Combating Motion Artifact: A Researcher's Guide to Detecting and Preventing Spurious Brain-Behavior Associations

Abstract

In-scanner head motion is a pervasive source of spurious findings in brain-behavior association studies, posing a significant threat to the validity of neuroimaging research and its translation to drug development. This article provides a comprehensive guide for researchers and scientists on the latest methodologies for quantifying and mitigating motion-related bias. Drawing on recent advances, including the novel SHAMAN framework and findings from large-scale studies like ABCD, we cover the foundational impact of motion, introduce trait-specific motion impact scores, outline optimization strategies for denoising and censoring, and present validation techniques to ensure robust and reproducible results. The content is tailored to empower professionals in distinguishing true neurobiological signals from motion-induced artifacts, thereby strengthening the foundation for biomarker discovery and clinical translation.

The Stealthy Confounder: Understanding Motion's Systemic Bias in Brain-Wide Association Studies

The Pervasive Challenge of In-Scanner Head Motion in fMRI

Head motion during functional magnetic resonance imaging (fMRI) scans represents one of the most significant methodological challenges in contemporary neuroimaging research. Even sub-millimeter movements can introduce substantial artifacts that systematically distort functional connectivity measures, morphometric analyses, and diffusion imaging results [1] [2]. These motion-induced artifacts create spurious correlations in resting-state fMRI data, primarily affecting short-range connections and potentially mimicking trait correlates of behavior [2] [3]. The problem is particularly pronounced in specific populations, including children, elderly patients, and individuals with neurodevelopmental or psychiatric disorders who often exhibit increased movement [4] [5]. Despite advanced processing pipelines, residual motion artifacts persist and can confound study results, making motion mitigation an essential consideration for robust research design, especially in studies investigating brain-behavior relationships [4] [2].

Quantifying the Impact: How Motion Distorts Data

Mechanisms of Artifact Generation

Head motion affects fMRI data through multiple physical mechanisms. When a subject moves in the scanner, it perturbs the spatial frequencies (k-space) of the MRI, introducing errors that propagate throughout the image and manifest as ghosting, ringing, and blurring artifacts [4]. Motion changes the tissue composition within a voxel, distorts the magnetic field, and disrupts the steady-state magnetization recovery of spins in slices that have moved [6]. In resting-state fMRI, these disruptions lead to distance-dependent biases in signal correlations, spuriously increasing correlations between nearby voxels while leaving long-range connections relatively unaffected [2].

Consequences for Structural and Functional Analyses

The impact of motion extends across various MRI modalities. In structural imaging, motion artifacts have been shown to systematically reduce estimated grey matter volumes and cortical thickness, with particularly pronounced effects in developmental studies [4]. In functional connectivity analyses, motion-induced signal changes can persist for more than 10 seconds after the physical movement has ceased, creating extended periods of data corruption [2]. These artifacts are often shared across nearly all brain voxels, making them particularly challenging to remove through standard processing techniques [2].

Table 1: Effects of Motion on Different Neuroimaging Modalities

Imaging Modality	Primary Effects of Motion	Key References
Structural MRI	Decreased grey matter volume estimates, reduced cortical thickness, blurred tissue boundaries	[4]
Resting-state fMRI	Distance-dependent correlation biases, increased short-range connectivity, reduced long-range connectivity	[2] [3]
Task-based fMRI	Signal dropouts, altered activation maps, reduced statistical power	[6]
Diffusion MRI	Altered fractional anisotropy, changed mean diffusivity, corrupted tractography	[4]
Magnetic Resonance Spectroscopy (MRS)	Degraded spectral quality, line broadening, incorrect metabolite quantification	[7]

Identifying Motion-Prone Populations

Understanding which factors predict increased head motion is crucial for effective study design and appropriate implementation of motion mitigation strategies. Recent large-scale analyses have identified several key indicators that can help researchers anticipate motion-related challenges in their specific cohorts.

Table 2: Subject Factors Associated with Increased Head Motion in fMRI

Factor Category	Specific Factor	Association with Motion	Effect Size/Notes
Demographic	Age (children <10 years)	Strong increase	Non-linear cortical thickness associations disappear with motion correction [4]
Demographic	Age (adults >40 years)	Moderate increase	Motion increases at extreme age ranges [4]
Anthropometric	Body Mass Index (BMI)	Strong positive correlation	A 10-point BMI increase corresponds to 51% motion increase [5]
Clinical	Psychiatric disorders (ASD, ADHD)	Variable increase	Effect sizes attenuated with motion correction [4] [5]
Clinical	Hypertension	Significant increase	p = 0.048 in adjusted models [5]
Behavioral	Cognitive task performance	Increased motion	t = 110.83, p < 0.001 [5]
Behavioral	Prior scan experience	Reduced motion	t = 7.16, p < 0.001 [5]

Large-scale studies have revealed that BMI and ethnicity demonstrate the strongest associations with head motion, with a ten-point increase in BMI (approximately the difference between "healthy" and "obese" classifications) corresponding to a 51% increase in motion [5]. Interestingly, disease diagnoses alone (including psychiatric, musculoskeletal disorders, and diabetes) were not reliable predictors of increased motion, suggesting that individual characteristics outweigh diagnostic categories in motion prediction [5].

Prospective Motion Mitigation Strategies

Behavioral Interventions

Behavioral strategies represent the first line of defense against head motion artifacts. Research has demonstrated that simple interventions can significantly reduce motion, particularly in challenging populations:

• Movie Watching: Presenting engaging movie content during scans significantly reduces head motion compared to rest conditions, especially in younger children (5-10 years) [1]. However, investigators should note that movie watching alters functional connectivity patterns compared to standard resting-state scans, making these conditions not directly comparable [1].

• Real-time Feedback: Providing visual feedback about head movement allows subjects to modify their behavior during scanning. This approach reduces motion in children, though the effects are age-dependent, with children older than 10 years showing minimal benefit [1].

• Subject Preparation: Adequate pre-scan training, acclimation sessions, and clear instruction significantly improve subject compliance and reduce motion [5]. Subjects with prior scan experience exhibit significantly reduced motion compared to first-time scanners [5].

Technological Solutions

Advanced acquisition methods provide powerful tools for prospective motion correction:

• Real-time Prospective Motion Correction (PMC): These systems use external tracking (optical cameras, NMR probes) or internal navigators to continuously update scan parameters based on head position [7] [8]. PMC simultaneously corrects for localization errors and B0 field inhomogeneities, which is particularly crucial for magnetic resonance spectroscopy (MRS) [7].

• Integrated Multimodal Correction: State-of-the-art systems combine prospective motion correction with parallel transmission techniques to address both motion artifacts and B1+ field inhomogeneity, particularly valuable at ultra-high field strengths (7T and above) [8].

Diagram 1: Motion Correction Strategy Classification

Retrospective Correction Methods

Standard Processing Approaches

Retrospective correction methods aim to remove motion artifacts after data acquisition through various processing techniques:

• Motion Parameter Regression: Including the estimated motion parameters (typically 6-24 regressors) as nuisance variables in general linear models. This approach only partially removes motion artifacts and leaves distance-dependent biases in functional connectivity [2].

• Global Signal Regression: Removing the global mean signal across the brain effectively reduces motion-related artifacts but remains controversial due to potential introduction of artificial anti-correlations and removal of neural signals of interest [2].

• Volume Censoring ("Scrubbing"): Identifying and removing motion-corrupted volumes based on framewise displacement (FD) or other quality metrics. This approach can effectively reduce motion-related group differences to chance levels when applied throughout the processing stream [2].

Advanced Computational Approaches

Recent methodological advances have introduced more sophisticated retrospective correction techniques:

• Structured Low-Rank Matrix Completion: This approach formulates artifact reduction as a matrix completion problem, enforcing a low-rank prior on a structured matrix formed from the time series samples. The method can recover missing entries from censoring while simultaneously performing slice-time correction, resulting in connectivity matrices with lower errors in pairwise correlation than standard pipelines [6] [9].

• Hybrid Motion Correction: Combining prospective motion correction with retrospective compensation for latency-induced errors, particularly valuable for addressing periodic motion such as breathing [8].

Experimental Protocols for Motion Management

Comprehensive Motion Correction Protocol for Resting-State fMRI

Based on current best practices, the following protocol provides a robust framework for motion management in resting-state fMRI studies:

• Pre-scan Preparation:

  • Conduct mock scanner session for naive participants
  • Provide clear instructions emphasizing importance of staying still
  • Use appropriate head stabilization (foam padding, custom molds)
  • For children or high-motion populations, implement movie watching or real-time feedback systems [1]

• Data Acquisition:

  • Incorporate navigator echoes or other prospective correction systems when available
  • Consider multiband acquisition to reduce scan time, but be aware of potential sensitivity to motion
  • Acquire additional timepoints to allow for subsequent data censoring

• Processing Pipeline:

  • Apply volume censoring with framewise displacement threshold of 0.2mm
  • Implement structured low-rank matrix completion to address data discontinuities from censoring [6] [9]
  • Include motion parameters, white matter, and CSF signals as nuisance regressors
  • Consider global signal regression based on specific research questions
  • Conduct quality control checks for residual motion-correlation relationships

Protocol for Structural and Multiparametric Mapping

For structural imaging and quantitative parameter mapping, a modified approach is necessary:

• Prospective Correction Integration:

  • Implement spiral navigators embedded in the acquisition sequence for real-time motion tracking [10]
  • Update slice position and orientation based on tracking data
  • For MRS, simultaneously update B0 shimming to maintain field homogeneity [7]

• Data Processing:

  • Apply rigid-body realignment with respect to a reference volume
  • Conduct visual inspection for residual artifacts
  • For multiparametric mapping, validate quantitative measurements against phantom data to ensure motion correction doesn't introduce bias [10]

Table 3: Research Reagent Solutions for Motion Management

Tool/Category	Specific Examples	Function/Purpose	Implementation Considerations
Software Packages	FSL (MCFLIRT), SPM, AFNI	Retrospective motion correction via image registration	Standard in most processing pipelines; provide motion parameter estimates [5]
Advanced Algorithms	Structured Low-Rank Matrix Completion	Recovery of censored data points using mathematical priors	Reduces discontinuities from scrubbing; improves connectivity estimation [6] [9]
Quality Metrics	Framewise Displacement (FD), DVARS	Quantification of inter-volume motion and signal changes	FD > 0.2mm indicates significant motion; used for censoring decisions [2] [3]
Behavioral Tools	Movie presentations, real-time feedback displays	Subject engagement and motion awareness	Particularly effective for children ages 5-10; alters functional connectivity [1]
Tracking Systems	Optical cameras, NMR probes, FID navigators	Prospective motion tracking for real-time correction	External tracking preferred for localization; navigators needed for B0 correction [7] [8]

Diagram 2: Motion Artifact Problem-Solution Framework

Frequently Asked Questions

Q1: What framewise displacement threshold should I use for volume censoring?

• For standard adult populations, a threshold of 0.2mm effectively identifies motion-corrupted volumes while retaining sufficient data [3]. However, stricter thresholds (0.1-0.15mm) may be necessary for studies focusing on short-range connectivity or involving high-motion populations. Studies have shown that prediction accuracy remains similar with both lenient (FD=0.5mm) and strict (FD=0.2mm) censoring thresholds [3].

Q2: Does global signal regression effectively remove motion artifacts?

• Global signal regression is highly effective at reducing motion-related variance in fMRI data, but it remains controversial because it can introduce artificial anti-correlations and potentially remove neural signals of interest [2]. The decision to use GSR should be theory-driven and consistent across all subjects in a study.

Q3: How does motion affect different populations, and should I exclude high-motion subjects?

• Motion affects clinical populations differentially, with increased motion typically observed in children, elderly individuals, and those with certain neuropsychiatric conditions [4] [5]. Rather than wholesale exclusion, researchers should implement rigorous motion correction strategies and include motion as a covariate in group analyses. Exclusion should be based on objective quality metrics after correction, not solely on group membership.

Q4: What is the most effective strategy for scanning children?

• For children aged 5-10 years, combined behavioral interventions including movie watching and real-time feedback significantly reduce motion [1]. Mock scanner training sessions and age-appropriate explanation of the importance of staying still also improve compliance. For very young children, shorter acquisition protocols with built-in breaks may be necessary.

Q5: Can functional connectivity predict an individual's head motion?

• Emerging evidence suggests that individual differences in functional connectivity, particularly within the cerebellum and default-mode network, can predict in-scanner head motion [3]. This suggests there may be neurobiological traits associated with motion control that extend beyond simple compliance.

Q6: What are the latest technical advances in motion correction?

• Current research focuses on hybrid methods combining prospective and retrospective correction, structured low-rank matrix completion for data recovery, and deep learning approaches for motion detection and quality assessment [6] [8]. These approaches show promise for more effective artifact reduction without the need for extensive data censoring.

Effectively addressing the challenge of in-scanner head motion requires a comprehensive, multi-layered approach incorporating both prospective prevention and retrospective correction strategies. Researchers must carefully consider their specific population characteristics, with particular attention to factors like BMI and age that strongly predict motion [5]. Implementation of behavioral interventions should be standard practice for challenging populations, while advanced processing techniques like structured matrix completion offer promising avenues for recovering signal from motion-corrupted data [6] [9]. Critically, motion management must be integrated into every stage of research design, from participant selection and scanning protocols to data processing and statistical analysis, to prevent the introduction of spurious brain-behavior relationships and ensure the validity of neuroimaging findings.

What are motion artifacts and why are they a critical issue in functional neuroimaging? Motion artifacts are disturbances in neuroimaging signals caused by the subject's movement. In functional near-infrared spectroscopy (fNIRS), head movements cause a decoupling between the source/detector fiber and the scalp, which is reflected in the measured signal, usually as a high-frequency spike and a shift from the baseline intensity [11] [12]. In functional magnetic resonance imaging (fMRI), head motion introduces bias in measured functional connectivity (FC) through both common effects across all pair-wise regional correlations as well as distance-dependent biases, where correlations are increased most for adjacent regions and relatively decreased for regions that are distant [13]. These artifacts are particularly problematic because they can create spurious brain-behavior relationships, especially when comparing groups that differ systematically in head motion (e.g., children vs. young adults, clinical populations vs. controls) [13].

What types of motion artifacts affect fNIRS signals? Motion artifacts in fNIRS can be generally classified into three categories [11] [12]:

• Spikes: High-amplitude, high-frequency artifacts easily detectable in the data-series
• Baseline shifts: Sustained deviations from the baseline signal
• Low-frequency variations: Slower artifacts that are harder to distinguish from normal hemodynamic fNIRS signals

These artifacts can be isolated events or temporally correlated with the hemodynamic response, with the latter being particularly challenging to correct [11].

Troubleshooting Guide: Motion Artifact Identification

FAQ: How can I quickly identify if my fNIRS data contains significant motion artifacts?

Visual inspection of the raw signal remains one of the most effective initial screening methods. Look for these characteristic signs [11] [12]:

• Abrupt, high-amplitude spikes that deviate sharply from the baseline signal
• Sudden shifts in baseline that persist for multiple timepoints
• Low-frequency variations that coincide with participant movement or task performance
• Signal saturation where the intensity reaches the maximum or minimum recordable value

For a more systematic approach, implement automated artifact detection algorithms that can identify segments exceeding predetermined amplitude or derivative thresholds [14].

FAQ: What specific movements most commonly corrupt fNIRS signals?

Recent research using computer vision to characterize motion artifacts has identified that [15]:

• Repeated movements, upward movements, and downward movements tend to most significantly compromise fNIRS signal quality
• The occipital and pre-occipital regions are particularly susceptible to MAs following upwards or downwards movements
• The temporal regions are most affected by bend left, bend right, left, and right movements
• Movements involving jaw motion (talking, eating) can cause significant artifacts in frontal regions [14]

Motion Correction Techniques for fNIRS

FAQ: What are the most effective motion correction methods for fNIRS data?

Table 1: Comparison of fNIRS Motion Correction Techniques

Method	Principle	Best For	Efficacy (AUC Reduction)	Limitations
Wavelet Filtering	Multi-scale decomposition & thresholding	All artifact types, especially task-correlated	93% of cases [11] [12]	Requires parameter optimization
Spline Interpolation	Identify artifacts & interpolate with cubic splines	Easily detectable spikes [11] [12]	Variable performance [11] [12]	Dependent on accurate artifact detection
PCA	Remove components with high variance	When motion is principal variance source [11] [12]	Variable performance [11] [12]	May remove physiological signals
Kalman Filtering	Adaptive filtering based on signal model	Real-time applications [14]	Not specified in studies	Requires model assumptions
CBSI	Leverages negative correlation between HbO/HbR	Hemodynamic signals [11] [12]	Good for HbO/HbR correlation [11] [12]	Only applicable to hemoglobin signals

fNIRS Motion Correction Workflow

Experimental Protocol: Comparing fNIRS Motion Correction Techniques

For researchers seeking to validate motion correction methods, this protocol adapted from Brigadoi et al. provides a robust framework [11] [12]:

• Participant Preparation: Secure fNIRS optodes on the head using a probe-placement method based on a physical model of the head surface to ensure consistent positioning across subjects.

• Task Design: Implement a cognitive paradigm likely to induce mild, task-correlated motion artifacts, such as a color-naming task where participants verbalize responses. This creates low-frequency, low-amplitude motion artifacts correlated with the hemodynamic response.

• Data Acquisition: Collect fNIRS data at sufficient sampling frequency (≥7.8 Hz) using a multi-channel, frequency-domain NIR spectrometer with dual wavelengths (690 nm and 830 nm) to compute concentration changes of oxy-hemoglobin (HbO) and deoxy-hemoglobin (HbR).

• Motion Correction Application: Apply multiple correction techniques (wavelet filtering, spline interpolation, PCA, Kalman filtering, CBSI) to the same dataset for comparative analysis.

• Performance Evaluation: Use these objective metrics to evaluate correction efficacy:

  • AUC₀–₂: Area under the curve during the first 2 seconds post-stimulus
  • AUC ratio: Ratio of AUC during different time windows
  • Within-subject standard deviation: Consistency of hemodynamic responses across trials

Motion Correction in fMRI

FAQ: What motion correction strategies work best for resting-state fMRI?

