated claims that particular disturbances occurred preferentially after damage to particular regions. Later cognitive neuropsychology refined the object of explanation by distinguishing a clinical syndrome from the operations implicated by a task. This historical development is important because anatomical labels such as frontal, temporal, or parietal can be too broad to specify the representations or computations that actually fail. Modern lesion research inherits the causal ambition of classical localization and subjects it to quantitative lesion coverage, reproducible imaging, and explicit behavioral measurement. Karnath et al. (2018) Nickels et al. (2011)

The case tradition also established a methodological asymmetry. A striking impairment following a small lesion can motivate a hypothesis about an indispensable contribution, but the absence of impairment after another lesion has several interpretations: tissue may have been spared, a task may be insensitive, recovery may have occurred, or other pathways may support the tested operation. Healthy performance is itself multidetermined. Accordingly, one dramatic case is most useful when the test profile is detailed, competing explanations are investigated, and the anatomical record is interpreted alongside converging evidence. Single cases remain valuable for discovering theoretically discriminating patterns, while systematic case series identify their frequency and boundary conditions. Crawford & Garthwaite (2005b) Nickels et al. (2011) Sperber & Karnath (2018)

Computed tomography and magnetic resonance imaging allowed investigators to characterize living patients' lesions more systematically and to compare injuries in common anatomical space. Bates and colleagues' 2003 account of voxel-based lesion–symptom mapping was a prominent methodological transition: behavioral scores could be analyzed voxel by voxel in relation to damage, making localization an explicit statistical problem rather than an impressionistic accumulation of overlapping lesions. This transition introduced greater spatial detail and reproducibility while also exposing dependence on sample coverage, registration, correlated lesions, and choices of statistical test. Quantification did not eliminate the assumptions of classical localization; it made many assumptions measurable and therefore contestable. Bates et al. (2003) Rorden et al. (2009) Nachev et al. (2008)

Cognitive Dissociation as a Theory-Testing Instrument

A single dissociation is typically claimed when one ability is impaired and another relatively spared in a patient with neurological injury. Its strength depends on the expected scores under healthy conditions, the individual's premorbid abilities, and whether measured differences exceed normal variability. Absolute claims of normality should be replaced by normative comparisons supported by appropriate uncertainty. Tasks may draw on unequal attentional resources, differ in sensitivity, or reach ceiling and floor levels at different rates. A selective deficit is therefore an observation that narrows models, not a self-interpreting demonstration of a discrete cognitive organ. Crawford & Garthwaite (2005b) Crawford & Garthwaite (2006) Plaut (1995)

Double dissociation compares two profiles: a patient who is relatively worse at operation A than B, and another who is relatively worse at B than A. The crossed pattern challenges explanations that place all tasks on a single monotonic severity dimension, yet the inference depends on the form of that dimension and on measurement properties. Connectionist simulation has demonstrated that differential vulnerability within a distributed system can generate double dissociations without sharply encapsulated modules. The resulting lesson is methodological: functional differentiation can be established at a level of behavioral organization even when the underlying implementation is graded, interactive, and distributed. The competing computational models should make predictions about errors, generalization, and graded damage. Plaut (1995) Nickels et al. (2011)

The interpretation of dissociations benefits from specifying which contrast is theoretically decisive. Reading regular versus exceptional words, semantic versus phonological naming errors, or spatial exploration versus elementary sensory detection each invokes different hypotheses about processing. An observed interaction between lesion group and task is stronger than separate statements that one group differs significantly from controls and the other does not. Measurement reliability, matched difficulty, and within-person dependence require explicit examination. Cognitive neuropsychology has increasingly used combinations of single cases and case series so that selective patterns can test a model while population variation can constrain its generality. Nickels et al. (2011) Crawford & Garthwaite (2005a) Dell et al. (2007)

Statistical Identification of Single-Case Impairment

Single-patient inference begins with a sampling problem: the control mean and standard deviation are estimates, especially when comparison groups are small. Treating them as fixed population parameters can inflate false-positive decisions. Crawford and Garthwaite evaluated modified tests that account for uncertainty in normative estimates and proposed criteria for assessing the difference between two task scores in the patient relative to the distribution of such differences among controls. This approach recognizes that a difference between raw or standardized task scores is not sufficient unless that difference is unusual for the reference population. Appropriate age, education, language and sensory matching remain important to the validity of comparisons. Crawford & Garthwaite (2005b) Crawford & Garthwaite (2005a)

The covariance of tests is central. A patient with moderately low performance on two highly correlated tasks may exhibit an unusual discrepancy, whereas a visibly different pair of scores on weakly correlated tests may be ordinary. Normative assessment therefore requires the joint distribution of scores, standardized deficits, and a test of their difference. Measurement error affects the power to identify genuine impairment; repeated assessments may expose unstable classifications. Classification thresholds should be reported alongside raw performance and uncertainty rather than substituted for them. The distinction between a classical dissociation and a strong dissociation has statistical content only when both impairment and discrepancy criteria are met and explicitly defined. Crawford & Garthwaite (2005b) Crawford & Garthwaite (2006) Crawford & Garthwaite (2005a)

Clinical measurement introduces further challenges. A motor limitation may depress timed naming even when lexical representations remain accessible; visual field loss can affect apparently conceptual tests; linguistic or cultural background can invalidate a normative threshold. Longitudinal learning, fatigue, and fluctuating attention can change the profile without new damage. A strong case study documents task administration, item properties, response accuracy, speed, errors, test order, and alternative perceptual or motor explanations. When these details are available, the patient becomes an informative constraint on a cognitive model rather than an example chosen only because the impairment appears dramatic. Nickels et al. (2011) Dell et al. (2007) Crawford & Garthwaite (2006)

Lesion Geometry, Imaging, and Registration

Reliable lesion–behavior analysis requires a defensible representation of damaged tissue. Structural magnetic resonance sequences, diffusion-weighted images in acute ischemia, and computed tomography at selected clinical stages provide different contrasts and can delineate different aspects of injury. Lesion masks are drawn or algorithmically segmented and should undergo review for anatomy, missing regions, and partial-volume effects. Differences in timing after stroke matter because edema, tissue evolution, hemorrhagic transformation, and subsequent atrophy change the image. Anatomical interpretation requires documenting injury type, hemisphere, onset-to-scan interval, scanning modality, and the rules for defining an injured voxel. Mah et al. (2014) Nachev et al. (2008) Karnath et al. (2018)