Table 2: Efficacy of fMRI Motion Correction Pipelines

Method	Residual Motion-FC Relationship	Data Loss	Test-Retest Reliability	Clinical Sensitivity
Volume Censoring	Excellent [16]	High [16]	Good [16]	Affects group differences [16]
ICA-AROMA	Good [16]	Moderate [16]	Good [16]	Moderate impact [16]
aCompCor	Only in low-motion data [16]	Low [16]	Variable [16]	Low impact [16]
Global Signal Regression	Improves most pipelines but increases distance-dependence [16]	Low [16]	Good [16]	Significant impact [16]
Basic Regression	Poor [16]	Low [16]	Poor [16]	Low impact [16]

FAQ: How does respiration affect motion parameters in fMRI?

Respiration influences realignment estimates in two ways [13]:

• True changes in head position caused by respiration-related movement
• Apparent (factitious) motion in the phase-encoding direction generated by perturbations of the main (B0) magnetic field caused by chest wall motion

The rate of respiration in adults (12-18 breaths per minute, 0.2-0.3 Hz) often aliases into frequencies from 0.1-0.2 Hz in single-band fMRI studies with TR=2.0-2.5s [13]. This high-frequency motion (HF-motion) is more common in older adults, those with higher body mass index, and those with lower cardiorespiratory fitness [13].

Solution: Implement a low-pass filtering approach (cutoff ~0.1-0.15 Hz) to remove HF-motion contamination from motion summary measures like framewise displacement (FD). This approach saves substantial amounts of data from FD-based frame censoring while still effectively reducing motion biases in functional connectivity measures [13].

fMRI Motion Artifact Sources

Research Reagent Solutions

Table 3: Essential Tools for Motion Artifact Research

Tool/Resource	Function	Application Context
Homer2/Homer3 Software	Open-source fNIRS analysis	Processing fNIRS data, implementing motion correction algorithms [17]
NIRS Toolbox	MATLAB-based fNIRS analysis	Flexible processing pipelines, motion correction implementation [18]
Accelerometers/IMUs	Motion tracking	Hardware-based motion artifact detection and correction [14]
Computer Vision Systems	Head movement quantification	Ground-truth movement data for artifact characterization [15]
ICA-AROMA	ICA-based motion removal for fMRI	Automated removal of motion components in resting-state fMRI [16]
Short-Separation Detectors	Superficial signal measurement	Reference channels for motion artifact regression in fNIRS [11]

Experimental Protocol: Validating Motion Correction with Ground-Truth Data

For method development and validation, this protocol using computer vision provides rigorous artifact characterization [15]:

• Experimental Setup: Position participants in front of a video recording system while wearing a whole-head fNIRS cap. Use the SynergyNet deep neural network or similar computer vision system to compute head orientation angles from video frames.

• Movement Paradigm: Instruct participants to perform controlled head movements along three main rotational axes (vertical, frontal, sagittal) with variations in speed (fast, slow) and type (half, full, repeated rotation).

• Data Synchronization: Precisely synchronize video frames with fNIRS data acquisition using trigger signals or timestamp alignment.

• Feature Extraction:

  • From video data: Extract maximal movement amplitude and speed from head orientation data
  • From fNIRS data: Identify spikes and baseline shifts using automated algorithms

• Correlation Analysis: Quantify the relationship between specific movement parameters (amplitude, velocity, direction) and artifact characteristics in the fNIRS signal.

Best Practices and Recommendations

FAQ: Should I reject motion-contaminated trials or correct them?

Always correct rather than reject when possible. Evidence from fNIRS studies shows that motion correction is always better than trial rejection, with wavelet filtering reducing the area under the curve where the artifact is present in 93% of cases [11] [12]. Trial rejection should only be considered when:

• The number of motion artifacts is low
• The total number of trials is high
• Artifacts are so severe that they defy correction

In challenging populations (infants, clinical patients, children) where trial numbers are often limited, correction is strongly preferred over rejection [11].

FAQ: What reporting standards should I follow for motion correction?

To ensure transparency and reproducibility:

• Clearly specify the exact motion correction algorithm and parameters used
• Report the amount of data censored or excluded due to motion
• Include quality metrics demonstrating correction efficacy
• Test and report the residual relationship between motion and functional connectivity after correction
• For fMRI studies: Conduct simple analyses to report the degree to which findings may be affected by motion-related artifact [16]

Effectively addressing motion artifacts requires a comprehensive approach spanning experimental design, data acquisition, processing methodology, and transparent reporting. The most successful strategies:

• Combine multiple correction approaches rather than relying on a single method
• Implement proactive measures during data collection (secure cap placement, participant instruction)
• Validate correction efficacy using objective metrics relevant to your research question
• Account for motion in group comparisons where motion differences may confound results

By adopting these practices, researchers can significantly reduce the risk of spurious brain-behavior relationships arising from motion artifacts rather than true neural phenomena.

Technical Support Center: Troubleshooting Motion Artifact in Brain-Behavior Research

This technical support center provides troubleshooting guides and FAQs for researchers investigating brain-behavior relationships, with a specific focus on mitigating motion-related artifacts that disproportionately affect studies of psychiatric and developmental populations.

Frequently Asked Questions: Motion Vulnerability

Q1: Why are studies of psychiatric and developmental populations particularly vulnerable to motion artifacts?

Research consistently shows that in-scanner head motion is systematically higher in individuals with certain psychiatric (e.g., ADHD, autism) and developmental conditions [19] [20]. This is not merely a behavioral inconvenience but a major methodological confound. Since motion introduces systematic bias into functional connectivity (FC) measures, and this motion is correlated with the trait of interest, it creates a high risk of spurious findings—where observed brain-behavior relationships are driven by motion artifact rather than true neural phenomena [19] [20]. For example, motion artifact systematically decreases FC between distant brain regions, which could lead a researcher to falsely conclude that a disorder causes reduced long-distance connectivity [19].

Q2: What is a "motion impact score," and how can I use it?

A motion impact score, such as that generated by the Split Half Analysis of Motion Associated Networks (SHAMAN) method, quantifies the degree to which a specific trait-FC relationship is impacted by residual head motion [19] [21]. It helps researchers distinguish between:

• Motion Overestimation: The artifact inflates the observed trait-FC effect.
• Motion Underestimation: The artifact masks or reduces the observed trait-FC effect [19]. Using this score allows for a trait-specific assessment of motion contamination, moving beyond generic, study-wide motion metrics.

Q3: After standard denoising, how prevalent is significant motion artifact?

Residual motion artifact remains a significant issue even after standard denoising pipelines are applied. One study assessing 45 traits from over 7,000 participants in the Adolescent Brain Cognitive Development (ABCD) Study found that after standard denoising without motion censoring:

• 42% (19/45) of traits had significant motion overestimation scores.
• 38% (17/45) of traits had significant motion underestimation scores [19] [21] [22]. This underscores that denoising alone is often insufficient for motion-correlated traits.

Q4: Does rigorous motion censoring solve the problem?

Censoring (removing high-motion volumes) is effective but imperfect. The same ABCD study showed that censoring at Framewise Displacement (FD) < 0.2 mm reduced the number of traits with significant overestimation scores from 42% to just 2% (1/45) [19] [21]. However, this aggressive censoring did not reduce the number of traits with significant underestimation scores, which remained prevalent [19]. This creates a tension: while censoring reduces false positives, it can also bias sample distributions by systematically excluding high-motion individuals who are critical for studying certain traits [19].

Quantitative Data on Risk Factors and Motion Impact

Table 1: Key Risk Factors for Developmental Vulnerability in Children (Ages 5-6) [23]

Risk Factor	Risk Ratio for Developmental Vulnerability
Poor Socioeconomic Status	1.58
Biological Sex - Male	1.51
History of Mental Health Diagnosis	1.46

Table 2: Impact of Denoising and Censoring on Motion Artifact (n=7,270) [19] [21]

Processing Step	Traits with Significant Motion Overestimation	Traits with Significant Motion Underestimation
After standard denoising (no censoring)	42% (19/45)	38% (17/45)
After censoring (FD < 0.2 mm)	2% (1/45)	38% (17/45)

Experimental Protocols for Motion Mitigation

Protocol 1: Implementing the SHAMAN Framework for Trait-Specific Motion Impact

Objective: To calculate a motion impact score for a specific trait-FC relationship to determine if it is spuriously overestimated or underestimated by residual head motion [19].

• Data Preparation: Start with preprocessed resting-state fMRI timeseries data and participant trait data (e.g., cognitive scores, symptom severity).
• Split Timeseries: For each participant, split the fMRI timeseries into two halves: one with higher relative motion and one with lower relative motion.
• Calculate Trait-FC Effects: Compute the correlation between the trait and FC strength separately for each half of the data.
• Compare Effects: Statistically compare the trait-FC effect size between the high-motion and low-motion halves. A significant difference indicates the trait-FC relationship is impacted by motion.
• Determine Direction:
  • Overestimation: The motion impact score aligns with the direction of the trait-FC effect.
  • Underestimation: The motion impact score opposes the direction of the trait-FC effect [19].
• Statistical Inference: Use permutation testing and non-parametric combining across connections to generate a p-value for the motion impact score [19].

Protocol 2: A Multi-Level Denoising and Censoring Pipeline

Objective: To minimize the influence of motion artifact in functional connectivity data.

• Preprocessing: Apply a standard denoising pipeline, which may include:
  • Motion correction (volume realignment)
  • Global signal regression
  • Regression of motion parameters and their derivatives
  • Nuisance variable regression (e.g., white matter, CSF signals)
  • Temporal filtering [20]
• Quantify Motion: Calculate Framewise Displacement (FD) for each volume to quantify head motion [20].
• Apply Censoring: Identify and remove (censor) volumes with FD exceeding a predetermined threshold (e.g., FD > 0.2 mm). Interpolate censored volumes using validated methods [19] [20].
• Quality Control: Ensure participants retain a sufficient amount of uncensored data (e.g., > 8-10 minutes) for reliable FC estimation. Report mean FD and the proportion of censored volumes for all groups to ensure comparability [20].

The Scientist's Toolkit: Essential Research Reagents & Materials

Table 3: Key Resources for fMRI Motion Mitigation Research

Item / Resource	Function & Explanation
SHAMAN Algorithm	A novel method for assigning a trait-specific motion impact score, distinguishing overestimation from underestimation of brain-behavior effects [19].
Framewise Displacement (FD)	A scalar summary measure of volume-to-volume head motion, derived from the realignment parameters. It is the primary metric for quantifying in-scanner motion and guiding censoring [20].
ABCD-BIDS Pipeline	A standardized, open-source denoising algorithm for fMRI data, incorporating global signal regression, respiratory filtering, and motion parameter regression [19].
High-Quality Population Datasets (e.g., ABCD Study)	Large-scale, open-access datasets (n > 10,000) with extensive phenotypic and neuroimaging data. They provide the statistical power necessary to detect true brain-behavior relationships and robustly quantify motion effects [19].

Workflow and Pathway Visualizations

Diagram 1: Experimental workflow for mitigating motion artifact, from raw data to interpretation of trait-specific motion impact.

Diagram 2: Logical relationships showing how population traits increase motion and risk of spurious findings.

Troubleshooting Guides & FAQs

FAQ: Experimental Design and Motion Artifacts

Q1: How can unintentional body movement create false brain-behavior correlations in our imaging data? A1: Movement artifacts can create spurious correlations by introducing structured noise that is misinterpreted as meaningful neural signal. Even minor head motion below 2mm can systematically bias functional connectivity estimates, particularly in regions with high susceptibility to artifacts like the prefrontal cortex. These motion-related signals can be misattributed to cognitive processes or clinical symptoms.

Q2: What are the most effective methods to control for motion artifacts in developmental populations? A2: A multi-layered approach is most effective:

• Proactive control: Implement rigorous in-scanner head restraint systems and provide extensive behavioral training using mock scanners.
• Acquisition parameters: Use sequences less susceptible to motion (multiband EPI, volumetric navigators).
• Post-processing: Include motion parameters as regressors in general linear models and implement stringent framewise displacement thresholds (<0.2mm).
• Quality metrics: Calculate quantitative measures like mean framewise displacement and signal-to-fluctuation for each participant.

Q3: Our infant movement analysis shows significant group differences in lower limb kinematics. Could this represent a false positive finding? A3: Possibly. According to recent research using marker-less AI video analysis, significant differences in lower limb features like 'Median Velocity' and 'Periodicity' were observed at 10 days old in infants later diagnosed with neurodevelopmental disorders (NDDs) with 85% accuracy [24]. However, these differences diminished over time, highlighting the importance of longitudinal assessment to distinguish transient from persistent motor signatures [24].

Troubleshooting Guide: Validating Motor Signatures

Problem: Inconsistent replication of early motor markers across research sites.

Root Cause Analysis: Variability in video recording conditions (camera angle, lighting, infant state) and differences in computational feature extraction pipelines can significantly impact kinematic measurements.

Solution Architecture:

• Standardized Protocol (Time: 2 weeks)
  • Establish fixed camera positions (90° zenith angle, 1.5m distance)
  • Control for infant state (quiet alert, post-prandial timing)
  • Implement calibration objects in frame for spatial reference

• Computational Validation (Time: 3 weeks)

  • Compare feature extraction against validated semi-automatic algorithms (e.g., Movidea software)
  • Target correlation coefficients: Pearson R >90% [24]
  • Calculate inter-rater reliability for manual coding (target Cohen's κ >0.8)

• Longitudinal Tracking (Time: 24 months)

  • Schedule assessments at multiple time points (10 days, 6, 12, 18, 24 weeks)
  • Monitor whether early differences persist or diminish over development

Quantitative Data Synthesis

Table 1: Lower Limb Kinematic Features Differentiating NDD and Typically Developing Infants at 10 Days Old [24]

Kinematic Feature	NDD Group Mean	TD Group Mean	Statistical Significance	Clinical Interpretation
Median Velocity	Higher	Lower	p < 0.05	Increased movement speed
Area Differing from Moving Average	Larger	Smaller	p < 0.05	Less smooth movement patterns
Periodicity	Reduced	Higher	p < 0.05	Less rhythmic movement organization

Table 2: Performance Metrics of SVM Classifier for Early NDD Identification [24]

Performance Metric	Result	Interpretation
Accuracy	85%	High overall correct classification
Sensitivity	64%	Moderate true positive rate
Specificity	100%	Perfect true negative rate
Sample Size	74 high-risk infants	Italian NIDA Network

Experimental Protocols

Protocol 1: Marker-Less Video Analysis of Infant Movements

Objective: To capture early motor signatures predictive of neurodevelopmental disorders using non-invasive video analysis.

Methodology:

• Participant Recruitment: Select high-risk infants (siblings of children with ASD, preterm newborns) from established networks (e.g., Italian NIDA Network) [24]
• Video Acquisition: Record five sessions at 10 days, 6, 12, 18, and 24 weeks of age
• Motion Tracking: Apply deep learning algorithms (e.g., OpenPose, AlphaPose, DeepLabCut) to extract body landmarks
• Feature Extraction: Calculate kinematic parameters (velocity, area, periodicity) from movement trajectories
• Validation: Compare against semi-automatic algorithms (target: Pearson R >90%, RMSE <10 pixels) [24]
• Outcome Assessment: Conduct comprehensive clinical evaluation at 36 months to establish diagnostic outcomes

Protocol 2: Controlling for Motion Artifacts in Neuroimaging

Objective: To minimize spurious brain-behavior relationships arising from head motion.

Methodology:

• Pre-scan Preparation: Implement mock scanner training with motion feedback
• Real-time Monitoring: Track head motion using volumetric navigators
• Acquisition Parameters: Use multiband acceleration, prospective motion correction
• Quality Control: Exclude participants with mean framewise displacement >0.5mm
• Statistical Control: Include 24 motion parameters (6 rigid-body + derivatives + squares) as nuisance regressors
• Validation: Test for correlations between motion and experimental variables of interest

Research Visualization

Experimental Workflow for Early NDD Detection

Motion Artifact Contamination Pathway

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials for Early NDD Motion Research

Research Tool	Function/Purpose	Specifications/Requirements
Marker-less AI Tracking Software (OpenPose, AlphaPose, DeepLabCut)	Automated extraction of body landmarks from video recordings	Capable of processing infant movement videos; outputs 25-point skeletal data [24]
High-resolution Digital Cameras	Video acquisition of infant spontaneous movements	Fixed position (90° angle, 1.5m distance); 30+ fps recording capability [24]
Validated Semi-automatic Algorithm (Movidea)	Benchmark for validation of automated tracking	Provides correlation metrics (target: Pearson R >90%, RMSE <10 pixels) [24]
Support Vector Machine (SVM) Classifier	Predictive modeling of NDD risk from kinematic features	Capable of handling multiple kinematic parameters; outputs probability scores [24]
Standardized Clinical Assessment Tools	Diagnostic confirmation at 36 months	ADOS, Mullen Scales, Vineland Adaptive Behavior Scales [24]

Frequently Asked Questions

What is the evidence that residual motion remains a problem after standard denoising? Even after applying comprehensive denoising pipelines like ABCD-BIDS, a significant proportion of signal variance (23%) can remain unexplained due to head motion [19]. Furthermore, analyses of specific trait-functional connectivity relationships show that a large number of traits (42%) can still exhibit significant motion overestimation scores after standard denoising, indicating that motion continues to inflate brain-behavior associations spuriously [19].

Why does motion create such persistent artifacts in functional connectivity estimates? Motion-correlated artifacts manifest in two primary forms: globally distributed artifacts that inflate connectivity estimates throughout the brain, and distance-dependent artifacts that preferentially affect short-range versus long-range connections [25]. This systematic bias causes decreased long-distance connectivity and increased short-range connectivity, most notably in default mode network regions [19]. The complex nature of these artifacts makes complete removal during standard denoising exceptionally challenging.

Which denoising strategies are most effective against residual motion? No single strategy completely eliminates motion artifacts, but combinations often work best. A comparative study found that combining FIX denoising with mean grayordinate time series regression (as a proxy for global signal regression) was the most effective approach for addressing both globally distributed and spatially specific artifacts [25]. However, censoring high-motion timepoints remains uniquely effective for reducing distance-dependent artifacts [26].

How can I validate that motion is not driving my brain-behavior findings? The SHAMAN framework provides a method to compute trait-specific motion impact scores that distinguish between motion causing overestimation or underestimation of trait-FC effects [19]. Additionally, Quality Control-Functional Connectivity correlations can evaluate whether connectivity values correlate with motion indicators across subjects, with Data Quality scores above 95% typically indicating minimal associations [27].

Troubleshooting Guides

Problem: Significant motion artifacts persist after standard denoising

Assessment:

• Check the motion-FC effect matrix correlation with average FC matrix; strong negative correlations (e.g., Spearman ρ = -0.58) indicate persistent motion effects [19]
• Calculate the proportion of signal variance explained by motion after denoising; values above 5-10% suggest inadequate denoising [19]
• Evaluate Quality Control-Functional Connectivity correlations using tools like CONN's Data Quality score [27]

Solutions:

• Implement additional censoring: Apply framewise displacement thresholding (e.g., FD < 0.2 mm), which can reduce significant motion overestimation from 42% to 2% of traits [19]
• Combine denoising methods: Use hybrid approaches such as FIX denoising with mean grayordinate time series regression [25]
• Apply trait-specific motion impact analysis: Use SHAMAN to identify which specific trait-FC relationships are affected by residual motion [19]

Problem: Motion artifacts differentially affect experimental conditions

Assessment:

• Compare mean framewise displacement between conditions (e.g., rest vs. task)
• Check for systematic differences in the number of censored volumes across groups
• Evaluate condition-specific changes in distance-dependent connectivity

Solutions:

• Use condition-balanced denoising: Apply aCompCor optimized to increase noise prediction power, which shows better balancing of artifacts between conditions like rest and task [26]
• Avoid over-reliance on censoring: While censoring reduces distance-dependent artifacts, it can significantly reduce network identifiability and statistical power [26]
• Implement simultaneous regression and filtering: Use techniques like "Simult" in CONN to avoid frequency mismatch between nuisance regression and filtering steps [27]

Quantitative Data on Motion Effects

Table 1: Effectiveness of Denoising Strategies in Reducing Motion-Related Variance

Denoising Method	Residual Variance Explained by Motion	Reduction in Global Artifacts	Reduction in Distance-Dependent Artifacts
Minimal processing (motion-correction only)	73% [19]	-	-
ABCD-BIDS denoising	23% [19]	Moderate	Moderate
Censoring (FD < 0.2 mm)	Not reported	Small	Small [25]
FIX denoising	Not reported	Substantial remaining [25]	Effective [25]
Mean grayordinate time series regression	Not reported	Significant [25]	Substantial remaining [25]
FIX + MGTR combination	Not reported	Most effective [25]	Most effective [25]

Table 2: Impact of Residual Motion on Trait-FC Associations in ABCD Study Data (n=7,270)

Motion Impact Type	Percentage of Traits Affected (Before Censoring)	Percentage of Traits Affected (After FD < 0.2 mm Censoring)
Significant overestimation	42% (19/45 traits) [19]	2% (1/45 traits) [19]
Significant underestimation	38% (17/45 traits) [19]	No decrease [19]
Total traits affected	80% (36/45 traits) [19]	Reduced but substantial underestimation remains [19]

Experimental Protocols

Protocol 1: SHAMAN Framework for Quantifying Trait-Specific Motion Impact

Purpose: To assign a motion impact score to specific trait-FC relationships that distinguishes between motion causing overestimation or underestimation of effects [19].

Procedure:

• Data Preparation: Process resting-state fMRI data using standard denoising (e.g., ABCD-BIDS pipeline including global signal regression, respiratory filtering, spectral filtering, despiking, and motion parameter regression) [19]
• Split-Half Analysis: Divide each participant's fMRI timeseries into high-motion and low-motion halves based on framewise displacement [19]
• Connectivity Calculation: Compute functional connectivity matrices separately for high-motion and low-motion halves
• Trait-FC Effect Estimation: Calculate the correlation between trait measures and FC for both halves
• Motion Impact Score: Compute the difference in trait-FC effects between high-motion and low-motion halves
• Statistical Testing: Use permutation testing and non-parametric combining across pairwise connections to obtain p-values for motion impact scores [19]

Interpretation:

• Motion impact score aligned with trait-FC effect direction indicates motion overestimation
• Motion impact score opposite to trait-FC effect direction indicates motion underestimation
• Significant p-value (p < 0.05) indicates trait-FC relationship is substantially impacted by residual motion [19]

Protocol 2: Comprehensive Denoising Strategy Evaluation

Purpose: To systematically evaluate the effectiveness of different denoising pipelines in removing motion artifacts.

Procedure:

• Data Acquisition: Collect resting-state fMRI data with associated motion parameters (framewise displacement, DVARS)
• Pipeline Application: Apply multiple denoising strategies to the same dataset:
  • Realignment/tissue-based regression
  • PCA-based methods (aCompCor)
  • ICA-based methods (ICA-AROMA)
  • Global signal regression
  • Censoring of motion-contaminated volumes [26]
• Benchmark Assessment: Evaluate each pipeline using:
  • QC-FC correlations: Correlation between motion metrics and functional connectivity [27]
  • Distance-dependent analysis: Examine how motion-connectivity correlations vary with connection distance [25]
  • Network identifiability: Ability to recover known functional networks [26]
• Data Quality Scoring: Calculate Data Validity (DV), Data Quality (DQ), and Data Sensitivity (DS) scores for each pipeline [27]

Optimization Criteria:

• Target Data Quality scores >95% indicating minimal motion-FC associations [27]
• Balance artifact reduction with network identifiability preservation [26]
• Ensure similar effectiveness across different experimental conditions (e.g., rest vs. task) [26]

Visualizing Motion Artifact Relationships

Diagram 1: Motion artifact types and their impact on brain-behavior research.

Diagram 2: Comprehensive workflow for assessing and addressing residual motion effects.

Research Reagent Solutions

Table 3: Essential Tools for Motion Detection and Correction in fMRI Research

Tool/Category	Specific Examples	Function	Key Considerations
Motion Quantification Metrics	Framewise displacement (FD), DVARS	Quantifies volume-to-volume head motion	FD < 0.2 mm threshold effective for reducing overestimation [19]
Denoising Pipelines	ABCD-BIDS, FIX, aCompCor, ICA-AROMA	Removes motion-related variance from BOLD signal	Combined approaches (FIX+MGTR) most effective [25]
Motion Impact Analysis	SHAMAN framework	Quantifies trait-specific motion effects	Distinguishes overestimation vs. underestimation [19]
Quality Control Metrics	Data Validity (DV), Data Quality (DQ), Data Sensitivity (DS) scores	Evaluates denoising effectiveness	Target scores >95% for DV and DQ [27]
Censoring Tools	Volume removal, Scrubbing	Removes high-motion timepoints	Reduces distance-dependent artifacts but decreases power [26]
Processing Software	CONN, FSL, MRtrix	Implements denoising pipelines	Consider processing order (denoising before vs. after distortion correction) [28]

Beyond Denoising: Implementing the SHAMAN Framework for Trait-Specific Motion Impact Scores

SHAMAN Technical Support Center

SHAMAN (Split Half Analysis of Motion Associated Networks) is a novel method for computing a trait-specific motion impact score that operates on one or more resting-state fMRI (rs-fMRI) scans per participant. It is designed to help researchers determine if their observed trait-functional connectivity (trait-FC) relationships are spuriously influenced by residual in-scanner head motion, thereby avoiding false positive results in brain-wide association studies (BWAS). [19]

This technical support center provides troubleshooting guides and FAQs to help researchers successfully implement SHAMAN in their analyses of brain-behavior relationships.

---

Frequently Asked Questions (FAQs)

Q1: What is the primary purpose of the SHAMAN algorithm? SHAMAN assigns a motion impact score to specific trait-FC relationships to distinguish between motion causing overestimation or underestimation of trait-FC effects. This is crucial for researchers studying traits associated with motion (e.g., psychiatric disorders) to avoid reporting spurious findings. [19]

Q2: How does SHAMAN differ from standard motion correction approaches? Most standard approaches for quantifying motion are agnostic to the hypothesis under study. SHAMAN specifically quantifies trait-specific motion artifact in functional connectivity, which is particularly important when studying motion-correlated traits like ADHD or autism. [19]

Q3: What are the software requirements for running SHAMAN? SHAMAN is implemented in MATLAB and requires access to the GitHub repository (DosenbachGreene/shaman). The code includes functionality for generating simulated data to test implementations. [29]

Q4: How does SHAMAN handle the trade-off between data quality and bias? There is a natural tension between removing motion-contaminated volumes to reduce spurious findings and systematically excluding individuals with high motion who may exhibit important trait variance. SHAMAN helps researchers quantify this specific trade-off for their trait of interest. [19]

Q5: What type of data input does SHAMAN require? SHAMAN requires preprocessed resting-state fMRI timeseries data for each participant. The method capitalizes on having at least one rs-fMRI scan per participant, though it can accommodate multiple scans. [19]

---

Troubleshooting Guides

Issue 1: Installation and Setup Problems

Problem: Difficulty cloning repository or starting SHAMAN in MATLAB.

Solution:

• Ensure Git is installed on your system.
• Run the following commands in your terminal or command line:

• Within MATLAB, generate simulated data to test your setup:
  [29]

Issue 2: Interpreting Motion Impact Scores

Problem: Understanding what a significant "motion overestimation" vs. "motion underestimation" score means for a specific trait-FC relationship.

Solution:

• Motion Overestimation Score: A significant positive score aligned with the direction of the trait-FC effect indicates motion is causing inflation of the observed effect. Your reported effect size is likely larger than the true biological effect. [19]
• Motion Underestimation Score: A significant negative score opposite the direction of the trait-FC effect indicates motion is obscuring or suppressing the observed effect. The true biological effect might be stronger than what your analysis detected. [19]

Issue 3: Framewise Displacement (FD) Censoring Thresholds

Problem: Uncertainty about selecting an appropriate FD censoring threshold and how it affects SHAMAN results.

Solution:

• Evidence from the ABCD Study suggests censoring at FD < 0.2 mm is effective. This threshold was shown to reduce the number of traits with significant motion overestimation scores from 42% (19/45) to just 2% (1/45). [19]
• However, note that this same censoring threshold did not decrease the number of traits with significant motion underestimation scores. Therefore, relying on censoring alone is insufficient, and SHAMAN should be used to check for residual underestimation effects. [19]

---

Table 1: Impact of Head Motion on Functional Connectivity (ABCD Study Data, n=7,270)

Metric	Value	Context / Interpretation
Signal variance explained by motion after minimal processing	73%	Square of Spearman's rho; indicates motion is a massive source of artifact. [19]
Signal variance explained by motion after ABCD-BIDS denoising	23%	A relative reduction of 69%, but residual artifact remains substantial. [19]
Correlation (Spearman ρ) between motion-FC effect and average FC matrix	-0.58	Strong, negative systematic bias: participants who moved more had weaker long-distance connections. [19]
Traits with significant motion overestimation scores (before FD < 0.2 mm censoring)	42% (19/45)	A large proportion of traits are vulnerable to inflated effect sizes. [19]
Traits with significant motion overestimation scores (after FD < 0.2 mm censoring)	2% (1/45)	Aggressive censoring is highly effective at mitigating overestimation. [19]
Traits with significant motion underestimation scores (before FD < 0.2 mm censoring)	38% (17/45)	Motion can also hide true effects. [19]
Traits with significant motion underestimation scores (after FD < 0.2 mm censoring)	38% (17/45)	Censoring alone does not resolve motion-caused underestimation. [19]

Table 2: SHAMAN Algorithm Inputs and Outputs

Component	Description	Purpose
Input: DataProvider Object	Points to folder containing fMRI timeseries data (e.g., in .mat files).	Interfaces the algorithm with the user's specific dataset. [29]
Input: Trait Names	An array of strings specifying the behavioral traits to analyze (e.g., ["trait"]).	Tells the algorithm which trait-FC relationships to test. [29]
Core Step: Split-Half Analysis	Splits each participant's timeseries into high-motion and low-motion halves.	Capitalizes on trait stability over time to isolate motion-related changes in FC. [19]
Core Step: Difference Matrix Calculation	For each participant, subtracts the high-motion FC matrix from the low-motion FC matrix.	In the absence of motion artifact, the difference should be zero. The residual reflects motion impact. [29]
Output: Motion Impact Score	A score and p-value from regressing the trait against the difference matrices.	Quantifies and tests the significance of motion's influence on a specific trait-FC relationship. [19] [29]

---

Detailed Experimental Protocols

Protocol 1: Core SHAMAN Workflow for a Single Trait

Objective: To compute a motion impact score for a single trait-FC relationship.

Methodology:

• Data Preparation: Start with preprocessed resting-state fMRI timeseries data for all participants. Ensure head motion parameters (e.g., Framewise Displacement) have been calculated for each timepoint. [19]
• Data Splitting: For each participant, split their entire fMRI timeseries into two halves: a "high-motion" half (containing the timepoints with the highest motion) and a "low-motion" half (containing the timepoints with the lowest motion). [19]
• Connectivity Matrix Generation: Generate a separate functional connectivity matrix from each half of each participant's data. [29]
• Motion Regression: Regress out between-participant differences in head motion from the connectivity matrices. This covariate aims to mop up some of the residual motion artifact. [29]
• Difference Calculation: For each participant, subtract the high-motion connectivity matrix from their low-motion connectivity matrix. This creates a single "difference matrix" per participant. [29]
• Motion Impact Regression: Regress the trait of interest (e.g., cognitive score) against the stack of participant difference matrices. The resulting statistic is the motion impact score, and its significance is assessed via permutation testing. [19] [29]

Protocol 2: Validating SHAMAN with Simulated Data

Objective: To confirm a correct installation and understand the output using a controlled, simulated dataset.

Methodology:

• Generate Simulation: Use the provided cdsimulate and simulate functions in MATLAB to create simulated fMRI and trait data, which is written to sub*.mat files. [29]
• Construct Data Provider: In MATLAB, create a SimulatedDataProvider object that points to the folder containing the newly created simulated data. [29]
• Initialize SHAMAN: Feed the data into the main Shaman algorithm object, specifying the name of the simulated trait to analyze. [29]
• Run with Few Permutations: For a quick result, set the number of permutations to a low number (e.g., shaman.permutations.nperm = 32;). [29]
• Retrieve Results: Execute shaman.get_scores_as_table to get a table of false positive motion impact scores. A successful run indicates the software is functioning correctly. [29]

---

Essential Visualizations

Diagram 1: SHAMAN Analytical Workflow

Diagram 2: Motion Impact Decision Logic

---

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Tools for SHAMAN Analysis

Item / Tool Name	Function / Purpose	Relevance to SHAMAN Protocol
High-Quality rsfMRI Data	The primary input; should have sufficient length (e.g., >8 mins) and be preprocessed.	SHAMAN requires timeseries data to perform the split-half analysis. Data from large cohorts like ABCD is ideal. [19]
Framewise Displacement (FD)	A scalar summary of head motion at each timepoint, calculated from realignment parameters.	Used to split the timeseries into high- and low-motion halves. A fundamental metric for the analysis. [19]
SHAMAN GitHub Repository	The official codebase (DosenbachGreene/shaman) implementing the algorithm.	Required to run the analysis. Contains core functions and example scripts for simulation. [29]
MATLAB Software	The numerical computing environment in which SHAMAN is implemented.	A system requirement for executing the provided code. [29]
DataProvider Object	A software object within the SHAMAN code that interfaces with the user's dataset.	Critical for feeding your specific data into the SHAMAN algorithm workflow. [29]
Permutation Testing Framework	A non-parametric statistical method within SHAMAN to assess significance.	Used to compute the p-value for the motion impact score, protecting against false inferences. [19]

Frequently Asked Questions

• FAQ 1: What is the fundamental trait-motion problem in brain-behavior research? Many psychological traits are assumed to be stable, but the behavioral and physiological data used to measure them, such as brain activity recorded by fMRI, are inherently variable and contaminated with motion artifacts. This creates a risk of identifying false, motion-driven correlations rather than genuine brain-behavior relationships [30].

• FAQ 2: How does the SHAMAN framework define and handle "trait stability"? SHAMAN does not assume traits are perfectly static. It incorporates a dynamic perspective, where a trait is conceptualized as a stable, central tendency (e.g., a mean value) around which there is natural, meaningful fluctuation. This approach reconciles long-term stability with short-term variability, preventing the misclassification of state-specific measures as stable traits [30].

• FAQ 3: What specific "motion variability" does the framework address? The framework addresses two types:

  • Physiological Motion: Head movement during fMRI scans, which can introduce significant spurious signal changes.
  • Psychological "Motion": The natural, moment-to-day fluctuation in self-reported states, affect, and behaviors, which can be studied using Ecological Momentary Assessment (EMA) methods to capture temporal dynamics [30].

• FAQ 4: What is the consequence of ignoring motion variability in my analysis? Ignoring motion variability can induce false positive results, where a correlation between a supposed "trait" and brain function is actually driven by a third, motion-related variable. This undermines the validity and replicability of findings [30].

• FAQ 5: Are there experimental paradigms that naturally embody this trait-stability/motion-variability principle? Yes. Research on shamanic trance provides a powerful model. The shamanic practitioner represents a stable trait-like role, while the journey into a trance state involves predictable, high-motion variability in brain network configuration. Studying this controlled transition helps isolate true neuro-correlates of an altered state from general motion artifacts [31].

---

Troubleshooting Guides

Guide 1: Resolving Spurious Correlations from Head Motion in fMRI

• Problem: A strong brain-behavior correlation is observed, but it is suspected to be a false positive driven by head motion.
• Investigation Protocol:
  • Quantify Motion: Calculate a framewise displacement (FD) time series for each participant's fMRI data.
  • Correlate with Behavior: Check the correlation between the mean FD (a subject-level motion metric) and your behavioral trait measure.
  • Scrub High-Motion Volumes: Identify and remove fMRI volumes where FD exceeds a threshold (e.g., 0.5mm).
  • Re-analyze: Re-run your primary analysis after this scrubbing procedure.
• Interpretation & Solution:
  • If the significant correlation disappears or substantially weakens after scrubbing, it was likely spurious and motion-related.
  • Solution: Implement motion scrubbing as a standard preprocessing step. Include mean FD as a nuisance covariate in your group-level statistical models to control for residual motion effects.