Population analysis requires images in a common coordinate framework, yet the very tissue being aligned may be absent or deformed. Naive nonlinear registration can move tissue unpredictably near an abnormal region, distorting the apparent location of lesion–behavior relations. Nachev and colleagues evaluated enantiomorphic normalization, which uses information from homologous contralesional anatomy to substitute for missing tissue during registration. Their simulations illustrated the advantage of making the abnormal signal tractable rather than asking a registration algorithm to interpret it as ordinary anatomy. Such methods remain contingent on lesion extent, symmetry assumptions, original image quality, and individual neurological anatomy. Nachev et al. (2008)

Lesion coverage is the intersection of available injured and uninjured cases at each location. Areas never damaged in a sampled cohort cannot be tested for necessity by that cohort, and areas damaged in almost everyone cannot support a stable comparison. Spatial smoothing, resampling, thresholding of lesion probability, and minimum-count requirements influence the tested brain space. Maps should present a lesion overlap or coverage image, exclusion masks, and the distribution of total lesion volumes. Without these, an apparently silent region might reflect the absence of informative observations rather than evidence of preserved functional capacity. Image-processing uncertainty belongs in the inferential account rather than the figure legend alone. Bates et al. (2003) Rorden et al. (2009) Sperber & Karnath (2018)

Why Vascular Anatomy Shapes Statistical Inference

Ischemic strokes follow arterial territories, and lesions within a patient are spatially correlated. A behavioral deficit may correlate with injury to several locations simply because one vascular event frequently damages them together. Therefore a peak statistic does not necessarily identify the tissue most directly responsible for the psychological operation. Mah and colleagues analyzed high-dimensional lesion patterns in an acute ischemic cohort and demonstrated that their structure could systematically displace localization results. The phenomenon arises from nonrandom lesion generation, so repeating the analysis in a sample with similar vascular distribution may reproduce the same misleading emphasis. Mah et al. (2014) Sperber (2020)

The distinction between association and intervention matters even when brain injury is an external perturbation. Patients are not randomly assigned to equivalent lesion locations, intensities, or sizes; biological and vascular causes determine the joint exposure. Severity, comorbidity, lesion size, and involvement of correlated structures influence both outcome and sampling. Statistical adjustment is therefore a model of possible confounding, not a guarantee that all causal alternatives have been removed. The inferential target must be stated: localization of a neural contribution, identification of a predictive damage pattern, or comparison of candidate models of a function. These targets may require different analysis designs. Mah et al. (2014) Sperber & Karnath (2018) Sperber (2020)

Some proposed remedies change the unit of analysis from voxels to parcels or networks. Coarser units can reduce dimensionality but may hide vascular dependence within a region of interest, while strongly regularized multivariate methods may redistribute weights across correlated features. Robustness therefore requires simulation with plausible ground truth, alternative lesion distributions, careful accounting for feature collinearity, and independent samples where feasible. A scientific map should show uncertainty and explain the conditions under which its localization would change. The best-supported claim identifies an anatomical contribution with a defined scope rather than treating an isolated statistical coordinate as an anatomical verdict. Mah et al. (2014) Karnath et al. (2018) Sperber (2020)

Voxel-Based Lesion–Symptom Mapping

Voxel-based lesion–symptom mapping evaluates a relation between behavioral performance and lesion status at each adequately sampled voxel. In a basic analysis, patients whose lesion includes a location are compared with patients in whom the same location is spared, or a regression predicts continuous scores from lesion status while incorporating covariates. The resulting map displays evidence of association between localized damage and the outcome under the selected model. This statistical design is especially valuable when a continuous deficit varies across individuals and classic syndrome labels would discard information. It depends on selecting a behaviorally meaningful scale and on demonstrating that damage classification and image alignment are sufficiently reliable. Bates et al. (2003) Rorden et al. (2007) Karnath et al. (2018)

Mass-univariate analyses yield many correlated tests because the brain is spatially extensive. A nominal voxelwise threshold therefore offers inadequate protection against chance findings unless multiplicity is addressed. Permutation tests can construct an empirical null distribution of a maximum statistic or cluster measure under a defensible exchangeability assumption. Corrections also affect sensitivity: a very conservative threshold can erase weaker contributions in sparsely sampled areas. The method section must specify the test, threshold, sidedness, permutation scheme, minimum lesion prevalence, and rule for controlling family-wise error. Selected thresholded maps should be accompanied by continuous effects or test statistics and an account of where inference was possible. Rorden et al. (2009) Rorden et al. (2007)

A significant voxel cannot automatically be called a necessary and sufficient processor of the measured capacity. The score itself may aggregate distinct mechanisms, and a voxel may belong to a tract whose interruption is the operative cause. A region could appear because injury to it is strongly correlated with injury elsewhere. Conversely, two distinct regions might support compensating routes and fail to appear in a simple map despite functional importance. These possibilities make a voxelwise result the beginning of mechanistic interpretation. Replication across different deficit measures, cross-validated analysis, focal cases, and other perturbational methods can distinguish some interpretations. Mah et al. (2014) Thiebaut de Schotten et al. (2020) Sperber (2020)

Behavioral Phenotyping and Construct Validity

Behavioral variables determine what a lesion map can mean. A global impairment score may be statistically reliable yet combine attention, executive demand, language comprehension, motor speed, motivation, and task familiarity. A more focused measure can increase interpretability but may be too narrow to represent functioning outside the testing room. Good phenotyping records item-level errors, accuracy, response latency, and context; it also tests whether separate outcomes measure distinct constructs. The level of anatomical precision cannot compensate for a poorly specified psychological variable. Cognitive neuropsychology has long treated selective error patterns as mechanistic evidence, and case-series modeling extends that principle across individuals. Nickels et al. (2011) Dell et al. (2007) Mirman et al. (2015)

Psychometric quality includes internal consistency, interrater reliability, practice effects, test–retest stability, normative validity, and sensitivity over a relevant severity range. Ceiling effects conceal mild impairment, while floor effects obscure meaningful differences among severe cases. When patients cannot perform tasks for noncognitive reasons, missing data are frequently informative. Researchers should document exclusions and the relation between testing capability and lesion distribution, because selecting only patients who can complete a complex test can distort the sample's anatomical coverage. Scores should be comparable across language and education groups, or analyses must be explicitly restricted to the populations in which the instruments are interpretable. Crawford & Garthwaite (2005b) Crawford & Garthwaite (2006) Sperber (2020)