Guide 2: Validating Trait Stability versus State Variability

• Problem: Uncertainty over whether a questionnaire score measures a stable trait or a variable state, leading to misinterpretation of its relationship with brain function.
• Investigation Protocol:
  • Implement EMA: Administer brief, repeated measures of the construct (e.g., negative affect) up to 24 times daily for one week using a smartphone app [30].
  • Calculate Stability Metrics: For each participant, calculate both the mean (estimating the stable trait level) and the standard deviation (estimating the state variability) of their EMA scores [30].
  • Cross-Validate: Correlate the traditional one-time questionnaire score with the EMA-derived mean score.
• Interpretation & Solution:
  • A high correlation validates the questionnaire as a trait measure.
  • A low correlation suggests the questionnaire is sensitive to state variability.
  • Solution: Use the EMA-derived mean for trait analysis. The EMA-derived standard deviation can be used as a separate variable of interest, representing an individual's lability.

---

Experimental Protocols & Data

Protocol 1: EMA for Quantifying Trait and Variability

This methodology is adapted from research on personality and affect dynamics [30].

• Objective: To disentangle the stable core of a psychological construct from its inherent moment-to-day variability.
• Procedure:
  • Participant Recruitment: Recruit participants from your target population.
  • Baseline Trait Assessment: Administer standard one-time trait questionnaires (e.g., Big Five Inventory, PANAS for affect).
  • EMA Phase: For 7-14 days, prompt participants via a mobile app up to 24 times per day at random intervals.
  • EMA Items: At each prompt, present short-form versions of the traits/affects (e.g., "Rate how 'extraverted' you have been since the last prompt" or "How 'nervous' do you feel right now?") on a Likert scale [30].
  • Data Aggregation: For each participant, aggregate their EMA data to calculate person-specific mean (trait stability) and standard deviation (motion variability) for each construct.

Protocol 2: fMRI of Controlled State Transitions

This methodology is based on a study investigating the shamanic trance state [31].

• Objective: To study brain network reconfiguration during a well-defined transition from a normal state to an altered state, controlling for motion.
• Procedure:
  • Participants: Experienced practitioners capable of voluntarily entering a specific state (e.g., shamanic trance).
  • Stimulus: Use a standardized, rhythmic auditory stimulus (e.g., drumming at ~4 Hz) to induce the state [31].
  • fMRI Data Acquisition: Acquire fMRI data across multiple 8-minute runs, alternating between the state (trance) and a control (non-trance) condition in a counterbalanced order.
  • Subjective Measures: After each run, administer a phenomenological inventory to quantify the depth of the state experience [31].
  • Motion Monitoring & Control: Record head motion in real-time. Apply motion scrubbing and include motion parameters as regressors in the general linear model.
  • Analysis: Use functional connectivity analyses (e.g., eigenvector centrality mapping, seed-based correlation) to identify state-specific brain network changes, specifically controlling for motion parameters [31].

Quantitative Data from Foundational Studies

Table 1: Key Findings from the Shamanic Trance fMRI Study [31]

Brain Region	Function	Change During Trance	Interpretation
Posterior Cingulate Cortex (PCC)	Default Mode Network hub	↑ Eigenvector Centrality	Amplified internal focus and self-referential thought.
Dorsal Anterior Cingulate (dACC)	Control/Salience Network	↑ Eigenvector Centrality	Enhanced control and maintenance of the internal train of thought.
Left Insula/Operculum	Control/Salience Network	↑ Eigenvector Centrality	Increased awareness of internal bodily states.
Auditory Pathway	Sensory Processing	↓ Functional Connectivity	Perceptual decoupling from external, repetitive drumming.

Table 2: EMA-Derived Metrics for Traits and Affect [30]

Construct	Stability Metric (Mean)	Variability Metric (Std. Dev.)	Clinical/Research Implication
Extraversion	Average rating across prompts	Fluctuation around personal mean	High variability may correlate with creativity or stress.
Positive Affect	Average positive emotion level	Lability of positive emotion	High variability is linked to mood disorders.
Negative Affect	Average negative emotion level	Lability of negative emotion	High variability is a marker of neuroticism and emotional dysregulation.

---

The Scientist's Toolkit

Table 3: Essential Research Reagents & Materials

Item	Function in Experiment
Ecological Momentary Assessment (EMA) App	Enables real-time, in-the-moment data collection on traits and affect in a participant's natural environment, capturing temporal dynamics [30].
fMRI Scanner	Provides high-resolution data on brain activity and functional connectivity during controlled state transitions or task performance.
Rhythmic Auditory Stimulator	Delivers standardized, repetitive auditory stimuli (e.g., drumming at 4 Hz) to reliably induce predictable altered states of consciousness for study [31].
Motion Tracking System	Precisely quantifies head movement during fMRI scans, providing critical data for identifying and correcting motion artifacts.
Phenomenological Inventory	A standardized questionnaire (e.g., the Phenomenology of Consciousness Inventory) to quantitatively measure subjective experience after a state induction [31].

---

Framework Visualization

SHAMAN Core Mechanics Logic

SHAMAN Analysis Workflow

A technical support center guide for researchers

This guide provides troubleshooting and methodological support for researchers aiming to implement the Split Half Analysis of Motion Associated Networks (SHAMAN) framework to calculate distinct overestimation and underestimation scores, thereby preventing spurious brain-behavior relationships in functional connectivity (FC) research.

---

FAQs: Core Concepts and Procedures

1. What is the fundamental principle behind calculating separate overestimation and underestimation scores?

The method capitalizes on a key difference between the nature of a trait and motion: a trait (e.g., cognitive score) is stable over the timescale of an MRI scan, whereas head motion is a state that varies from second to second [19] [21]. The SHAMAN framework measures the difference in the correlation structure between split high-motion and low-motion halves of each participant's fMRI timeseries. A significant difference indicates that state-dependent motion impacts the trait's connectivity [19].

• Motion Overestimation Score: When the direction of the motion impact score is aligned with the direction of the trait-FC effect, it suggests motion is causing an overestimation of the trait-FC effect [19] [21].
• Motion Underestimation Score: When the motion impact score is opposite to the direction of the trait-FC effect, it suggests motion is causing an underestimation of the trait-FC effect [19] [21].

2. After standard denoising, how prevalent is the confounding effect of head motion on trait-FC associations?

Analyses of the ABCD Study dataset (n=7,270) after standard denoising with ABCD-BIDS but without motion censoring revealed that motion significantly confounds a large proportion of traits [19] [21]:

• 42% (19/45) of traits had significant motion overestimation scores.
• 38% (17/45) of traits had significant motion underestimation scores.

This confirms that residual motion artifact is a widespread source of potential bias, capable of inflating or masking true effects [19].

3. Does aggressive motion censoring eliminate both overestimation and underestimation bias?

No. The same study found that motion censoring strategies have an asymmetric effect [19] [21]. After censoring at framewise displacement (FD) < 0.2 mm:

• Significant motion overestimation was drastically reduced to 2% (1/45) of traits.
• However, the number of traits with significant motion underestimation was not decreased.

This highlights the critical need to quantify both types of bias, as underestimation may persist even after standard cleaning protocols are applied [19].

4. In what other experimental domains is the overestimation/underestimation distinction critically important?

The distinction is a common challenge across fields:

• Motor Neuroscience: Older adults and patients with Parkinson's disease often overestimate their physical performance (e.g., step distance), which is a safety hazard linked to falls. This overestimation is related to declines in walking ability [32].
• Visual Perception: When a moving object is temporarily occluded, observers consistently overestimate the duration of its occluded motion (or underestimate its speed), a robust bias observed in both action and perception tasks [33].
• Remote Sensing: In forest canopy height estimation using PolInSAR, the Random Volume over Ground (RVoG) model simultaneously underestimates heights in tall forests and overestimates them in sparse, low forests, requiring separate correction methods [34].

---

Experimental Protocol: SHAMAN Workflow

The following workflow is adapted from Kay et al. (2025) for implementing the SHAMAN framework [19] [21].

Objective: To compute trait-specific motion overestimation and underestimation scores for resting-state functional connectivity (FC) data.

Procedure in Detail:

• Data Input: Begin with preprocessed resting-state fMRI timeseries for all participants, corresponding framewise displacement (FD) timeseries as a measure of head motion, and the trait data of interest (e.g., cognitive scores) [19].
• Split Timeseries: For each participant, split their fMRI timeseries into two halves: one with the highest-motion timepoints and one with the lowest-motion timepoints, based on the FD metric [19] [21].
• Calculate FC Matrices: Compute separate functional connectivity matrices for the high-motion and low-motion halves for each participant [19].
• Compute Trait-FC Effects: For every connection (edge) in the FC matrix, perform a regression (e.g., using a generalized linear model that can include covariates like age and sex) with the trait as the predictor and the FC strength as the outcome. Perform this regression separately for the group-level FC data derived from the high-motion halves and the low-motion halves. This yields two vectors of beta coefficients (βhigh and βlow), representing the trait-FC effect size in high-motion and low-motion states, respectively [19].
• Calculate Motion Impact Score: For each FC edge, calculate the motion impact score as the difference: βhigh - βlow [19].
• Classify Bias Direction:
  • Overestimation: If the sign (positive/negative) of the motion impact score for an edge is the same as the sign of the trait-FC effect calculated using the full, unsplit timeseries, it indicates motion is causing an overestimation of this effect [19] [21].
  • Underestimation: If the sign of the motion impact score is opposite to the sign of the full-sample trait-FC effect, it indicates motion is causing an underestimation [19] [21].
• Statistical Inference: Use permutation testing (e.g., by shuffling trait labels) and non-parametric combining across all edges in the network to determine whether the overall overestimation and underestimation scores for the trait are statistically significant [19].

---

The table below summarizes key quantitative findings from the ABCD Study, demonstrating the effect of motion censoring on overestimation and underestimation scores [19] [21].

Table 1: Prevalence of Significant Motion Bias Before and After Motion Censoring (n=7,270, 45 traits)

Analysis Condition	Traits with Significant Overestimation	Traits with Significant Underestimation
After denoising (ABCD-BIDS), No Censoring	42% (19/45)	38% (17/45)
After denoising + Censoring (FD < 0.2 mm)	2% (1/45)	38% (17/45)

Key Interpretation: Censoring is highly effective at removing false positive inflation (overestimation) but is ineffective at recovering effects that are masked or suppressed (underestimation) by motion artifact [19].

---

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Resources for Implementing the SHAMAN Framework

Item	Function in the Protocol	Example/Note
High-Quality rs-fMRI Dataset	Provides the primary input timeseries data. Large sample sizes are crucial for reliable trait-FC effect estimation.	Adolescent Brain Cognitive Development (ABCD) Study [19]; Human Connectome Project (HCP) [19].
Framewise Displacement (FD)	A scalar quantity that summarizes head motion between volumes. Used to split timeseries into high- and low-motion halves [19].	Standard output from preprocessing tools (e.g., FSL, AFNI).
Denoising Pipeline	Removes major sources of noise and artifact from the BOLD signal before FC calculation.	ABCD-BIDS pipeline (includes global signal regression, respiratory filtering, motion parameter regression) [19].
Motion Censoring (Optional)	A post-processing step to exclude (censor) individual fMRI volumes with excessive motion.	Thresholds like FD < 0.2 mm are common. Note the asymmetric effect on bias [19].
SHAMAN Algorithm	The core computational method for calculating split-half motion impact scores.	Code implementation is required, capitalizing on the stability of traits over time [19] [21].
Permutation Testing Framework	Provides non-parametric statistical significance (p-values) for the calculated motion impact scores.	Essential for inference, typically involving thousands of permutations [19].

We hope this technical guide enhances the rigor and reliability of your brain-behavior association studies. For further troubleshooting, consult the primary reference: Kay et al. (2025) Nature Communications [19].

Frequently Asked Questions (FAQs)

Q1: What is the primary function of the SHAMAN resource monitor? The shaman-monitor service is a high-availability tool that runs on each node in a cluster. Its primary function is to enumerate and maintain a comprehensive list of all resources (virtual machines and containers) on its node. One service in the cluster is elected as the master, which manages the relocation and restarting of resources from failed nodes to healthy ones to ensure continuous operation [35].

Q2: What are the common issues that can occur with SHAMAN resources? Several issues can arise from inconsistencies in SHAMAN's resource tracking [35]:

• Broken Resources: A resource is marked as broken and moved to a /.shaman/broken directory if a command for it previously returned an error. These resources are ignored during high-availability events.
• Unregistered Resources: A resource exists in a node's SHAMAN repository, but the corresponding virtual environment is not registered on that node. This can lead to duplication during a failover.
• Orphaned Resources: A virtual environment is registered and running on one node, but its SHAMAN resource is located on a different node. Such resources will not be relocated during a node failure.
• Duplicate Resources: Two SHAMAN resources in different repositories point to the same virtual environment, causing conflicts.
• Missing Resources: A virtual environment's SHAMAN resource is missing entirely, which effectively disables high-availability for that resource.

Q3: How can I verify and fix broken or out-of-sync SHAMAN resources? You can use command-line tools to troubleshoot and repair SHAMAN resources [35]:

• Identify Broken Resources: Use the shaman stat command. Resources marked with a "B" in the output are broken.
• Clean Up Broken Resources: Run shaman cleanup-broken to re-register all broken resources on a node.
• Synchronize Resources: Execute shaman sync on each cluster node. This command will:
  • Re-register a VE on its local node if the resource was missing or on another node.
  • Delete a SHAMAN resource if the corresponding VE is not present on the node.
  • Update resource parameters to match the VE's parameters if they differ.

Troubleshooting Guides

Guide 1: Resolving "Broken" Resource States

Problem: A resource is marked as "broken" and is inactive after a storage access issue during a high-availability event [35].

Solution:

• Locate Broken Resources: Run shaman stat on your nodes and note any VE records marked with "B" [35].
• Check VE Status: Verify if these VEs are shown in the output of prlctl list on the nodes where they reside [35].
• Execute Repair:
  • If the VE is present, restarting it will often automatically fix its SHAMAN resource [35].
  • Alternatively, run # shaman [-c <cluster_name>] cleanup-broken to re-register all broken resources on a node [35].

Guide 2: Fixing Resource Location Mismatches and Duplicates

Problem: SHAMAN resources are located on the wrong node, or duplicate resources point to the same virtual environment, preventing successful migration or causing duplication during failover [35].

Solution:

• Manual Clean-up: Before syncing, you must manually clean up any unneeded (stopped or invalid) VE duplicates from the hypervisor [35].
• Cluster-wide Synchronization: Run the # shaman [-c <cluster_name>] sync command on every node in the cluster [35].
• Sync Logic: This command ensures a VE's SHAMAN resource is located on the same node where the VE itself is present, removing resources where the VE is absent and creating them where it is found [35].

The Scientist's Toolkit: Research Reagent Solutions

Item/Reagent	Primary Function in Analysis
shaman-monitor Service	Provides high-availability management by tracking and relocating resources across a cluster to prevent downtime [35].
shaman stat Command	A diagnostic tool used to list all resources and quickly identify any marked as broken ("B") [35].
shaman sync Command	A repair tool that synchronizes the SHAMAN resource repository with the actual state of virtual environments on a node [35].
shaman cleanup-broken Command	A targeted repair tool that attempts to re-register all resources previously marked as broken on a node [35].

SHAMAN Troubleshooting Command Reference

The table below summarizes the key commands for maintaining SHAMAN resource consistency.

Command	Primary Use Case	Key Outcome
shaman stat	Audit and discovery of resource states.	Lists all resources and identifies broken ones for further action [35].
shaman sync	Correcting location mismatches and duplicates.	Ensures SHAMAN resources are present on, and only on, the node where the VE exists [35].
shaman cleanup-broken	Repairing resources after storage/command errors.	Re-registers broken resources, attempting to return them to a managed state [35].

Experimental Protocol: SHAMAN Resource Consistency Check

Objective: To proactively verify the consistency of SHAMAN resources across a cluster and rectify any broken, misplaced, or duplicate entries to ensure reliable high-availability failover.

Methodology:

• Initial Audit: Execute the shaman stat command on every node within the cluster. Record the output, paying specific attention to any resources flagged with a "B" status [35].
• Data Comparison: Cross-reference the lists from each node to identify any virtual environments (VEs) that appear in the repository of one node but are actively running on a different node, as shown by prlctl list [35].
• Cleanup Phase: Manually remove any stopped or invalid VE duplicates present in the hypervisor environment. This is a critical prerequisite before running the sync command [35].
• Synchronization Phase: Run the shaman sync command sequentially on every node in the cluster. This will force the local SHAMAN repository to align with the actual VE state on that node [35].
• Validation: Re-run shaman stat on all nodes to confirm that broken resources have been cleared and that all resource locations are correct.

SHAMAN Resource Management Workflow

The following diagram illustrates the logical process for diagnosing and resolving common SHAMAN resource issues, from initial discovery to final synchronization.

Frequently Asked Questions

Q1: What does a "significant motion impact score" actually tell me about my finding? A significant motion impact score indicates that the observed relationship between a trait (e.g., a cognitive score or psychiatric diagnosis) and functional connectivity (FC) is not independent of head motion. The SHAMAN (Split Half Analysis of Motion Associated Networks) method specifically tests whether the trait-FC relationship changes significantly between high-motion and low-motion halves of your data. A significant score means this relationship is confounded by motion, raising doubts about whether your finding reflects true neurobiology or a motion-related artifact [21] [19].

Q2: What is the difference between "overestimation" and "underestimation"? The SHAMAN method distinguishes the direction of motion's bias:

• Motion Overestimation Score: The trait-FC effect is stronger in the high-motion half. This can lead to false positive results, making you believe an effect exists when it is partly or wholly driven by motion [19].
• Motion Underestimation Score: The trait-FC effect is weaker in the high-motion half. This can lead to false negative results, obscuring a genuine underlying brain-behavior relationship [19].

Q3: My finding has a significant motion impact score. Should I exclude it? A significant score is a major red flag requiring caution, not necessarily automatic exclusion. It means your result is not reliable. You should:

• Re-evaluate your denoising and censoring strategy. Applying stricter motion censoring (e.g., framewise displacement FD < 0.2 mm) has been shown to drastically reduce the number of traits with significant overestimation scores [21] [19].
• Interpret results with extreme caution. A finding with a significant motion impact score should not be presented as a robust, primary result without strong additional validation [21].
• Formally account for missing data in your analysis, as exclusion based on motion can systematically bias your sample [36].

Q4: Can stringent motion censoring solve all my motion-related problems? No. While censoring high-motion frames (e.g., FD < 0.2 mm) is highly effective at reducing motion overestimation, research on the ABCD dataset shows it does not significantly reduce the number of traits with motion underestimation scores [21] [19]. Censoring also creates a trade-off by potentially biasing your sample, as participants with higher motion often differ systematically in behavioral, demographic, and health-related variables [36].

Q5: Where can I find tools to help implement these methods and perform quality control? Several open-source software packages can assist with fMRI quality control:

• AFNI: Offers a suite of tools and an automated HTML report (APQC HTML) for systematic quality assessment of processing steps [37].
• pyfMRIqc: A Python package that generates user-friendly HTML reports on raw fMRI data quality, highlighting issues like artefacts and motion [38].
• Color Oracle: A tool to simulate how your figures appear to users with color vision deficiencies, ensuring your FC visualizations are accessible to all readers [39].

---

Quantitative Data on Motion Impact from the ABCD Study

The following data, derived from a large-scale analysis of the Adolescent Brain Cognitive Development (ABCD) Study (n=7,270), illustrates the prevalence and impact of motion on trait-FC findings.

Table 1: Prevalence of Significant Motion Impact Scores Before and After Censoring

This table shows the percentage of 45 assessed traits with significant (p < 0.05) motion impact scores, following standard denoising with ABCD-BIDS [21] [19].

Condition	Motion Overestimation	Motion Underestimation
No Motion Censoring	42% (19/45 traits)	38% (17/45 traits)
With Censoring (FD < 0.2 mm)	2% (1/45 traits)	38% (17/45 traits)

Key Takeaway: Stringent censoring is highly effective against overestimation bias but does not address underestimation bias [19].

Table 2: Recommended Quality Control Metrics and Tools

Incorporating these tools and metrics into your workflow is essential for detecting and mitigating motion-related bias.

Tool / Metric	Function	Relevance to Motion Impact
Framewise Displacement (FD)	Quantifies head motion from one volume to the next.	Primary metric for triggering frame censoring (scrubbing) [21].
SHAMAN Method	Provides a trait-specific motion impact score.	Directly tests if your trait-FC finding is confounded by motion [19].
AFNI's APQC HTML	Automated, interactive quality control report.	Checks alignment, motion correction, and other processing steps that affect data quality [37].
Color Oracle Simulator	Simulates colorblindness for figures.	Ensures FC maps and results are interpretable by a wider audience, promoting reproducible science [39].

---

Experimental Protocol: The SHAMAN Methodology

The SHAMAN (Split Half Analysis of Motion Associated Networks) procedure is designed to calculate a motion impact score for a specific trait-FC finding [19].

• Data Preparation: Start with preprocessed resting-state fMRI data and a phenotypic trait of interest.
• Split Time Series: For each participant, split their fMRI time series into two halves based on motion. The "high-motion" half contains the time points with the highest framewise displacement (FD), and the "low-motion" half contains the time points with the lowest FD.
• Calculate Half-Connectivity Matrices: Compute separate functional connectivity (FC) matrices for the high-motion and low-motion halves for every participant.
• Estimate Trait-FC Effects per Half: For each FC edge and for each half of the data independently, estimate the correlation (or regression coefficient) between the trait and the FC strength. This produces two trait-FC effect maps: one from the high-motion half and one from the low-motion half.
• Compute Motion Impact Score: The motion impact score for each FC edge is the difference between the trait-FC effects from the two halves (e.g., effect in high-motion half minus effect in low-motion half).
• Statistical Testing & Directional Interpretation:
  • Use permutation testing across participants to determine if the motion impact score is statistically significant.
  • Interpret the direction:
    • A significant positive score (aligned with the trait-FC effect) suggests motion overestimation.
    • A significant negative score (opposite the trait-FC effect) suggests motion underestimation [19].

The following diagram illustrates the core logic and workflow of the SHAMAN method:

---

Item	Function in Context	Key Detail / Best Practice
SHAMAN Framework	Provides a trait-specific motion impact score to distinguish overestimation from underestimation.	The primary methodological framework for directly testing the integrity of your trait-FC finding against motion [19].
Framewise Displacement (FD)	A scalar measure of head motion between volumes. Used for censoring and split-half analysis.	The standard metric for quantifying in-scanner head motion. A threshold of FD < 0.2 mm is commonly used for censoring [21].
Motion Censoring (Scrubbing)	The post-hoc removal of high-motion fMRI volumes from analysis.	Effective against overestimation but can introduce sample bias and does not fix underestimation [21] [36].
ABCD-BIDS Pipeline	A standardized denoising algorithm for resting-state fMRI data.	Includes global signal regression, respiratory filtering, and motion parameter regression. Reduces but does not eliminate motion artifact [19].
Colorblind-Safe Palettes	Color schemes for figures that are interpretable by readers with color vision deficiencies.	Use tools like ColorBrewer or Paul Tol's schemes. Avoid red/green combinations. Show greyscale individual channels for microscope images [39].
Automated QC Tools (e.g., AFNI, pyfMRIqc)	Software to systematically evaluate data quality at all processing stages.	Generates HTML reports with images and quantitative measures (e.g., motion parameters, TSNR) to flag problematic datasets [37] [38].

Practical Strategies for Motion Mitigation: From Censoring Thresholds to Analytical Pitfalls

Frequently Asked Questions (FAQs)

Q1: Why is in-scanner head motion such a critical issue for functional connectivity (FC) research? Head motion is the largest source of artifact in fMRI signals. It introduces systematic bias into resting-state functional connectivity data, which is not completely removed by standard denoising algorithms. This bias can lead to spurious brain-behavior associations, a particularly critical problem for researchers studying traits inherently correlated with motion, such as psychiatric disorders. For example, motion artifact systematically decreases FC between distant brain regions, which has previously led to false conclusions that autism decreases long-distance FC, when the results were actually driven by increased head motion in autistic participants [19].

Q2: What is the "motion impact score" and how does the SHAMAN method help? The motion impact score is derived from the Split Half Analysis of Motion Associated Networks (SHAMAN) method. SHAMAN is a novel approach that assigns a specific motion impact score to individual trait-FC relationships. Its key advantage is the ability to distinguish whether residual motion artifact is causing an overestimation or underestimation of a true trait-FC effect. This allows researchers to identify if their specific findings are impacted by motion, helping to avoid reporting false positives [19].

Q3: How effective is the ABCD-BIDS denoising pipeline at removing motion artifacts? While the ABCD-BIDS pipeline significantly reduces motion-related variance, it does not eliminate it entirely. In a large-scale assessment [19]:

• After only minimal processing (motion-correction by frame realignment), 73% of the fMRI signal variance was explained by head motion.
• After full denoising with ABCD-BIDS (which includes global signal regression, respiratory filtering, motion timeseries regression, and despiking), the variance explained by motion was reduced to 23%. This represents a 69% relative reduction, but the remaining residual motion is still sufficient to cause spurious associations [19].

Q4: Does aggressive motion censoring (e.g., FD < 0.2 mm) solve the problem? Motion censoring helps but is not a perfect solution. Research shows that censoring at framewise displacement (FD) < 0.2 mm dramatically reduces motion-related overestimation of trait-FC effects (from 42% to 2% of traits). However, it does not reduce the number of traits with significant motion underestimation scores. This highlights a complex relationship where censoring can fix one type of bias while potentially exacerbating another [19].

Q5: I am using preprocessed data from a large study like ABCD. Should I re-process it with a different pipeline like fMRIPrep? This is a common dilemma. The ABCD minimally preprocessed data has already undergone several steps, including motion correction and distortion correction. Some experts recommend against re-running motion correction on already motion-corrected data, as multiple interpolations can introduce new artifacts. A suggested alternative is to use tools like fMRIPrep primarily for generating new confound regressors (e.g., for CompCor) or for surface-based analysis if those outputs are not already available. The best practice is to align your processing stream with your specific analytical goals rather than applying redundant steps [40].

Q6: Is there a single "best" denoising pipeline that works for all studies? No. Evidence suggests that no single pipeline universally excels across different cohorts and objectives. The efficacy of a denoising strategy can depend on the specific dataset and the research goal—whether the priority is maximizing motion artifact removal or augmenting valid brain-behavior associations. Pipelines that combine multiple methods (e.g., ICA-FIX with global signal regression) often represent a reasonable trade-off [41].

---

Troubleshooting Guides

Issue: Suspecting Spurious Results Due to Motion

Problem: You are concerned that a significant brain-behavior relationship in your analysis might be driven by residual head motion, especially if the behavioral trait is known to correlate with motion (e.g., inattention, age).

Investigation and Solution:

• Quantify the Motion Impact: If possible, apply a method like SHAMAN [19] to your results. This will provide a specific p-value and score indicating whether and how motion is impacting your trait-FC finding.
• Benchmark Your Pipeline: Compare your primary results against those obtained from several different denoising pipelines. A finding that is robust across multiple pipelines is more trustworthy. The table below summarizes pipelines and their components [41].

Pipeline / Tool Name	Key Denoising Components	Primary Function / Context
ABCD-BIDS [42]	Global signal regression, respiratory filtering, motion parameter regression, despiking, spectral filtering.	Default for ABCD Study preprocessed data. A combination of HCP minimal pre-processing and DCAN Labs tools.
fMRIPrep [40]	Generation of motion regressors, tissue segmentation for CompCor, ICA-AROMA.	Robust pre-processing and confound variable generation; often used for surface-based analyses.
ICA-FIX + GSR [41]	Independent Component Analysis (ICA)-based artifact removal, Global Signal Regression (GSR).	A combination often identified as a good trade-off between motion reduction and behavioral prediction.
DiCER [41]	Diffuse Cluster Estimation and Regression.	An alternative method for noise component regression.
Standard Confound Regression	White matter & cerebrospinal fluid signal regression, motion parameter regression.	A common baseline approach in many studies.

• Systematically Vary Censoring: Re-run your analysis using different framewise displacement (FD) censoring thresholds (e.g., no censoring, FD < 0.3 mm, FD < 0.2 mm). If your effect size changes dramatically or disappears with stricter censoring, it may be motion-related [19].
• Check Spatial Correlation: Regress average head motion (mean FD) against FC across your participants to create a motion-FC effect matrix. A strong spatial correlation (e.g., Spearman ρ < -0.5) between this motion-FC matrix and your trait-FC effect matrix suggests contamination [19].

Issue: Working with Multi-Scanner Datasets (e.g., ABCD)

Problem: Inconsistencies in data processing or timing files when dealing with data from different scanner manufacturers (Siemens, Philips, GE).

Investigation and Solution:

• Identify Scanner and Software Version: Check the metadata (e.g., in the JSON sidecar files) for each subject to determine the scanner manufacturer and software version.
• Handle Dummy Scans Appropriately: The number of initial volumes (dummy scans) to discard varies by scanner [40]:
  • Siemens/Philips: Discard the first 8 volumes.
  • GE DV25: Discard the first 5 volumes.
  • GE DV26: Discard the first 16 volumes.
  • Note: Tools like fMRIPrep can automatically detect and flag non-steady state volumes for you.
• Use Provided Tools for Timing: For task-based fMRI in the ABCD Study, use the official MATLAB functions provided by the consortium (abcd_extract_eprime) to correctly generate timing files from the shared event files (*_events.tsv). Time=0 in these files is defined as the start of the first non-discarded TR [40].

---

The Scientist's Toolkit: Research Reagent Solutions

Essential Material / Resource	Function in Experimentation
ABCD-BIDS Community Collection (ABCC) [43]	A BIDS-standardized, rigorously curated collection of ABCD Study MRI data and derivatives, enabling reproducible and cross-study integrative research.
SHAMAN (Split Half Analysis of Motion Associated Networks) [19]	A specific methodological "reagent" to quantify the impact of motion on individual trait-FC associations, distinguishing over- from underestimation.
FIRMM (Real-time motion monitoring) [19]	Software for real-time motion analytics during brain MRI acquisition to improve data quality.
QSIPrep [42] [43]	An integrative platform for preprocessing and reconstructing diffusion MRI (dMRI) data.
fMRIPrep [40] [43]	A robust, standardized pipeline for functional MRI pre-processing, valued for its confound generation and surface-based processing.
XCP-D [43]	An extensible pipeline for post-processing and computing functional connectivity from resting-state fMRI data.

---

Quantitative Data on Pipeline Efficacy

Table 1: Efficacy of ABCD-BIDS Denoising on Motion-Related Variance [19]

Processing Stage	Variance in fMRI Signal Explained by Motion	Relative Reduction vs. Minimal Processing
Minimal Processing	73%	--
After ABCD-BIDS Denoising	23%	69%

Table 2: Impact of Motion and Censoring on Trait-FC Associations (n=7270, 45 traits) [19]

Condition	Traits with Significant Motion Overestimation	Traits with Significant Motion Underestimation
After ABCD-BIDS (No Censoring)	42% (19/45)	38% (17/45)
With Censoring (FD < 0.2 mm)	2% (1/45)	38% (17/45)

---

Detailed Experimental Protocol: The SHAMAN Method

The following workflow diagram outlines the SHAMAN procedure for calculating a motion impact score.

Title: SHAMAN workflow for motion impact score

Protocol Steps:

• Input Data: Begin with one or more resting-state fMRI (rs-fMRI) scans per participant that have undergone standard denoising (e.g., via ABCD-BIDS). The method capitalizes on the fact that traits are stable over the timescale of an MRI scan, while motion is a varying state [19].
• Split-Half Analysis: For each participant, split their fMRI timeseries into two halves: one with higher motion and one with lower motion [19].
• Calculate Functional Connectivity (FC): Compute the correlation structure (FC matrix) for both the high-motion and low-motion halves of the data [19].
• Compare Halves: Measure the difference in the FC between the two halves. A significant difference indicates that state-dependent motion is impacting the connectivity [19].
• Determine Impact Direction:
  • If the direction of the FC difference (high-motion vs. low-motion) is aligned with the direction of the trait-FC effect across participants, it suggests motion is causing an overestimation of the true effect.
  • If the direction of the FC difference is opposite to the trait-FC effect, it suggests motion is causing an underestimation of the true effect [19].
• Statistical Inference: Use permutation of the timeseries and non-parametric combining across pairwise connections to generate a final motion impact score and an associated p-value, distinguishing significant from non-significant motion impacts [19].

The Censoring Dilemma represents a fundamental challenge in brain-behavior research: how to remove motion-contaminated data to reduce false positive findings without introducing sample bias by systematically excluding participants prone to movement. This technical guide addresses this critical balance, providing methodologies to detect and correct for motion artifacts while preserving statistical power and population representativeness. Recent large-scale studies demonstrate that even after standard denoising, 42% of brain-behavior relationships show significant motion overestimation effects, while 38% show underestimation effects [19]. Through this support center, researchers gain practical tools to implement rigorous motion correction protocols that protect against spurious findings while maintaining sample integrity.

Evidence and Data Compendium

Quantitative Impact of Motion on Brain-Behavior Associations

Table 1: Motion Impact on Behavioral Traits After Standard Denoising (ABCD Study, n=7,270)

Motion Impact Type	Percentage of Traits Affected	Traits Examples	Primary Direction of Bias
Significant Overestimation	42% (19/45 traits)	Psychiatric diagnostic scales, attention measures	False positive associations
Significant Underestimation	38% (17/45 traits)	Cognitive performance measures	False negative associations
Minimal Motion Impact	20% (9/45 traits)	Non-motion-correlated measures	Minimal bias

Data from the ABCD Study reveals that even after application of standard denoising pipelines (ABCD-BIDS), motion continues to significantly impact the majority of trait-functional connectivity relationships. Researchers studying traits associated with motion (e.g., psychiatric disorders, attention measures) face particularly high risks of reporting spurious results without appropriate censoring methods [19].

Censoring Method Efficacy and Trade-offs

Table 2: Censoring Threshold Performance on Motion Artifact Reduction

Censoring Threshold (Framewise Displacement)	Residual Motion Overestimation	Residual Motion Underestimation	Sample Retention Impact	Recommended Application Context
No Censoring (FD > 0.9mm)	42% of traits affected	38% of traits affected	Maximum retention	Initial exploratory analysis only
Liberal (FD < 0.3mm)	15% of traits affected	35% of traits affected	Moderate retention	Studies requiring maximal power
Moderate (FD < 0.2mm)	2% of traits affected	35% of traits affected	Reduced retention	Standard analysis for motion-correlated traits
Stringent (FD < 0.1mm)	<1% of traits affected	Unknown effects	Severe retention issues	Final confirmatory analysis only

Application of framewise displacement censoring at FD < 0.2mm dramatically reduces motion overestimation effects (from 42% to 2% of traits) but does not address motion underestimation effects, which persist in 35% of traits. This illustrates the fundamental censoring dilemma: aggressive motion removal eliminates false positives but may not address all bias types while reducing statistical power through data exclusion [19].

Experimental Protocols

SHAMAN Methodology for Trait-Specific Motion Impact Assessment

The Split Half Analysis of Motion Associated Networks (SHAMAN) framework provides a specialized approach for assigning motion impact scores to specific trait-FC relationships [19].

Protocol Steps:

• Data Segmentation: Split each participant's fMRI timeseries into high-motion and low-motion halves based on framewise displacement (FD)
• Correlation Structure Analysis: Measure differences in functional connectivity between split halves
• Trait-FC Effect Comparison: Compare motion impact direction with trait-FC effect direction
• Statistical Testing: Use permutation testing and non-parametric combining across connections to generate motion impact scores and p-values
• Bias Direction Determination:
  • Motion impact score aligned with trait-FC effect direction = motion overestimation
  • Motion impact score opposite trait-FC effect direction = motion underestimation

Implementation Considerations:

• Requires one or more resting-state fMRI scans per participant
• Adaptable to include covariates in modeling
• Capitalizes on trait stability versus motion state variability
• Provides quantitative threshold for acceptable motion impact levels

Prospective Motion Correction for MRS Studies

For magnetic resonance spectroscopy (MRS) studies, prospective motion correction methods simultaneously update localization and B0 field to improve data quality [7].