Construct validation is strengthened by experimental manipulations. A suspected retrieval deficit may change with semantic cueing, stimulus frequency, or phonological complexity, while a suspected spatial-attention deficit changes with visual competition and location. Those interactions allow a theory to specify what information processing was disrupted. A lesion analysis of an experimentally isolated effect can be more informative than a map of undifferentiated diagnosis. Nevertheless, differences across conditions must be assessed at the level of within-patient interactions, using methods that respect the covariance of repeated measures. Deficit specificity and the biological location of its cause remain separate propositions that require independent support. Dell et al. (2007) Mirman et al. (2015) Nickels et al. (2011)

Lesion Volume, Severity, and Covariate Decisions

Large lesions damage more tissue and commonly co-occur with broader impairment. Total lesion volume can therefore predict behavior and correlate with the lesion status of many individual voxels. Omitting volume from analyses may create diffuse, highly significant maps that reflect overall injury burden rather than selective anatomy. DeMarco and Turkeltaub assessed support-vector regression lesion mapping and several correction strategies, finding strong brainwide biases when volume was inadequately handled. The implication is not that a single correction is universally correct; adjusting for volume can also remove genuine effects when lesion extent forms part of the causal pathway. The research question and causal structure must guide the adjustment. DeMarco & Turkeltaub (2018) Sperber (2020)

Covariate control can be implemented by regressing nuisance effects from behavioral scores, from lesion features, or from both, but these operations answer different conditional questions. A parameter chosen only because it sharpens an attractive map risks confirmation bias. Age, time after injury, sex, education, motor disability, clinical severity, scanning protocol, and site may require consideration according to the outcome and design. Some variables measured after stroke are consequences of the lesion and can induce new distortions when controlled improperly. Reporting unadjusted and predeclared adjusted results, with an explicit causal rationale, makes readers able to evaluate the stability and meaning of anatomical associations. DeMarco & Turkeltaub (2018) Sperber & Karnath (2018) Mah et al. (2014)

Because voxel features are correlated, standard regression coefficients can become unstable even when the model predicts behavior well. Different adjustment rules redistribute explanatory weight, while the actual spatial variation of lesions limits which regional effects are identifiable. Sensitivity analyses should vary lesion-volume handling, nuisance specification, coverage thresholds, and imaging quality filters, and should report how the primary inference changes. Where collinearity prevents assigning a unique contribution to nearby structures, the appropriate conclusion is that a combined injury pattern is associated with the deficit. Greater apparent spatial precision from an algorithm must be supported by design and validation, not inferred from the number of voxels examined. Sperber (2020) DeMarco & Turkeltaub (2018) Mah et al. (2014)

Multivariate Lesion–Symptom Models

Multivariate lesion–symptom mapping estimates behavioral outcomes from many lesion features simultaneously. Support vector regression, regularized linear models, and other predictive algorithms can represent spatially distributed combinations of damage that a series of independent voxel tests may inadequately capture. Their target may be prediction, explanatory feature importance, or a representation of interacting injuries; those goals require separate validation. A model can improve the prediction of performance by exploiting correlations in the dataset without proving that every high-weight location is functionally indispensable. Consequently, model coefficients, predictive accuracy, and causal contribution should not be treated as interchangeable results. Karnath et al. (2018) DeMarco & Turkeltaub (2018) Sperber (2020)

Regularization reduces variance in a high-dimensional setting, but it also selects among correlated features according to optimization constraints. Two adjacent white-matter locations may convey almost identical injury information; a sparse model may retain one, and a different sample may retain the other. Multicollinearity consequently makes spatial explanation unstable unless feature-selection uncertainty is evaluated. Mah and colleagues' work showed how dependence among lesion patterns can undermine apparently precise inferences, while subsequent methodological accounts stress that multivariate approaches require an equally careful interpretation of their spatial solutions. Feature-level claims should be checked against plausible alternate models and independent evidence about anatomy. Mah et al. (2014) Sperber (2020) Karnath et al. (2018)

The scientific value of a multivariate model increases when its design is reproducible. Investigators should specify input masks, anatomical parcellations, scaling, missing-data handling, regularization, tuning strategy, permutation testing, nuisance covariates, and statistical uncertainty. Train-only preprocessing must be repeated within every cross-validation fold; otherwise information from held-out patients can leak into features or parameter choices. If lesion volume is a strong baseline predictor, multivariate performance should be compared with that baseline. An accurate model can justify further clinical investigation even when its spatial weights remain uncertain; its explanatory interpretation requires additional experiments and cross-population testing. DeMarco & Turkeltaub (2018) Teghipco et al. (2024) Hope et al. (2013)

Cross-Validation, Stability Selection, and External Generalization

Performance estimated on the same participants used to select features can be optimistic. Nested cross-validation separates inner model tuning from outer testing, while external validation measures transport to a genuinely distinct cohort. These steps are especially consequential in lesion data because the number of candidate spatial predictors can exceed the number of patients by several orders of magnitude. Stratification must account for correlated observations, repeated scans of the same individual, and center-specific acquisition patterns. A study should report predictive error, uncertainty, calibration where relevant, and comparison with simple baselines rather than interpreting a single correlation between predictions and observed outcomes as comprehensive validation. Teghipco et al. (2024) Hope et al. (2013)

Teghipco and colleagues' 2024 study introduced stable multivariate lesion–symptom mapping, using feature reliability and stability selection within a predictive framework. Its reported analyses involved chronic and independent acute-stroke data and emphasized more consistent feature identification and prediction of aphasia severity compared with conventional approaches. The methodological advance is important because feature selection and model training can be evaluated jointly. However, the method's advantages must be interpreted relative to the studied outcomes, anatomical distributions, sample sizes, and tuning choices. Stability across resampling of one cohort is not the same as anatomical truth or external causal confirmation. Teghipco et al. (2024)

Temporal and geographic transport are distinct demands. A model trained on a single tertiary rehabilitation center may encounter different lesion sizes, populations, imaging practices, and severity distributions in a population-based acute-stroke service. Predictive success may also depend on the clinical stage: early deficits reflect acute physiology and evolving injury, whereas chronic performance reflects adaptation, compensatory strategies, and rehabilitation history. External evaluation should state whether both input imaging and outcomes are gathered comparably across cohorts. A well-calibrated predictive model in one setting can still be unsafe or ineffective for individual prognostic decisions in another. Hope et al. (2013) Grefkes & Fink (2014) Teghipco et al. (2024)