Hardware Requirements:

• Internal navigator sequences or external tracking (optical cameras, NMR probes)
• Real-time shim updating capability
• Scanner frequency and 1st order shim updates every sequence repetition

Performance Specifications:

• Translation detection precision: <0.17 mm along three orthogonal axes
• Rotation detection precision: 0.26-2.9 degrees (depending on voxel location)
• Target metabolite stability: ≥5% for primary metabolites (NAA, Cr, Cho)
• Recommended for clinical populations with high motion (children, movement disorders)

Visual Workflows

Motion Artifact Management Decision Pathway

Motion Impact Detection and Resolution Workflow

Research Reagent Solutions

Table 3: Essential Tools for Motion Management in Brain-Behavior Research

Research Tool Category	Specific Solutions	Primary Function	Implementation Considerations
Motion Tracking Systems	Optical motion tracking, Inertial Measurement Units (IMU), Accelerometers, Camera-based systems	Real-time head movement quantification	External systems require integration; internal navigators more seamless [7] [14]
Prospective Correction Tools	Volumetric navigators (vNavs), FIELDMAP, Fat Navs, Spherical navigator echoes	Real-time adjustment of imaging parameters	Requires compatible scanner hardware and sequences [7]
Retrospective Correction Algorithms	SHAMAN, ABCD-BIDS, fMRIPrep, HCP Pipelines	Post-acquisition artifact correction	Varying sensitivity to different artifact types; parameter optimization needed [19]
Quality Control Frameworks	Framewise displacement (FD), DVARS, Quality Index (QI), Visual QC protocols	Data quality assessment and exclusion criteria	Standardized metrics enable cross-study comparisons [19] [44]
Statistical Control Methods	Motion matching, Global signal regression, Motion parameter inclusion, Covariate adjustment	Statistical mitigation of residual motion effects	Can introduce new biases; careful implementation required [19]

Frequently Asked Questions

Technical Implementation

Q: What censoring threshold should I use for framewise displacement in my analysis of children with ADHD?

A: For motion-correlated traits like ADHD, we recommend a multi-tiered approach:

• Begin with moderate censoring (FD < 0.2mm) for primary analysis, as this reduces motion overestimation from 42% to 2% of traits
• Conduct sensitivity analyses with liberal censoring (FD < 0.3mm) to evaluate sample bias
• Implement SHAMAN analysis to calculate trait-specific motion impact scores for your specific cohort
• Report results across multiple censoring thresholds to demonstrate robustness [19]

Q: How can I determine if my significant brain-behavior finding is actually driven by motion artifacts?

A: Implement the SHAMAN protocol to calculate motion impact scores:

• Split each participant's data into high-motion and low-motion halves
• Compare correlation structures between halves
• A significant difference indicates motion impact
• Direction relative to your trait-FC effect determines overestimation vs. underestimation This method specifically tests whether motion is driving your particular finding rather than relying on generic motion metrics [19].

Methodological Guidance

Q: My sample includes patients with movement disorders - how can I manage motion without excluding my entire clinical population?

A: Consider these approaches for high-motion populations:

• Prioritize prospective correction during acquisition using real-time tracking and shim updates [7]
• Apply deep learning-based retrospective correction that can improve cortical surface reconstruction quality even in Parkinson's disease patients [44]
• Use motion matching techniques rather than exclusion - match high-motion patients with high-motion controls
• Implement uncensored analyses with robust motion controls and demonstrate consistent findings across approaches

Q: What is the minimum performance specification I should require for motion correction in my MRS study?

A: For magnetic resonance spectroscopy, motion correction systems should achieve:

• Translation detection precision better than 0.17 mm
• Rotation detection precision between 0.26-2.9 degrees (depending on voxel location)
• Metabolite measurement stability of ≥5% for primary metabolites (NAA, Cr, Cho)
• Real-time update capability for both localization and B0 shimming [7]

Data Interpretation

Q: Why do I still see motion effects in my data after applying rigorous censoring (FD < 0.2mm)?

A: This expected result reflects the complexity of motion artifacts:

• Censoring at FD < 0.2mm primarily addresses motion overestimation effects (reducing from 42% to 2% of traits)
• Motion underestimation effects persist in approximately 35% of traits even after this censoring
• Different motion artifacts require specialized approaches - consider combining censoring with:
  • SHAMAN analysis for trait-specific impacts [19]
  • Prospective correction methods [7]
  • Advanced statistical controls

Q: How complete does mortality information need to be in real-world data to avoid significant bias in survival analysis?

A: Simulation studies reveal:

• With fully captured mortality, censoring at data cutoff provides unbiased median survival estimates
• With incomplete mortality data (sensitivity ~87%), censoring at last activity date underestimates median survival
• As missing mortality information increases, bias worsens with data cutoff censoring
• We recommend data providers validate and report mortality capture rates to inform methodological choices [45]

FAQs on Framewise Displacement and Motion Mitigation

Q1: Why is framewise displacement (FD) a critical metric in fMRI studies, especially for brain-behavior associations?

Head motion is the largest source of artifact in fMRI signals, and FD quantifies this motion from one volume to the next. Its critical importance stems from the fact that residual motion artifact, even after standard denoising, can systematically bias functional connectivity (FC) estimates. This bias is not random; it often decreases long-distance connectivity and increases short-range connectivity [19]. Because certain traits (e.g., symptoms of psychiatric disorders) and populations (e.g., children) are associated with more head motion, this artifact can create spurious brain-behavior relationships, leading to false positive or false negative findings if not properly addressed [46] [19].

Q2: What is the evidence that FD thresholds like 0.2 mm are effective?

Research on large datasets provides concrete evidence. One study on the ABCD dataset found that after standard denoising, 42% (19/45) of tested traits showed significant motion-related overestimation of brain-behavior effects. After applying censoring with an FD threshold of 0.2 mm, the number of traits with significant overestimation dropped dramatically to 2% (1/45) [19]. This shows that a 0.2 mm threshold is highly effective at mitigating one type of spurious finding. The same study also noted that motion-corrected data showed a weaker negative correlation between an individual's average FD and their overall functional connectivity strength, indicating a reduction in motion-based bias [19].

Q3: Does a strict threshold (e.g., FD < 0.2 mm) completely solve the motion problem?

No. While a 0.2 mm threshold is very effective against certain artifacts, it is not a panacea.

• Underestimation Artifacts: The same study that showed the benefits of the 0.2 mm threshold for overestimation found that it did not reduce the number of traits with significant motion-related underestimation scores [19]. This means motion can still suppress true effects.
• Residual Lagged Effects: Research indicates that even very small framewise displacements can be followed by structured, global changes in the BOLD signal that persist for 20-30 seconds [46]. This lagged structure can influence connectivity estimates even after the high-motion volumes themselves have been censored [46].

Q4: What alternative or supplementary methods can be used to manage motion artifacts?

Given that censoring alone is insufficient, a multi-pronged approach is recommended:

• Prospective Motion Correction (PMC): This technique uses an optical tracking system to update scan parameters in real-time during data acquisition. Studies show PMC improves the temporal signal-to-noise ratio (tSNR) of data acquired under motion and helps preserve the spatial and temporal characteristics of resting-state networks [47].
• Global Signal Regression (GSR): GSR has been shown to largely attenuate the artifactual lagged structure in the BOLD signal that follows head movements [46].
• Trait-Specific Validation: Methods like the Split Half Analysis of Motion Associated Networks (SHAMAN) can be used to compute a motion impact score for a specific trait-FC finding, testing whether it is likely spurious due to residual motion [19].

---

Data and Threshold Recommendations

The table below summarizes quantitative findings on the effectiveness of different denoising and censoring strategies.

Table 1: Efficacy of Motion Denoising and Censoring Strategies

Processing Step	Dataset	Key Metric	Result	Interpretation
Minimal Processing (motion correction only)	ABCD Study [19]	% of BOLD signal variance explained by FD	73%	The vast majority of signal variance is motion-related before cleanup.
Standard Denoising (ABCD-BIDS: GSR, respiratory filtering, etc.)	ABCD Study [19]	% of BOLD signal variance explained by FD	23%	Denoising achieves a 69% relative reduction in motion-related variance.
Standard Denoising (without censoring)	ABCD Study [19]	Traits with significant motion overestimation	42% (19/45)	Residual motion still strongly impacts nearly half of all traits studied.
Standard Denoising + Censoring (FD < 0.2 mm)	ABCD Study [19]	Traits with significant motion overestimation	2% (1/45)	Strict censoring is highly effective at eliminating overestimation artifacts.

Table 2: Characteristics of Common Motion Mitigation Strategies

Strategy	Mechanism	Key Advantages	Key Limitations / Considerations
Framewise Censoring (e.g., FD < 0.2 mm)	Removes high-motion volumes from analysis.	Highly effective at reducing motion-related overestimation of effects [19].	Can bias sample by excluding high-motion participants; does not address lagged artifact or underestimation [46] [19].
Global Signal Regression (GSR)	Removes the global mean signal from data.	Attenuates widespread motion-related artifact and lagged BOLD structure [46].	Controversial as it may remove neural signal; can induce negative correlations.
Prospective Motion Correction (PMC)	Updates scan parameters in real-time using head tracking.	Addresses the problem at acquisition; improves tSNR and network specificity under motion [47].	Requires specialized hardware; cannot be applied retrospectively to existing data.
Motion Parameter Regression	Models the relationship between motion parameters and BOLD signal over the entire run.	Standard, widely available method for mitigating motion-related variance.	Does not fully model the complex, temporally-lagged relationship between motion and BOLD signal [46].

---

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Tools for fMRI Motion Management

Tool or Method	Primary Function	Application Context
Framewise Displacement (FD) [46] [19]	A scalar index of volume-to-volume head movement derived from image realignment parameters.	The standard metric for quantifying head motion in any fMRI dataset and for implementing motion censoring.
Motion Censoring ("Scrubbing")	The post-hoc removal of fMRI volumes where FD exceeds a specified threshold (e.g., 0.2 mm).	A critical step to eliminate the most severely motion-contaminated data points from analysis [19].
Prospective Motion Correction (PMC) [47]	An acquisition technique that uses an MR-compatible camera to track head motion and update the imaging sequence in real-time.	Used during scanning to improve data quality, particularly in populations or studies where high motion is anticipated.
SHAMAN Method [19]	A statistical method (Split Half Analysis of Motion Associated Networks) that assigns a motion impact score to specific trait-FC relationships.	Used for validation to test whether a discovered brain-behavior association is likely spurious due to residual motion artifact.
Global Signal Regression (GSR) [46]	A preprocessing step that removes the global mean BOLD signal from every time point.	Used to mitigate widespread, motion-related artifacts in functional connectivity data.
Physiological Noise Modeling [46]	Modeling of noise from respiration and cardiac cycles using external recordings (e.g., respiratory belt).	Helps disentangle physiological noise from motion artifacts, though it requires additional hardware and data that are not always available.

---

Experimental Protocols and Workflows

Validating Trait-FC Relationships with the SHAMAN Method

The SHAMAN method provides a tailored approach to test if a specific finding is driven by motion [19].

Protocol:

• Split Data: For each participant, split the preprocessed fMRI timeseries into two halves: a high-motion half (volumes with higher FD) and a low-motion half (volumes with lower FD).
• Compute Connectivity: Calculate separate functional connectivity matrices for the high-motion and low-motion halves for each participant.
• Trait-FC Effect: Calculate the trait-FC effect (e.g., correlation) for the full dataset.
• Motion Impact Score: For each functional connection, calculate the difference in correlation structure between the high- and low-motion halves. A significant difference indicates a motion impact.
  • Motion Overestimation: Motion impact score aligned with the trait-FC effect.
  • Motion Underestimation: Motion impact score opposite the trait-FC effect.
• Statistical Testing: Use permutation testing and non-parametric combining across connections to generate a overall motion impact score and p-value for the trait-FC relationship.

Conceptualizing Lagged Motion Artifact

The following diagram illustrates the prolonged impact of motion on the BOLD signal, a key reason why simple censoring is insufficient.

---

Frequently Asked Questions

What are motion overestimation and underestimation in brain-behavior research? In the context of functional MRI (fMRI), motion overestimation occurs when residual head motion artifact causes a false inflation of a true brain-behavior relationship, making it appear stronger than it is. Motion underestimation occurs when motion artifact suppresses or masks a true relationship, making it appear weaker than it is [48] [21] [22].

Why is it crucial to distinguish between these two effects? Distinguishing between them is vital because they require different analytical responses. If motion is causing overestimation, a finding may be a false positive and require more aggressive motion correction. If motion is causing underestimation, a real and important effect might be missed or dismissed, potentially leading to false negatives. For researchers studying traits linked to movement (e.g., in psychiatric disorders), knowing the direction of motion's impact is essential for reporting accurate results [48].

What is motion censoring (scrubbing)? Motion censoring, or scrubbing, is a data processing technique that involves removing individual volumes (timepoints) from an fMRI scan where head motion exceeds a specific threshold, such as a Framewise Displacement (FD) of 0.2 mm. This prevents highly motion-corrupted data points from influencing the analysis [49].

What is the SHAMAN method? SHAMAN (Split Half Analysis of Motion Associated Networks) is a method devised to assign a motion impact score to specific trait-functional connectivity relationships. It can distinguish whether motion is causing an overestimation or an underestimation of a particular brain-behavior effect [48] [21] [22].

---

Troubleshooting Guide: Addressing Motion-Related Artifacts

Problem: High rates of false positive findings in your analysis.

• Potential Cause: Motion-induced overestimation of trait-FC relationships.
• Solution:
  • Implement Motion Censoring: Apply scrubbing with a Framewise Displacement (FD) threshold, for example, FD < 0.2 mm. Research has shown this can drastically reduce the number of traits with significant motion overestimation scores [48].
  • Quantify the Issue: Use a method like SHAMAN to calculate a motion impact score for your key findings. This provides quantitative evidence of whether overestimation is a concern for your specific trait of interest [48] [22].

Problem: True effects appear weak or non-existent; concerns about false negatives.

• Potential Cause: Motion-induced underestimation of trait-FC relationships.
• Solution:
  • Do Not Rely on Censoring Alone: Be aware that motion censoring, while effective against overestimation, may not resolve underestimation. In one large study, censoring did not decrease the number of traits with significant motion underestimation scores [48].
  • Diagnose with SHAMAN: Use the SHAMAN method to check if your non-significant results are being suppressed by motion artifact. A significant underestimation score suggests the true effect might be stronger than measured [48].
  • Explore Advanced Denoising: Consider that standard denoising pipelines may be insufficient. Even after a denoising algorithm reduced motion-related signal variance by 69%, substantial residual motion artifact remained that could lead to underestimation [48].

---

Evidence and Data: The Asymmetric Impact of Censoring

The following data, derived from a large-scale study on the ABCD dataset, quantifies the problem and the asymmetric effect of censoring.

• Study Context: Analysis of 45 traits from n=7,270 participants in the Adolescent Brain Cognitive Development (ABCD) Study after standard denoising [48] [22].
• Method: Application of the SHAMAN method to calculate motion impact scores with and without motion censoring (FD < 0.2 mm) [48].

Table 1: Prevalence of Motion Impact Before and After Censoring

Motion Impact Type	Significant Effects (Before Censoring)	Significant Effects (After Censoring at FD < 0.2 mm)
Overestimation	42% (19 out of 45 traits)	2% (1 out of 45 traits)
Underestimation	38% (17 out of 45 traits)	38% (17 out of 45 traits)

Interpretation: This data clearly shows the asymmetric effect. Motion censoring is highly effective at mitigating false positives from overestimation but does not address the problem of underestimation, which can persist in over a third of traits [48].

Table 2: Key Research Reagents & Analytical Solutions

Item	Function in Research
Framewise Displacement (FD)	A scalar quantity that summarizes head motion between volumes. It is the primary metric for identifying and censoring high-motion volumes [49].
SHAMAN Method	The specific analytical tool that calculates a motion impact score to diagnose overestimation and underestimation for individual trait-FC relationships [48] [22].
ABCD-BIDS Pipeline	A standardized, extensive denoising algorithm that includes global signal regression, respiratory filtering, and motion parameter regression. Serves as a baseline for preprocessing [48].
Motion Censoring (Scrubbing)	The procedural technique of removing high-motion volumes (based on FD) from analysis to reduce artifact [49].

---

Experimental Protocols

Protocol 1: The SHAMAN Methodology for Quantifying Motion Impact

SHAMAN (Split Half Analysis of Motion Associated Networks) operates on the principle that traits (e.g., cognitive scores) are stable during a scan, while motion is a varying state [48].

Workflow:

• Data Split: For each participant's fMRI timeseries, split the data into high-motion and low-motion halves.
• Correlation Calculation: Measure the difference in the correlation structure (functional connectivity) between the two halves.
• Trait-FC Effect Alignment:
  • If the motion impact score aligns with the trait-FC effect's direction, it indicates motion overestimation.
  • If the motion impact score is opposite to the trait-FC effect's direction, it indicates motion underestimation.
• Statistical Significance: Permutation testing and non-parametric combining across brain connections yield a p-value for the motion impact score [48].

Protocol 2: Implementing Motion Censoring in Task or Resting-State fMRI

This protocol outlines the steps for the motion censoring (scrubbing) procedure shown to effectively reduce overestimation [49].

Workflow:

• Calculate Framewise Displacement: Compute the FD for every volume in the preprocessed fMRI timeseries.
• Set Censoring Threshold: Define an FD threshold (common examples include 0.2 mm or 0.5 mm). A stricter threshold (e.g., 0.2 mm) removes more data but is more aggressive against motion artifact [48] [49].
• Identify Volumes to Censor: Flag all volumes where FD exceeds the threshold. It is also common practice to censor one or two volumes before and after the high-motion volume to account for the spin-history effect [49].
• Perform Censoring: In the General Linear Model (GLM) analysis, these flagged volumes are withheld from estimation. They are not included in the calculation of model parameters [49].

---

Key Takeaways for Researchers

• Motion's bias is not uniform: It can artificially inflate or suppress brain-behavior relationships.
• Censoring is not a panacea: While it is a powerful tool against false positives (overestimation), it is ineffective against false negatives (underestimation).
• Diagnosis is critical: Before drawing conclusions, use methods like SHAMAN to diagnose the direction and significance of motion's impact on your specific findings.
• Report motion impact: For full transparency, especially when studying motion-correlated traits, reporting motion impact scores should become a best practice in the field.

FAQs

What are the consequences of incomplete denoising on functional connectivity results?

Incomplete denoising leaves behind two primary types of motion artifacts that systematically bias functional connectivity (FC) estimates. Global artifacts inflate correlation estimates across the entire brain, while distance-dependent artifacts specifically increase short-distance correlations and decrease long-distance correlations [25] [50]. These artifacts create spurious correlations that can be misinterpreted as genuine neural signals.

Even after applying standard denoising pipelines like ABCD-BIDS (which includes global signal regression, respiratory filtering, and motion parameter regression), a significant portion of motion-related variance can persist. One study found that while denoising reduced motion-explained variance from 73% to 23%, this remaining 23% residual variance continued to significantly impact trait-FC relationships [19].

Why are certain traits particularly vulnerable to motion-related confounds?

Certain behavioral and clinical traits are inherently correlated with head motion, creating a built-in confound that standard denoising cannot fully eliminate. For example, individuals with conditions such as attention-deficit/hyperactivity disorder or autism spectrum disorder tend to move more in the scanner [19]. When researchers then compare these groups to neurotypical controls, differences in FC may reflect residual motion artifacts rather than genuine neural differences.

This vulnerability occurs because motion acts as an unmeasured confounding variable that systematically biases results. Studies have shown that after standard denoising, 42% of examined traits had significant motion overestimation scores, while 38% had significant underestimation scores [19]. This demonstrates that residual motion can either inflate or deflate apparent brain-behavior relationships.

How can I determine if my trait of interest is compromised by motion artifacts?

The SHAMAN (Split Half Analysis of Motion Associated Networks) framework provides a method to calculate a motion impact score for specific trait-FC relationships [19]. This approach works by:

• Splitting each participant's fMRI timeseries into high-motion and low-motion halves
• Comparing the correlation structure between these halves
• Calculating whether motion causes overestimation or underestimation of trait-FC effects
• Generating a p-value indicating significance of motion impact

A significant motion overestimation score occurs when the motion impact aligns with the trait-FC effect direction, while a significant underestimation score occurs when it opposes the trait-FC effect [19]. This method specifically addresses trait-specific motion contamination beyond general motion-FC relationships.

What denoising strategies most effectively address different artifact types?

Different denoising strategies target different types of motion artifacts, with combined approaches proving most effective. Research evaluating multiple techniques found that:

Denoising Method	Impact on Global Artifacts	Impact on Spatially Specific Artifacts
FIX (ICA-based)	Limited reduction	Substantial reduction
MGTR (Global signal regression)	Significant reduction	Limited reduction
Censoring high-motion time points	Small reduction	Small reduction
Motion regression	Limited reduction	Limited reduction

The most effective approach combined FIX denoising with mean grayordinate time series regression (MGTR), which addressed both global and spatially specific artifacts [25]. This highlights the importance of using complementary methods rather than relying on a single denoising technique.

What common pitfalls should I avoid in nuisance regression?

Nuisance regression requires careful implementation to avoid introducing new biases while removing artifacts:

• Implement pre-whitening to achieve valid statistical inference of noise model fit parameters [51]
• Incorporate temporal filtering into the noise model to properly account for changes in degrees of freedom [51]
• Use temporal shifting of regressors cautiously with optimization and validation of a single temporal shift [51]
• Regularly assess appropriateness of noise models for each new dataset [51]

Failure to implement these practices can result in inadequate noise removal or the introduction of new statistical biases into the denoised data.

Experimental Protocols & Methodologies

Protocol: Evaluating Denoising Strategy Efficacy

This protocol evaluates the effectiveness of different denoising strategies in reducing motion artifacts, based on methodologies from the Human Connectome Project [25].

Materials:

• Resting-state fMRI data with matched motion parameters
• Processing pipelines for multiple denoising strategies
• Quality control metrics for artifact assessment

Procedure:

• Calculate framewise displacement (FD) for each time point as a measure of head motion
• Apply multiple denoising strategies in parallel to the same dataset:
  • FIX ICA-based denoising
  • Mean grayordinate time series regression (MGTR)
  • Censoring of high-motion time points (FD > 0.2 mm)
  • Motion parameter regression
• Compute functional connectivity matrices for each approach
• Calculate QC-FC correlations between motion and connectivity measures
• Compare motion-group differences between high- and low-motion participants

Validation Metrics:

• Reduction in distance-dependent artifacts
• Reduction in global artifacts
• Differences between high- and low-motion groups
• QC-FC correlation plots

Protocol: Implementing SHAMAN for Trait-Specific Motion Impact Assessment

This protocol assesses whether specific trait-FC relationships are compromised by motion artifacts using the SHAMAN framework [19].

Materials:

• Resting-state fMRI data with associated trait measures
• Framewise displacement calculations
• Processing pipeline for split-half analysis

Procedure:

• For each participant, identify high-motion and low-motion time points based on FD
• Split the timeseries into high-motion and low-motion halves
• Compute separate FC matrices for each half
• Calculate motion impact for each edge as the difference between high-motion and low-motion FC values
• Test for alignment between motion impact and trait-FC effects:
  • Positive alignment indicates motion overestimation of trait-FC effect
  • Negative alignment indicates motion underestimation of trait-FC effect
• Perform permutation testing to establish statistical significance

Interpretation:

• Significant motion overestimation: trait-FC effect is likely inflated by motion
• Significant motion underestimation: trait-FC effect is likely obscured by motion
• Non-significant result: trait-FC effect is likely robust to motion artifacts

Data Presentation

Table 1. Efficacy of Denoising Strategies Across Artifact Types

Denoising Strategy	Global Artifact Reduction	Distance-Dependent Artifact Reduction	Impact on High-Low Motion Group Differences	Key Limitations
FIX (ICA-based)	Limited	Substantial	Limited reduction	Leaves substantial global artifacts
MGTR (Global signal regression)	Significant	Limited	Substantial reduction	Leaves substantial spatially specific artifacts
Censoring (FD < 0.2 mm)	Small	Small	Moderate reduction	Can bias sample by excluding high-motion individuals
Motion parameter regression	Limited	Limited	Limited reduction	Ineffective against complex motion artifacts
Combined FIX + MGTR	Significant	Substantial	Substantial reduction	Increased complexity of implementation

Data synthesized from evaluation of denoising strategies in HCP data [25]

Table 2. Motion Impact on Traits After Different Denoising Levels

Processing Level	Traits with Significant Motion Overestimation	Traits with Significant Motion Underestimation	Example Vulnerable Traits
Minimal processing + no censoring	42% (19/45)	38% (17/45)	Attention, impulsivity measures
ABCD-BIDS denoising (standard)	22% (10/45)	28% (13/45)	Clinical symptom scores
ABCD-BIDS + censoring (FD < 0.2 mm)	2% (1/45)	27% (12/45)	Cognitive performance measures

Data based on analysis of 45 traits in n=7,270 participants from the ABCD Study [19]

Research Reagent Solutions

Table 3. Essential Tools for fMRI Motion Artifact Research

Tool Name	Type	Function	Key Features
FIX (FMRIB's ICA-based X-noiseifier)	Software package	Automated ICA-based denoising	Classifies independent components as signal or noise; trained on manual classifications
SHAMAN	Analytical framework	Quantifies trait-specific motion impact	Split-half analysis; distinguishes overestimation vs. underestimation
CICADA	Automated ICA denoising tool	Comprehensive noise reduction	Uses manual classification guidelines; 97.9% classification accuracy
Res-MoCoDiff	Deep learning model	MRI motion artifact correction	Residual-guided diffusion model; only 4 reverse diffusion steps
JDAC	Iterative learning framework	Joint denoising and motion correction	3D processing; integrates noise level estimation and artifact correction
HALFpipe	Software platform	Pipeline comparison and evaluation	Multi-metric approach; summary performance index for denoising strategies

Workflow Diagrams

Diagram 1: Motion Artifact Impact and Detection Workflow

Diagram 2: Comprehensive Denoising Strategy Implementation

Ensuring Robustness: Validation Techniques and Comparative Analysis of Motion Correction Methods

Troubleshooting Guides & FAQs

Frequently Asked Questions

Q1: My correlation between a brain measurement and a behavioral score is statistically significant, but I am concerned it might be spurious. What is the first thing I should check?

A: The first and most critical step is to visually inspect a scatterplot of your data. Look for outliers, as a single outlying data point can create a false positive correlation or mask a true one [52]. Following this, assess the impact of head motion, as it is a major source of systematic artifact that can induce false brain-behavior relationships [19]. Consider using robust correlation measures like Spearman correlation or skipped correlations, which are less sensitive to outliers [52].

Q2: What is the fundamental difference between Cronbach's Alpha and split-half reliability, and why does it matter for reaction-time tasks?

A: Cronbach's Alpha requires linking individual item scores across participants, which is often not possible in cognitive tasks where trials are presented in random order and error trials are removed. Furthermore, applying Cronbach's Alpha to averaged trial sets for reaction-time difference scores often produces "highly inaccurate and negatively biased reliability estimates" [53]. Split-half methods, in contrast, are designed to work with aggregates of trials and are therefore more suitable for the data structure of typical cognitive tasks [54].

Q3: I use an evidence-accumulation model (e.g., DDM, LBA). I've found a strong negative correlation between boundary separation and non-decision time difference scores across participants. Is this a meaningful finding?

A: Exercise caution, as this is a known spurious correlation. Simulation studies show that a pronounced negative correlation (around r = -0.70 or larger) can emerge between these parameter difference scores even in the absence of any true underlying differences at the population level [55]. This artifact can occur when only the drift rate is truly manipulated between conditions, or when no true differences exist. It is recommended not to base substantive conclusions solely on such correlational patterns without further validation.

Q4: How can I evaluate if my brain-behavior correlation is being driven by head motion?

A: The Split Half Analysis of Motion Associated Networks (SHAMAN) framework is designed specifically for this purpose. It calculates a motion impact score for a specific trait-FC (functional connectivity) relationship by comparing the correlation structure between high-motion and low-motion halves of a participant's fMRI timeseries [19]. A significant motion impact score indicates that the observed trait-FC relationship is likely influenced by residual motion artifact even after standard denoising.

Troubleshooting Guides

Problem: Low reliability of a cognitive task score (e.g., an approach-avoidance bias score). Goal: To accurately estimate the internal reliability of a task score and improve it.

• Step 1: Choose the right splitting method. Avoid simple first-half/second-half or odd-even splits, as they can be confounded with time effects and task design, respectively [54]. Instead, use a resampling method.
• Step 2: Use a permutation-based split-half approach. This involves randomly splitting all trials into two halves thousands of times (e.g., 5,000 replications), calculating the correlation for each split, and then averaging these correlations [53] [54].
• Step 3: Apply the Spearman-Brown correction. The average split-half correlation must be corrected for the shorter test length using a modified Spearman-Brown formula to estimate the reliability of the full task [53].
• Step 4: Ensure stratification. When splitting trials, ensure an approximately equal number of trials from each stimulus type or condition ends up in each half. This prevents confounds with the task design and improves the accuracy of the reliability estimate [53] [54].
• Step 5: Check for stability. The accuracy of the final reliability estimate depends on computing a sufficient number of split-half correlations. Research suggests that for a task with 30 participants and 256 trials, around 5,400 random splits are needed for a stable estimate [53].

Problem: A significant brain-behavior correlation is suspected to be a false positive driven by head motion. Goal: To test whether a specific trait-functional connectivity (FC) finding is spuriously influenced by motion.

• Step 1: Preprocess data with denoising. Apply a standard denoising pipeline (e.g., motion parameter regression, global signal regression, despiking). Acknowledge that this reduces but does not completely eliminate motion artifact [19].
• Step 2: Split timeseries by motion. For each participant, split their resting-state fMRI timeseries into two halves: one with higher framewise displacement (FD) and one with lower FD [19].
• Step 3: Calculate split-half trait-FC effects. Compute the correlation between the trait (e.g., a behavioral score) and every functional connection (edge) separately for the high-motion and low-motion halves.
• Step 4: Compute the motion impact score. For each functional connection, calculate the difference in the trait-FC effect between the two halves (low-motion half vs. high-motion half). Aggregate these differences across connections to get an overall motion impact score [19].
• Step 5: Perform permutation testing. Repeatedly shuffle the assignment of the trait values across participants and re-compute the motion impact score to build a null distribution. Compare your true motion impact score to this null distribution to obtain a p-value [19] [56].
• Step 6: Interpret the direction. A motion impact score that aligns with the trait-FC effect suggests motion causes overestimation. A score in the opposite direction suggests motion causes underestimation [19].

Data Presentation

Table 1: Comparison of Splitting Methods for Estimating Split-Half Reliability [54]

Splitting Method	Description	Controls for Time Effects?	Controls for Task Design?	Risk of Trial-Sampling Error?
First-Second Half	Splits trials by their order in the sequence.	No	No	High (single split)
Odd-Even	Splits trials based on odd vs. even trial numbers.	Yes	No (can be confounded)	High (single split)
Permutated Split-Half	Creates many random splits of trials without replacement.	Yes	Can be combined with stratification	Low (averaged over splits)
Monte Carlo Split-Half	Creates many random splits of trials with replacement.	Yes	Can be combined with stratification	Low (averaged over splits)

Table 2: Motion Impact on Functional Connectivity Traits (ABCD Study Data) [19]

Analysis Condition	Traits with Significant Motion Overestimation	Traits with Significant Motion Underestimation
After denoising (no censoring)	42% (19/45 traits)	38% (17/45 traits)
With censoring (FD < 0.2 mm)	2% (1/45 traits)	38% (17/45 traits)

Experimental Protocols

Protocol 1: Estimating Permutation-Based Split-Half Reliability for a Cognitive Task

• Data Preparation: For each participant, compile all valid trial-level data (e.g., reaction times for correct trials only).
• Stratification: Group trials by their condition (e.g., stimulus category: spider vs. leaf) to prepare for stratified splitting.
• Random Splitting: For a large number of replications (e.g., 5,000):
  • Randomly assign each trial from every condition into one of two halves (Split A and Split B), ensuring the halves are balanced for conditions.
  • Calculate the participant's score (e.g., mean RT difference) separately for Split A and Split B.
• Compute Correlation: For each replication, calculate the correlation coefficient (e.g., Pearson's r) of the scores between Split A and Split B across all participants.
• Average and Correct: Calculate the mean of all split-half correlations. Apply the Spearman-Brown prophecy formula to this mean correlation to estimate the reliability of the full task [53] [54].
  • Spearman-Brown formula: Reliability_full = (2 * r_half) / (1 + r_half)

Protocol 2: SHAMAN Protocol for Trait-Specific Motion Impact Score

• Data Input: Required data for each participant: preprocessed resting-state fMRI timeseries, framewise displacement (FD) timeseries, and a trait measurement (e.g., cognitive test score).
• Split Timeseries by Motion: For each participant, calculate the median FD. Label timepoints with FD above the median as the "high-motion half" and those below as the "low-motion half."
• Calculate Half-Brain Networks: Compute functional connectivity matrices (e.g., correlation matrices between brain regions) separately for the high-motion and low-motion halves.
• Trait-FC Correlation: For each functional connection (edge), calculate the correlation between the trait score and the connectivity strength across participants. Do this separately for the high-motion and low-motion halves, resulting in two trait-FC effect maps.
• Compute Motion Impact Score: For each edge, subtract the trait-FC effect in the high-motion half from the effect in the low-motion half. The distribution of these difference scores across all edges constitutes the motion impact score [19].
• Statistical Significance Testing:
  • Permute the trait values across participants.
  • Recompute the entire motion impact score with the permuted data.
  • Repeat this process thousands of times to build a null distribution.
  • Compare the true motion impact score to this null distribution to obtain a p-value.

Mandatory Visualization

Split-Half Reliability Workflow

Motion Impact Validation with SHAMAN

The Scientist's Toolkit

Table 3: Key Research Reagent Solutions for Robust Analysis

Reagent / Tool	Function / Purpose	Key Consideration
Permutation-Based Split-Half Reliability	Estimates internal consistency of cognitive task scores by averaging thousands of random split-half correlations. More accurate than Cronbach's alpha for RT data [53].	Requires a sufficient number of splits (e.g., ~5,400) for stability and should be combined with stratification by task condition.
SHAMAN (Split Half Analysis of Motion Associated Networks)	Assigns a motion impact score to specific brain-behavior relationships, distinguishing between overestimation and underestimation [19].	Effective even on preprocessed data from large repositories; helps decide if additional motion censoring is needed for a specific trait.
Skipped Correlation	A robust correlation technique that combines multivariate outlier detection with Spearman correlation. Reduces false positives/negatives from outliers [52].	More computationally intensive than Pearson or Spearman, but provides greater confidence by automatically identifying and handling bivariate outliers.
Framewise Displacement (FD)	A scalar quantity summarizing head motion between consecutive fMRI volumes. Used to identify high-motion timepoints [19].	Serves as the primary metric for splitting timeseries in SHAMAN and for implementing motion censoring (e.g., FD < 0.2 mm).
Spearman-Brown Formula	A statistical prophecy formula used to correct a split-half correlation, estimating the reliability of a test if it were doubled in length [53] [54].	Critical final step in split-half reliability analysis; without it, the reliability is underestimated.

SHAMAN Technical Support Center

Troubleshooting Guides & FAQs

This section addresses common challenges researchers face when implementing the Split Half Analysis of Motion Associated Networks (SHAMAN) framework and interpreting its results.

FAQ 1: My motion impact scores are consistently non-significant. Is SHAMAN working correctly?

• Problem: A lack of significant motion impact scores across multiple traits.
• Diagnosis: This can occur for two main reasons:
  • Your denoising pipeline has successfully removed most motion-related variance, leaving little residual artifact for SHAMAN to detect.
  • The traits you are studying are not strongly correlated with head motion, so trait-FC relationships are not susceptible to motion-induced bias.
• Solution:
  • Verify the integrity of your framewise displacement (FD) timeseries data. Ensure it is calculated correctly and matches the preprocessed fMRI data.
  • Run a positive control analysis. Correlate participant's mean FD with a trait known to be motion-correlated (e.g., age in developmental cohorts or inattention symptoms). A significant correlation confirms your motion metric is valid.
  • Check the distribution of your trait data. Restricted range in the trait can reduce the power to detect significant motion impacts.

FAQ 2: How do I interpret a significant "motion underestimation" score?

• Problem: Understanding what it means when motion causes an underestimation of a trait-FC effect.
• Diagnosis: A significant motion underestimation score indicates that residual head motion artifact is systematically weakening or obscuring a true brain-behavior relationship [19]. The observed effect size in your data is likely smaller than the genuine biological effect.
• Solution:
  • Do not dismiss this trait-FC relationship. An underestimation score suggests the association is likely real but is being masked by noise.
  • Apply more aggressive motion censoring (e.g., FD < 0.2 mm) and re-run your primary analysis. The effect size of the trait-FC relationship may increase after reducing motion contamination [19].
  • Report this finding transparently, as it indicates your initial analysis may be conservative.