Sampling, Missingness, and Reproducible Research

Sampling is a hidden determinant of lesion maps. Stroke services recruit different distributions of vascular injury from neurosurgical cohorts, and rehabilitation studies often oversample survivors who are available for extended cognitive testing. Aphasia, neglect, weakness, visual-field loss, or fatigue can prevent participation in assessments most relevant to a lesion analysis. This creates selection mechanisms linked to both anatomy and measured behavior. The resulting inference pertains to the enrolled population, not automatically to every person with neurological injury. Demographic composition, inclusion criteria, mortality, severe-case exclusion, and completion rates must be documented before interpreting generalizability. Sperber (2020) Mah et al. (2014) Karnath et al. (2018)

Statistical power varies spatially because lesion prevalence and functional effect sizes differ across brain locations. A nominal sample of many patients may provide weak information about rarely injured structures. The number of independent lesion patterns matters as much as the total number of voxel observations. Prospective power planning should use realistic lesion distributions and simulated outcomes when possible, followed by transparent reporting of which voxels or parcels were statistically testable. Multiple outcomes and numerous alternative mapping pipelines inflate the analytic search space; preregistered primary contrasts and disclosure of exploratory analyses reduce the risk of selective interpretation. Rorden et al. (2009) Mah et al. (2014) Sperber & Karnath (2018)

Reproducibility includes more than making a thresholded map available. It requires access, when consent and privacy permit, to lesion masks, preprocessing decisions, behavioral dictionaries, analysis code, library versions, parcellation files, and the exact statistical contrasts. Deidentified images can still be sensitive, so governance and controlled sharing may be necessary. Results should be checked using independent segmentation review and alternative analytic choices rather than only rerunning the original pipeline. When raw data cannot be released, sufficient simulated datasets, code, parameter files, and aggregate coverage maps can make the inferential procedure auditable. The strongest methodological result is one whose limitations and alternative explanations remain visible. Nachev et al. (2008) DeMarco & Turkeltaub (2018) Sperber (2020)

White-Matter Disconnection as an Explanatory Mechanism

An infarct confined to a relatively small region can impair performance across functions whose participating cortical regions remain structurally intact. The reason may be interruption of the pathways that allow information to travel between those regions. A lesion is therefore both a local injury and a potential disconnection event. Classical disconnection reasoning anticipated this distinction, and modern tractography-based analysis gives it a measurable expression. Foulon and colleagues developed tools for anatomical lesion and tract-level mapping; Thiebaut de Schotten and colleagues connected structural disruption to functional and behavioral consequences. These approaches expand the explanatory space beyond gray-matter lesion centers. Foulon et al. (2018) Thiebaut de Schotten et al. (2020)

Disconnectome analysis projects an observed lesion onto an estimate of white-matter trajectories or the structural connections between defined parcels. It can estimate which routes are likely interrupted even if portions of those routes are remote from the visibly injured tissue. Some methods use a population-based tractography atlas, whereas patient-specific diffusion imaging may capture individual variation under different acquisition constraints. Both involve uncertainty because tractography approximates pathways, crossing fibers can be challenging, and premorbid anatomy is generally unavailable. Network interruption should be treated as a probabilistic explanatory feature, accompanied by sensitivity analyses for parcellation and atlas choice. Foulon et al. (2018) Thiebaut de Schotten et al. (2020)

Anatomical disconnection can affect transmission, timing, and functional coupling without producing identical symptoms in all patients. The consequences depend on which operations were being coordinated, the presence of alternative routes, the time since injury, and the current task. A critical methodological issue is distinguishing direct tract interruption from correlated cortical injury: lesions following arterial territories often affect both. Comparative mapping that includes local tissue damage and disconnection measures can test incremental explanatory power. Evidence that connectivity predicts outcome more accurately supports a network model, while specific claims about causal information flow require additional physiological and intervention evidence. Siegel et al. (2016) Thiebaut de Schotten et al. (2020) Grefkes & Fink (2014)

Structural and Functional Network Perturbation

Brain networks are described by anatomical edges, physiological coupling, and task-dependent coordinated activity. These quantities are related but not identical. Structural connections constrain communication, whereas functional correlations also reflect shared inputs, state, and indirect interactions. Stroke can change both local activity and long-range coupling, with consequences well beyond the direct lesion. Siegel and colleagues reported that disruption of network connectivity predicted impairments across multiple behavioral domains after stroke. Their results demonstrate the utility of systems-level explanatory variables while leaving open which specific physiological alterations mediate each observed deficit. Siegel et al. (2016) Thiebaut de Schotten et al. (2020)

Network models can explain why ostensibly comparable lesions yield different profiles. A lesion's influence may depend on whether it interrupts a highly central pathway, whether other regions can compensate, and whether the affected operation engages multiple networks. Conversely, different lesion locations may converge on the same functional disconnection and therefore cause similar syndromes. These hypotheses require careful modeling of network definitions and comparison with the behavioral data. The language of hubs, edges, modularity, and global efficiency is descriptive until a demonstrated association with a specified psychological operation is established. Network topology alone cannot substitute for construct-valid testing. Grefkes & Fink (2014) Foulon et al. (2018) Thiebaut de Schotten et al. (2020)

Functional connectivity after injury is also affected by alertness, head motion, vascular response, medication, and time since stroke. A low correlation between regions can arise from altered neural coupling, physiological noise, or measurement differences. Structural lesions often distort the normal assumptions used in functional imaging, so quality assurance and physiological controls are indispensable. Investigators should describe the temporal sequence of imaging and behavioral assessment and the stability of results under alternative parcellations. The strongest explanatory synthesis connects structural damage, remote physiological change, and a functionally discriminating behavioral outcome within one coordinated design. Siegel et al. (2016) Grefkes & Fink (2014) Karnath et al. (2018)

The Language System as a Test Case

Language disorders have been central to the development of lesion–symptom mapping because language yields measurable contrasts in comprehension, naming, repetition, fluency, and semantic control. Bates and colleagues' early voxel-based analysis illustrated how graded language performance could be related to the distribution of damaged tissue in aphasic patients. Later studies such as Mirman and colleagues' investigation of spoken language incorporated multiple behavioral dimensions to examine the organization of neural language systems. These studies move beyond a single aphasia label toward a matrix of operations and lesion patterns, while remaining sensitive to correlated injuries in the left perisylvian vascular territory. Bates et al. (2003) Mirman et al. (2015)

One person's poor picture naming can arise from difficulty retrieving a lexical item, producing its phonological form, processing the picture, or sustaining task goals. A naming score must therefore be interpreted alongside comprehension, repetition, phonological judgments, and the structure of naming errors. Dell, Martin, and Schwartz tested an interactive model of lexical access through a case series that compared naming and repetition errors, illustrating the value of computationally explicit behavioral models. Such models constrain what a neural lesion map is permitted to claim about language organization. A region correlated with naming performance need not be dedicated uniquely to naming or even to one stage of lexical access. Dell et al. (2007) Nickels et al. (2011) Mirman et al. (2015)

Language evidence is particularly informative when behavioral factors are separable and statistically contrasted, yet even distinct factors may depend on overlapping white-matter routes. The analysis should evaluate whether a lesion variable predicts one component after adjusting for correlated behavior, and whether that adjustment is theoretically warranted. Patient grouping by chronic severity may amplify selection effects because patients with severe symptoms can be less able to complete complex tasks. Emerging multivariate models offer improved prediction of aphasia severity in tested cohorts, but their performance does not resolve every question about language representations or causal neural necessity. Teghipco et al. (2024) DeMarco & Turkeltaub (2018) Mirman et al. (2015)

Attention, Neglect, and Spatially Structured Deficits

Hemispatial neglect provides an instructive case in which a neuropsychological deficit is distributed across tasks and anatomical sites. Clinical assessments may include cancellation, line bisection, exploration, target detection, and functional behavior; each recruits sensory, motor, attentional, and executive processes in different proportions. A lesion–symptom map based on one task can thus emphasize different structures from a map based on another. The relevant construct is the systematic failure to orient or respond appropriately to portions of space under specified conditions. Interpretation must distinguish this from elementary sensory loss or limb weakness and must account for severity and task demands. Karnath et al. (2018) Sperber & Karnath (2018)

Vascular lesion covariance is especially consequential in spatial disorders because large lesions can encompass cortical attention regions and connecting pathways. A lesion overlap maximum among patients with neglect may describe the typical location of a vascular event rather than the unique site of the disrupted operation. Mah and colleagues' high-dimensional critique applies to precisely such anatomical dependence. Combining voxelwise comparisons with multivariate models, lesion coverage, and network disconnection measures provides stronger tests, though none alone abolishes the correlation problem. Anatomical interpretation should be expressed as a spatially distributed pattern with uncertainty when unique localization is statistically unidentifiable. Mah et al. (2014) Sperber (2020) Thiebaut de Schotten et al. (2020)

A demanding neuropsychological comparison asks whether two spatial behaviors dissociate after matching stimulus conditions and considering common motor outputs. An apparent improvement under cueing may indicate residual processing capacity rather than the restoration of a destroyed subsystem. Longitudinal assessment adds an additional dimension because changes in arousal, strategy, and recovery can modify observed neglect. These factors support a system-level account in which attention is dynamically realized by interacting pathways. Lesion-based causal reasoning remains important, but the appropriate target is a task-conditioned neural contribution rather than a timeless one-region label. Grefkes & Fink (2014) Karnath et al. (2018) Nickels et al. (2011)

Motor Impairment as a Methodological Control Case

Motor deficits make the contrast between anatomical specificity and statistical association especially clear. Damage to descending pathways may correlate strongly with limb impairment, but a large middle cerebral artery stroke can also injure neighboring tissue that does not independently control the tested movement. Rorden, Fridriksson, and Karnath compared alternative lesion-analysis methods using hemiparesis and demonstrated that a method with high statistical sensitivity could also yield widespread associations in tissue not specific to the motor system. Their work illustrates why comparison-group selection and test statistics are part of scientific interpretation rather than neutral implementation details. Rorden et al. (2009)

A motor outcome appears objective when measured by strength or movement speed, yet the test may also depend on comprehension of instructions, sensory feedback, pain, premorbid mobility, and fatigue. These factors become especially relevant when motor performance is used as a covariate in language or attentional mapping. A single broad impairment scale can suppress differential processing profiles that would otherwise help identify mechanisms. Similarly, lesion-volume correction and behavioral adjustment may have different consequences in a motor task than in an abstract memory task. Careful definition of outcome and competing functional pathways is essential even for seemingly straightforward neurological symptoms. Rorden et al. (2007) DeMarco & Turkeltaub (2018) Karnath et al. (2018)

Motor recovery also underscores the temporal character of lesion evidence. Structural damage may persist while performance improves through reperfusion, reorganization, compensatory movements, and rehabilitation. An early anatomical deficit association can therefore differ from a chronic prediction model. Functional connectivity analysis and longitudinal behavioral testing investigate which preserved circuits support regained performance and how damage constrains those possibilities. The distinction between the substrate necessary for normal pre-injury performance and the substrate supporting compensated post-injury behavior should be stated explicitly, because chronic lesion cohorts observe both residual deficits and consequences of adaptation. Grefkes & Fink (2014) Hope et al. (2013) Siegel et al. (2016)

Recovery, Longitudinal Change, and Clinical Translation

Prediction after neurological injury differs from explanation of normal cognition. A prognostic model estimates future functioning conditional on observed lesions and clinical characteristics, whereas a mechanistic account identifies how damage alters neural operations. Hope and colleagues examined the prediction of stroke outcome and recovery using lesion information from magnetic resonance images, exemplifying the clinical interest in linking structural injury to later performance. However, recovery is determined by multiple factors, including initial severity, time after injury, vascular physiology, reserve, rehabilitation exposure, and social opportunity. A model that predicts one endpoint accurately can still offer incomplete explanation of the process producing it. Hope et al. (2013) Grefkes & Fink (2014)

Longitudinal research should establish when tissue and behavior were measured, how outcome instruments changed with severity, and whether missing follow-up data were associated with initial disability. Statistical adjustment for baseline performance can sharpen prediction but may also alter the target from total outcome to residual change. Learning effects and alternative task strategies may look like recovery of the original mechanism although the neural implementation has changed. When possible, repeated network and behavioral measurements can distinguish physiological normalization, compensatory recruitment, and changed task strategy. These distinctions matter for designing interventions and for interpreting apparent neuropsychological dissociations over time. Grefkes & Fink (2014) Siegel et al. (2016) Hope et al. (2013)

Clinical use demands evidence beyond statistical significance or improved within-sample prediction. Prognostic models should undergo independent validation, calibration analysis, and assessment of whether they change decisions or patient outcomes. Variables must be available at the relevant time and measured in a reproducible clinical workflow. Uncertainty is especially important when predictions concern individual people; a group-average relation can perform poorly at the extremes of the severity distribution. Lesion–symptom research can inform assessment and rehabilitation science without licensing individual diagnoses or treatment decisions directly from a research map. Hope et al. (2013) Teghipco et al. (2024) Sperber (2020)

Causal Interpretation and the Problem of Neural Necessity

A lesion can provide stronger evidence of a causal contribution than observational task activation, because injury changes the system whose operation is being studied. Yet the cause is an event with many components: tissue damage, vascular disturbance, disconnection, acute physiology, and adaptation. Whether a region is necessary depends on the task, time, population, and remaining circuitry. It is more precise to infer that injury affecting a specified anatomical pattern increases the probability or severity of a defined impairment under observed conditions than to assert that one voxel uniquely contains a mental faculty. Localization hypotheses acquire causal strength through competing predictions, targeted evidence, and independent convergence. Rorden et al. (2007) Sperber (2020) Thiebaut de Schotten et al. (2020)

Potential outcomes reasoning helps clarify why nonrandom lesions complicate the inference. The same patient cannot be observed simultaneously with and without an equivalent injury at a chosen location; researchers compare different people or different phases of recovery. Vascular anatomy determines which combinations of locations can be damaged together. Age, health, and survival influence which injuries are observed. These features limit the identification of a unique causal effect even when the lesion itself was an unplanned biological perturbation. Statistical methods provide conditional estimates under assumptions about exchangeability, measurement, and selection. A causal diagram can make those assumptions visible for evaluation. Mah et al. (2014) Sperber (2020) Sperber & Karnath (2018)

Zavaglia and colleagues applied lesion-based modeling to clinical assessments of damage after stroke, illustrating how anatomical disruption can be translated into functional predictions. The general challenge is to distinguish a structure participating in a task, a structure whose injury is predictive of poor performance, and a structure whose intact contribution is necessary under a specified functional arrangement. These three claims can overlap without becoming equivalent. Strong inference requires convergent behavioral specificity, anatomical evidence, perturbation models, and tests of alternative functional architectures. This plural evidentiary structure preserves the causal value of lesions while avoiding a simplistic map of independent psychological organs. Zavaglia et al. (2015) Karnath et al. (2018) Plaut (1995)

Distributed Models, Modularity, and Mechanistic Explanation

Modular descriptions treat some cognitive operations as relatively separable, which makes a dissociation theoretically informative. Distributed neural models instead represent operations through patterns of interacting units, graded specialization, and shared resources. These descriptions operate at different analytical levels. Plaut's connectionist analysis demonstrated how selective vulnerability and even double dissociations can emerge in a network that lacks the discrete, isolated components inferred by a simple modular reading. The existence of a double dissociation therefore motivates explanation of functional differentiation but does not choose a unique implementation architecture. Plaut (1995) Nickels et al. (2011)

A computational model becomes a test of neuropsychological evidence when it predicts error distributions, lesion sensitivity, and cross-task generalization. Simulated lesions allow investigators to vary the location, scale, and severity of disruption while holding a model's learned history constant. Human studies cannot usually manipulate these factors independently. At the same time, successful simulation establishes that a proposed architecture is capable of reproducing a behavioral profile, rather than demonstrating that it is the architecture realized in a particular patient's brain. Convergent anatomical data and differentiated behavioral measurements are needed to adjudicate between models that can each generate the same dissociation. Plaut (1995) Dell et al. (2007) Nickels et al. (2011)

Network accounts of stroke add a spatial and physiological dimension to cognitive models. Structural lesions may disconnect distributed processors, while preserved pathways undergo altered coordination. A network explanation must still specify a computation; simply describing distant activity or reduced connectivity is insufficient. The strongest models map task manipulations onto particular information-processing demands and predict how localized or distributed injury changes those demands. This combination of cognitive modeling, lesion anatomy, and network dynamics offers a research program in which anatomical and psychological descriptions constrain one another. It also makes room for individual variation and adaptation without treating every unique profile as a separate mental module. Thiebaut de Schotten et al. (2020) Siegel et al. (2016) Grefkes & Fink (2014)

Comparing Evidence Across Lesion Etiologies and Populations

Ischemic stroke, hemorrhage, tumor resection, penetrating injury, traumatic brain injury, and neurodegeneration differ in the distribution and temporal evolution of damage. Stroke samples provide a natural setting for common-space mapping but reflect vascular territories. Tumors may displace tissue and generate gradual adaptation; traumatic injury can combine focal lesions with diffuse damage; progressive disease changes networks over long periods. These mechanisms challenge direct aggregation into a single lesion map. Comparative claims should identify etiology, imaging time, spatial precision, disease stage, and mechanisms of selection before any cross-study synthesis is attempted. Mah et al. (2014) Karnath et al. (2018) Grefkes & Fink (2014)

Age, handedness, multilingual experience, premorbid education, cognitive reserve, comorbid disease, and access to rehabilitation may shape behavioral outcomes independently of lesion coordinates. Some factors influence task interpretation and others alter the likelihood of entering the study. Functional anatomy can be stable in broad organization while differing in network efficiency and compensatory capacity. Generalization should therefore be evaluated separately for behavioral measures, statistical associations, and mechanistic explanations. A model's success within an English-language chronic aphasia cohort does not automatically establish comparable prediction for an acute multilingual population, even when lesions are mapped into the same template. Teghipco et al. (2024) Mirman et al. (2015) Sperber (2020)

Ethical scientific reporting is necessary because individual lesion masks can be identifying when linked with clinical histories, rare syndromes, or longitudinal imaging. Open methods and repeatable analyses should be pursued within legitimate consent, institutional review, and privacy controls. Accessible behavioral testing is equally important: excluding people with severe communication or motor deficits may systematically remove the cases most informative about the underlying system. Study reports should make recruitment constraints visible and avoid treating absence of enrollment as absence of impairment. Better sampling and explicit limits of transportability are substantive improvements in the science, not merely administrative enhancements. Crawford & Garthwaite (2005b) Sperber & Karnath (2018) Teghipco et al. (2024)

Critical Appraisal of the Evidence Base

The strongest evidence for voxel-based mapping is methodological and observational: injury distributions and behavioral outcomes can be collected, statistically associated, and reproduced under controlled analytical conditions. Bates and colleagues established an influential voxelwise approach; Rorden and colleagues examined its inferential alternatives; Mah and colleagues exposed important spatial biases; later studies developed multivariate and disconnection analyses. Each advance also changes the interpretation of prior maps. The accumulating literature supports lesion methods as powerful constraints on neurocognitive theory, provided lesion distributions, behavioral construct validity, model selection, and uncertainty are treated as parts of the scientific result. Bates et al. (2003) Rorden et al. (2009) Mah et al. (2014)

Evidence supporting a specific causal localization is generally strongest when a proposed anatomical relation survives variation in segmentation, coverage, correction, covariates, and population, and when convergent methods predict a differentiated behavioral effect. A statistically corrected map from one clinical sample offers weaker grounds for universal necessity than a replicated relation combined with focal cases, network analysis, and plausible mechanistic models. Large datasets improve estimation but cannot alone cure systematic selection or vascular dependence. Prediction may become more stable as multivariate methods improve, while spatial feature interpretation may remain sensitive to correlated damage. These are distinct dimensions of evidence quality. Sperber (2020) DeMarco & Turkeltaub (2018) Teghipco et al. (2024)

Contradictory findings should be interpreted in light of measurement and design before becoming claims about biological disagreement. Two studies may use different task versions, damage windows, normalization methods, patient-selection rules, and contrasts while applying the same syndrome name. Network analyses may favor white-matter interruption where a cortical lesion map highlights neighboring gray matter, and both may be statistically dependent on a common lesion territory. A productive critical synthesis therefore documents uncertainty rather than averaging all reported coordinates into a supposedly definitive center. Sensitivity, sampling, convergent evidence, and explicit hypotheses determine how strongly a result should shape cognitive theory or clinical interpretation. Karnath et al. (2018) Foulon et al. (2018) Sperber & Karnath (2018)

Open Questions and a Program for Future Investigation

The next stage of lesion–symptom research requires identifying which aspects of spatial dependence can be resolved with realistic datasets. Larger cohorts and more heterogeneous etiologies may improve coverage, but investigators must also develop methods that communicate when lesion collinearity makes several explanations observationally equivalent. Stability selection, multi-model ensembles, and uncertainty maps offer promising tools for describing that ambiguity. External comparisons across acquisition sites, disease stages, and populations will establish which neural associations are transportable. A result should be judged by how well its assumptions and predictions survive out-of-sample tests, rather than by how focal the displayed image appears. Teghipco et al. (2024) Sperber (2020) Mah et al. (2014)

A second priority concerns the connection between psychological constructs and causal neural descriptions. Carefully designed tasks, item-level measurement, error analysis, and experiments that vary cognitive demand can distinguish multiple explanations for the same raw deficit score. Mechanistically specified computational models should predict selective vulnerability, graded impairment, and recovery trajectories. Joint analysis of lesion anatomy, inferred structural disconnection, functional connectivity, and behavioral interaction may then test those predictions. Such integration must retain distinct validation standards for each modality; an accurate network predictor alone cannot prove the functional necessity of a connection. Dell et al. (2007) Siegel et al. (2016) Thiebaut de Schotten et al. (2020)

Clinical translation requires prospective evaluations of whether a lesion-based model changes assessment, prognosis, or rehabilitation outcomes. Models should include uncertainty for individual decisions, evaluate external calibration, and compare added value with inexpensive clinical baselines. More equitable recruitment can also test whether reported dissociations and anatomical relations hold across language, disability, and social context. Finally, reproducible simulated lesions and computational damage experiments can clarify the consequences of alternative cognitive architectures. The central research agenda is a coordinated improvement of causal identification, psychological measurement, network modeling, and responsible prediction. Its success will be measured in explanations that generalize and remain open to experimental challenge. Hope et al. (2013) Plaut (1995) Teghipco et al. (2024)

Conclusion

Lesion–symptom mapping and neuropsychological dissociation remain foundational methods for relating brain injury to the organization of human psychological functions. Lesion studies establish experimentally relevant constraints because interruption of neural structures changes performance, while dissociations distinguish the functional demands on which successful performance depends. The value of each method is greatest when the psychological construct, the anatomical exposure, and the comparison population are defined precisely. The historical progression from case reports to voxelwise statistics, multivariate analysis, and disconnectome modeling has improved precision and made long-standing inferential vulnerabilities visible. Bates et al. (2003) Crawford & Garthwaite (2005b) Karnath et al. (2018)

Scientific synthesis favors a distributed and differentiated account of neural implementation. Local tissue damage, interruption of anatomical pathways, changes in physiological coupling, and postinjury adaptation can all influence an observed symptom. A behavioral dissociation constrains cognitive architecture without necessarily proving strict modularity; a statistical lesion peak identifies an association without independently proving singular causal necessity. Vascular covariation, uneven coverage, volume effects, measurement error, and sample selection limit the strength of localization claims. Contemporary multivariate and network approaches address portions of these challenges while adding their own requirements for validation, feature stability, and explanatory discipline. Plaut (1995) Mah et al. (2014) Thiebaut de Schotten et al. (2020) Teghipco et al. (2024)

The resulting methodological principle is that neural and psychological explanations must be tested together. Well-defined tasks provide information about cognitive operations, carefully characterized lesions establish perturbational constraints, and validated statistical models reveal the conditions under which anatomical variation predicts impairment or recovery. Independent replications, sensitivity analyses, longitudinal data, and convergent structural and functional evidence strengthen conclusions. This framework preserves lesion science as a rigorous route to understanding human cognition while recognizing the complexity of distributed systems and the practical limits of clinical inference. Future progress will depend on better causal identification, reliable measurement, open analytical procedures, and generalization beyond the original cohorts. Sperber (2020) Nickels et al. (2011) Grefkes & Fink (2014)

References

  • Bates, E., Wilson, S. M., Saygin, A. P., Dick, F., Sereno, M. I., Knight, R. T., & Dronkers, N. F. (2003). Voxel-based lesion–symptom mapping. Nature Neuroscience, 6(5), 448–450. https://doi.org/10.1038/nn1050
  • Crawford, J. R., & Garthwaite, P. H. (2005). Evaluation of criteria for classical dissociations in single-case studies by Monte Carlo simulation. Neuropsychology, 19(5), 664–678. https://doi.org/10.1037/0894-4105.19.5.664
  • Crawford, J. R., & Garthwaite, P. H. (2005). Testing for suspected impairments and dissociations in single-case studies in neuropsychology: Evaluation of alternatives using Monte Carlo simulations and revised tests for dissociations. Neuropsychology, 19(3), 318–331. https://doi.org/10.1037/0894-4105.19.3.318
  • Crawford, J. R., & Garthwaite, P. H. (2006). Detecting dissociations in single-case studies: Type I errors, statistical power and the classical versus strong distinction. Neuropsychologia, 44(12), 2249–2258. https://doi.org/10.1016/j.neuropsychologia.2006.05.019
  • Dell, G. S., Martin, N., & Schwartz, M. F. (2007). A case-series test of the interactive two-step model of lexical access: Predicting word repetition from picture naming. Journal of Memory and Language, 56(4), 490–520. https://doi.org/10.1016/j.jml.2006.05.007
  • DeMarco, A. T., & Turkeltaub, P. E. (2018). A multivariate lesion symptom mapping toolbox and examination of lesion-volume biases and correction methods in lesion-symptom mapping. Human Brain Mapping, 39(11), 4169–4182. https://doi.org/10.1002/hbm.24289
  • Foulon, C., Cerliani, L., Kinkingnéhun, S., Levy, R., Rosso, C., Urbanski, M., Volle, E., & Thiebaut de Schotten, M. (2018). Advanced lesion symptom mapping analyses and implementation as BCBtoolkit. GigaScience, 7(3), giy004. https://doi.org/10.1093/gigascience/giy004
  • Grefkes, C., & Fink, G. R. (2014). Connectivity-based approaches in stroke and recovery of function. The Lancet Neurology, 13(2), 206–216. https://doi.org/10.1016/S1474-4422(13)70264-3
  • Hope, T. M. H., Seghier, M. L., Leff, A. P., & Price, C. J. (2013). Predicting outcome and recovery after stroke with lesions extracted from MRI images. NeuroImage: Clinical, 2, 424–433. https://doi.org/10.1016/j.nicl.2013.03.005
  • Karnath, H.-O., Sperber, C., & Rorden, C. (2018). Mapping human brain lesions and their functional consequences. NeuroImage, 165, 180–189. https://doi.org/10.1016/j.neuroimage.2017.10.028
  • Mah, Y.-H., Husain, M., Rees, G., & Nachev, P. (2014). Human brain lesion-deficit inference remapped. Brain, 137(9), 2522–2531. https://doi.org/10.1093/brain/awu164
  • Mirman, D., Chen, Q., Zhang, Y., Wang, Z., Faseyitan, O. K., Coslett, H. B., & Schwartz, M. F. (2015). Neural organization of spoken language revealed by lesion-symptom mapping. Nature Communications, 6, 6762. https://doi.org/10.1038/ncomms7762
  • Nachev, P., Coulthard, E., Jäger, H. R., Kennard, C., & Husain, M. (2008). Enantiomorphic normalization of focally lesioned brains. NeuroImage, 39(3), 1215–1226. https://doi.org/10.1016/j.neuroimage.2007.10.002
  • Nickels, L., Howard, D., & Best, W. (2011). On the use of different methodologies in cognitive neuropsychology: Drink deep and from several sources. Cognitive Neuropsychology, 28(7), 475–485. https://doi.org/10.1080/02643294.2012.672406
  • Plaut, D. C. (1995). Double dissociation without modularity: Evidence from connectionist neuropsychology. Journal of Clinical and Experimental Neuropsychology, 17(2), 291–321. https://doi.org/10.1080/01688639508405124
  • Rorden, C., Karnath, H.-O., & Bonilha, L. (2007). Improving lesion-symptom mapping. Journal of Cognitive Neuroscience, 19(7), 1081–1088. https://doi.org/10.1162/jocn.2007.19.7.1081
  • Rorden, C., Fridriksson, J., & Karnath, H.-O. (2009). An evaluation of traditional and novel tools for lesion behavior mapping. NeuroImage, 44(4), 1355–1362. https://doi.org/10.1016/j.neuroimage.2008.09.031
  • Siegel, J. S., Ramsey, L. E., Snyder, A. Z., Metcalf, N. V., Chacko, R. V., Weinberger, K., Baldassarre, A., Hacker, C. D., Shulman, G. L., & Corbetta, M. (2016). Disruptions of network connectivity predict impairment in multiple behavioral domains after stroke. Proceedings of the National Academy of Sciences, 113(30), E4367–E4376. https://doi.org/10.1073/pnas.1521083113
  • Sperber, C., & Karnath, H.-O. (2018). On the validity of lesion-behaviour mapping methods. Neuropsychologia, 115, 17–24. https://doi.org/10.1016/j.neuropsychologia.2017.07.035
  • Sperber, C. (2020). Rethinking causality and data complexity in brain lesion-behaviour inference and its implications for lesion-behaviour modelling. Cortex, 126, 49–62. https://doi.org/10.1016/j.cortex.2020.01.004
  • Teghipco, A., Newman-Norlund, R., Gibson, M., Bonilha, L., Absher, J., Fridriksson, J., & Rorden, C. (2024). Stable multivariate lesion symptom mapping. Aperture Neuro, 4, Article 117311. https://doi.org/10.52294/001c.117311
  • Thiebaut de Schotten, M., Foulon, C., & Nachev, P. (2020). Brain disconnections link structural connectivity with function and behaviour. Nature Communications, 11, 5094. https://doi.org/10.1038/s41467-020-18920-9
  • Zavaglia, M., Forkert, N. D., Cheng, B., Gerloff, C., Thomalla, G., & Hilgetag, C. C. (2015). Mapping causal functional contributions derived from the clinical assessment of brain damage after stroke. NeuroImage: Clinical, 9, 83–94. https://doi.org/10.1016/j.nicl.2015.07.009