FAQ 3: The SHAMAN analysis is computationally expensive for my large dataset. Are there alternatives?

• Problem: Long processing times for large cohorts (e.g., n > 5000).
• Diagnosis: SHAMAN involves permutation testing and split-half comparisons across many connections, which is computationally intensive.
• Solution:
  • Start with a hypothesis-driven approach by focusing on specific brain networks of interest for your trait, rather than running a whole-brain analysis.
  • Utilize high-performance computing clusters to parallelize operations across many nodes.
  • As a preliminary check, you can use simpler methods like correlating the spatial map of your trait-FC effects with the spatial map of motion-FC effects [19]. However, note that this method does not distinguish between over- and underestimation.

FAQ 4: Can SHAMAN be used with task-based fMRI data?

• Problem: Applying the SHAMAN framework to task-fMRI paradigms.
• Diagnosis: The standard SHAMAN method is designed for resting-state data, where traits are stable and motion is a transient state. In task-fMRI, the evoked brain response is a structured, time-locked "state."
• Solution:
  • The direct application of SHAMAN to task-fMRI is not currently validated and could be misleading, as it might interpret task-evoked activity as motion-related artifact.
  • For task data, prioritize standard motion correction and denoising, and include mean FD as a nuisance covariate in your general linear model (GLM). The development of a task-fMRI version of SHAMAN is an area for future research.

Experimental Protocols & Methodologies

This section provides detailed, step-by-step protocols for key experiments and analyses involving SHAMAN.

Protocol 1: Calculating the Motion Impact Score with SHAMAN

Objective: To compute a trait-specific motion impact score that quantifies whether residual head motion causes overestimation or underestimation of a brain-behavior relationship.

Materials:

• Preprocessed resting-state fMRI timeseries data.
• Framewise displacement (FD) timeseries for each participant.
• Trait of interest data (e.g., cognitive scores, clinical measures).

Procedure:

• Data Preparation: For each participant, ensure fMRI data is denoised using a standard pipeline (e.g., ABCD-BIDS, which includes global signal regression, motion parameter regression, and despiking) [19].
• Timeseries Splitting: For each participant's fMRI data, split the cleaned timeseries into two halves: a "low-motion" half (timepoints with the lowest FD values) and a "high-motion" half (timepoints with the highest FD values) [19].
• Connectivity Calculation: Calculate two separate functional connectivity (FC) matrices for each participant: one from the low-motion half and one from the high-motion half.
• Trait-FC Correlation: For each of the two halves independently, compute the correlation (across participants) between the trait and the strength of each functional connection (edge). This produces two spatial maps of trait-FC effect sizes: one from low-motion data and one from high-motion data.
• Motion Impact Calculation:
  • For each functional connection, calculate the difference in trait-FC effect sizes (High-motion effect - Low-motion effect). This is the raw motion impact for that connection.
  • To get the overall Motion Overestimation Score, average these differences only for the connections where the direction of the difference aligns with the direction of the full-data trait-FC effect.
  • To get the overall Motion Underestimation Score, average the differences only for the connections where the direction of the difference is opposite to the full-data trait-FC effect [19].
• Statistical Significance: Determine the significance of these scores using non-parametric permutation testing (e.g., by shuffling the trait labels and repeating the process thousands of times to create a null distribution) [19].

Protocol 2: Validating Denoising Pipelines with Motion-FC Correlations

Objective: To quantify the amount of residual motion artifact remaining in fMRI data after applying a denoising pipeline.

Materials:

• Minimally preprocessed and fully denoised fMRI data.
• Participant-level mean framewise displacement (FD).

Procedure:

• FC Calculation: Generate a whole-brain FC matrix for each participant using both the minimally preprocessed and the denoised data.
• Motion-FC Effect: For each functional connection, compute the correlation (across participants) between their mean FD and the strength of that connection. This produces a spatial map of motion-FC effect sizes.
• Variance Explained: For an overall metric, regress the mean FD against the average FC strength (across all connections) for each participant. The proportion of variance explained (R²) by FD in this model indicates the amount of residual motion artifact [19].
• Spatial Correlation: Calculate the spatial correlation (e.g., Spearman's ρ) between the group-average FC matrix and the motion-FC effect matrix. A strong negative correlation indicates that participants who move more have systematically weaker connections, especially in long-range pathways [19].

Data Presentation: Quantitative Comparisons

Table 1: Efficacy of Motion Mitigation Strategies in the ABCD Study Dataset

This table summarizes the performance of different motion correction strategies, as evaluated using the SHAMAN framework on data from the Adolescent Brain Cognitive Development (ABCD) Study (n=7,270) [19].

Motion Mitigation Strategy	Traits with Significant Motion Overestimation	Traits with Significant Motion Underestimation	Key Interpretation
Standard Denoising (ABCD-BIDS) only	42% (19/45 traits)	38% (17/45 traits)	Standard denoising alone leaves a large proportion of traits vulnerable to both false positives and masked true effects.
Standard Denoising + Censoring (FD < 0.2 mm)	2% (1/45 traits)	38% (17/45 traits)	Aggressive censoring effectively controls false positives (overestimation) but is ineffective against motion-induced underestimation of effects.
Residual Variance Explained by FD after Denoising	23% of signal variance	N/A	Denoising reduces but does not eliminate motion-related variance, leaving substantial room for confounding.

Table 2: Comparison of Motion Quantification Methods

This table compares SHAMAN to other common methods for assessing the impact of head motion in functional connectivity studies.

Method	Primary Function	Distinguishes Over/Underestimation?	Key Advantage	Key Limitation
SHAMAN [19]	Quantifies trait-specific motion bias	Yes	Directly tests if motion is biasing a specific brain-behavior relationship; provides directionality of bias.	Computationally intensive; currently designed for resting-state fMRI.
Framewise Displacement (FD)	Quantifies volume-to-volume head motion	No	Simple, widely adopted metric for quantifying gross head motion and censoring bad volumes.	Agnostic to the study's hypothesis; does not directly assess impact on trait associations.
Motion-FC Correlation [19]	Quantifies how motion systematically alters connectivity	No	Provides a spatial map of which connections are most vulnerable to motion artifact.	Does not indicate how this spatial map affects the specific trait-FC relationship under study.
Distance-Dependent Correlation [19]	Measures motion's effect on short- vs. long-range connections	No	Reveals a characteristic signature of motion artifact (increased short-range, decreased long-range FC).	A global measure that may not be sensitive to confounding in studies of specific behavioral traits.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Tools for SHAMAN Analysis

A list of key "research reagents," including software, datasets, and metrics, essential for implementing motion detection methods like SHAMAN.

Item	Function	Example / Note
Framewise Displacement (FD)	A scalar summary of head motion between consecutive brain volumes. Calculated from rotational and translational derivatives [19].	The primary metric for quantifying in-scanner head motion and for defining censoring thresholds (e.g., FD < 0.2 mm).
Denoising Pipeline (e.g., ABCD-BIDS)	A set of algorithms to remove motion and other non-neural signals from fMRI data.	Typically includes motion parameter regression, global signal regression, band-pass filtering, and despiking [19].
High-Motion Population Dataset	A dataset that includes participants who are more likely to move, ensuring a wide range of motion values.	The ABCD Study [19] is a prime example, containing data from over 11,000 children.
SHAMAN Algorithm	The core code that performs the split-half analysis and computes the motion impact score.	Requires implementation in a programming environment like Python or R, and access to high-performance computing for large datasets.
Permutation Testing Framework	A non-parametric statistical method to evaluate the significance of the motion impact score.	Used to create a null distribution by repeatedly shuffing trait labels or fMRI timeseries [19].

Workflow Visualization: SHAMAN Analysis Process

The following diagram illustrates the logical workflow and decision points in the SHAMAN analysis process.

SHAMAN Analysis Workflow

Motion Management Strategy Decision Guide

The following diagram helps researchers choose an appropriate motion management strategy based on their study goals and the results of a SHAMAN analysis.

Motion Management Decision Guide

FAQs: Addressing Motion Artifact Challenges

Q1: What is the primary source of artifact in fMRI signals, and why is it a major concern for large-scale studies like ABCD? Head motion is the largest source of artifact in fMRI signals [19]. It introduces systematic bias into functional connectivity (FC) that is not completely removed by standard denoising algorithms. This is particularly problematic for studies investigating traits associated with motion, such as psychiatric disorders, as it can lead to spurious brain-behavior associations and false positive results [19].

Q2: After standard denoising, how prevalent are motion-related distortions in brain-behavior associations? A method called SHAMAN (Split Half Analysis of Motion Associated Networks) assessed 45 traits from n=7,270 ABCD Study participants. After standard denoising without motion censoring, a significant proportion of traits were impacted by residual motion [19]:

• 42% (19/45) of traits had significant motion overestimation scores.
• 38% (17/45) of traits had significant motion underestimation scores [19].

Q3: Does motion censoring (removing high-motion frames) eliminate these problems? Motion censoring helps but does not completely resolve the issue. Censoring at a common threshold of framewise displacement (FD) < 0.2 mm dramatically reduced significant overestimation to only 2% (1/45) of traits. However, it did not decrease the number of traits with significant motion underestimation scores [19]. This indicates that censoring strategies must be carefully considered, as they can mitigate one type of bias while potentially leaving another unaffected.

Q4: Why is head motion a particularly critical issue for the ABCD Study? The ABCD Study investigates children, who typically exhibit higher levels of in-scanner head motion [57]. Furthermore, levels of head motion have been shown to differ according to demographic factors such as sex, race/ethnicity, and socioeconomic status (SES) [58]. Therefore, data quality procedures that exclude participants or data frames based on motion can lead to differential exclusions across demographic groups, potentially biasing the sample and the study's findings [58].

Q5: What is a key strategy to mitigate motion-related artifacts in functional connectivity analyses? One recommended strategy is to strictly control the degrees of freedom by calculating functional connectivity measures with the exact same amount of data across participants [58]. To aid in this, the ABCD Study has released 5-minute- and 10-minute-trimmed functional connectivity datasets [58]. Researchers are encouraged to use these trimmed measures to mitigate artifacts due to variable data quality.

Quantitative Data on Motion Impact

Table 1: Impact of Residual Motion on Trait-FC Associations in the ABCD Study (n=7,270)

Analysis Condition	Motion Overestimation	Motion Underestimation	Key Finding
After denoising (no censoring)	42% (19/45 traits)	38% (17/45 traits)	High prevalence of both over- and underestimation of effects [19]
With censoring (FD < 0.2 mm)	2% (1/45 traits)	38% (17/45 traits)	Censoring reduces overestimation but not underestimation [19]

Table 2: Effectiveness of Denoising in Reducing Motion-Related Variance

Processing Stage	Signal Variance Explained by Motion	Relative Reduction vs. Minimal Processing
Minimal Processing	73%	Baseline [19]
After ABCD-BIDS Denoising	23%	69% reduction [19]

Experimental Protocol: The SHAMAN Methodology

The Split Half Analysis of Motion Associated Networks (SHAMAN) was developed to assign a motion impact score to specific trait-FC relationships [19]. The protocol is as follows:

• Principle: The method capitalizes on the fact that traits (e.g., weight, intelligence) are stable over the timescale of an MRI scan, whereas motion is a state that varies from second to second [19].
• Procedure: For each participant, the fMRI timeseries is split into high-motion and low-motion halves. SHAMAN measures the difference in the correlation structure between these halves concerning a specific trait [19].
• Interpretation: A significant difference indicates that state-dependent motion impacts the trait's connectivity.
  • A motion impact score aligned with the trait-FC effect indicates motion overestimation.
  • A score opposite the trait-FC effect indicates motion underestimation [19].
• Statistical Testing: Permutation of the timeseries and non-parametric combining across pairwise connections yields a motion impact score with a p-value [19].

SHAMAN Workflow

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Tools for Motion-Resilient fMRI Analysis

Tool / Reagent	Function / Purpose	Application in ABCD
ABCD-BIDS Pipeline	A standardized denoising algorithm for pre-processed ABCD data. Includes global signal regression, respiratory filtering, motion parameter regression, and despiking [19].	Default preprocessing to systematically reduce motion-related artifact [19].
Framewise Displacement (FD)	A scalar quantity that summarizes head motion between volumes. Used to quantify motion and set censoring thresholds [58].	Primary metric for quantifying in-scanner head motion and for frame censoring [19].
FIRMM Software	Frame-wise Integrated Real-time Motion Monitoring software. Assesses head motion in real-time during scan acquisition [57].	Used at Siemens sites to provide operators with feedback, helping to ensure sufficient low-motion data is collected [57].
SHAMAN (Method)	Split Half Analysis of Motion Associated Networks. A novel method to compute a trait-specific motion impact score [19].	Post-processing analysis to diagnose if specific trait-FC relationships are significantly impacted by residual motion [19].
Trimmed FC Datasets	Pre-processed functional connectivity datasets trimmed to a standard amount of data (e.g., 5 or 10 minutes) per subject [58].	Mitigates artifacts caused by variable data quality and degrees of freedom across participants [58].

Frequently Asked Questions (FAQs)

Q1: What is generalizability assessment in neuroimaging and why is it critical? Generalizability assessment is the process of testing whether a predictive model or brain-behavior relationship discovered in one dataset holds true in a separate, independent dataset. This is a critical step for preventing spurious findings and ensuring that results are not driven by dataset-specific quirks or confounds, such as in-scanner head motion. Successful cross-dataset validation increases confidence that a finding reflects a true biological relationship rather than a statistical artifact [59].

Q2: How can I tell if my brain-behavior finding is a false positive caused by head motion? Systematic methods like SHAMAN (Split Half Analysis of Motion Associated Networks) have been developed to assign a "motion impact score" to specific trait-FC relationships. This score distinguishes whether residual motion is causing an overestimation or underestimation of your effect [19] [21]. One study found that after standard denoising, 42% of traits examined showed significant motion overestimation, a problem that can be largely mitigated by rigorous motion censoring [19].

Q3: What are the best practices for motion correction to improve generalizability? A combination of prospective and retrospective correction is recommended.

• Prospective Motion Correction (PmC): This technique adjusts the imaging sequence in real-time to account for head movement. It is particularly effective at reducing spin-history effects and other physical distortions that retrospective methods cannot fix, leading to increased sensitivity and fewer false positives, especially in cases of substantial motion [60].
• Retrospective Motion Correction: This standard approach involves realigning functional volumes during data processing. It is robust and widely available in analysis toolboxes but cannot fully correct all motion-related artifacts [60].
• Motion Censoring: Excluding (or "scrubbing") high-motion volumes from analysis is a highly effective post-hoc step. Censoring at a threshold of framewise displacement (FD) < 0.2 mm has been shown to drastically reduce the number of traits with significant motion overestimation [19].

Q4: My model works well on HCP data but fails on my local dataset. Why? This is a common challenge. The high-quality, standardized data from the Human Connectome Project (HCP) are often not representative of data acquired in typical clinical or research settings. Key factors include:

• Data Quality: HCP data benefits from advanced acquisition protocols, longer scan times, and highly compliant participants, leading to higher signal-to-noise ratio (tSNR) [61].
• Sample Characteristics: Models trained on HCP's carefully selected, healthy young adult cohort may not generalize to populations with different demographics, clinical conditions, or higher levels of in-scanner motion [59] [61].
• Protocol Differences: Variations in scanners, sequences, and preprocessing pipelines can significantly impact model performance.

Q5: What is Connectome Fingerprinting and how is its generalizability? Connectome fingerprinting is a procedure that uses an individual's unique pattern of functional connectivity to identify them from a group. While it achieves high accuracy (up to 90-98%) in high-quality datasets like the HCP, its performance drops significantly in datasets with more standard imaging quality. Furthermore, its specificity—the ability to truly distinguish one individual's connectome from all others—may be lower than initially thought, especially as the size of the dataset grows [61].

Troubleshooting Guides

Problem 1: A Significant Brain-Behavior Relationship Becomes Non-Significant After Cross-Dataset Validation

This indicates that the original finding may not be robust or was influenced by factors unique to the discovery dataset.

Step	Action	Diagnostic Question	Tools & Solutions
1	Check for Confounding Variables	Are there systematic differences in demographics, data quality, or preprocessing between datasets?	Compare participant age, sex, in-scanner motion (mean FD), and tSNR between cohorts [59] [62].
2	Quantify Motion Impact	Could head motion have artificially inflated the effect in the original dataset?	Apply the SHAMAN method or similar to your discovery dataset to calculate a motion impact score [19].
3	Re-effect Model Complexity	Is the model overfitted to noise in the discovery dataset?	Simplify the model. Use regularization techniques. Ensure the number of features (connections) is much smaller than the number of participants [59].
4	Assess Measurement Reliability	Is the behavioral trait measured reliably and consistently across both sites?	Check the test-retest reliability of the behavioral instrument. Ensure task paradigms are identical.

Problem 2: Poor Model Performance When Applying a HCP-Trained Model to a Clinical Dataset

Clinical populations often present with data quality challenges that HCP-trained models have not encountered.

Step	Action	Diagnostic Question	Tools & Solutions
1	Quantify Data Quality Mismatch	Is the tSNR or motion level in the target dataset worse than in HCP?	Calculate and compare temporal Signal-to-Noise Ratio (tSNR) and mean Framewise Displacement (FD) between HCP and your dataset [19] [62].
2	Implement Harmonization	Can I reduce the technical differences between the datasets?	Apply data harmonization techniques like ComBat to remove site-specific effects before applying the model [61].
3	Fine-Tune the Model	Can I adapt the pre-trained model to my specific dataset?	Use transfer learning: take the HCP-derived network model and re-train the final prediction layer on a small subset of your own data [59].
4	Consider a New Model	Is the underlying brain-behavior relationship different in my clinical population?	Build a new model from scratch using data from your target population, if sample size allows [59].

Experimental Protocols for Robust Generalizability

Protocol 1: Assessing Motion Impact with SHAMAN

This protocol helps determine if a specific trait-functional connectivity (FC) association is spuriously driven by head motion [19].

Methodology:

• Data Preparation: Process your resting-state fMRI data with your standard denoising pipeline (e.g., including motion parameter regression, global signal regression).
• Split Timeseries: For each participant, split the preprocessed fMRI timeseries into two halves: a "low-motion" half (volumes with FD below the participant's median FD) and a "high-motion" half (volumes with FD above the median).
• Calculate Trait-FC Effects: Compute the correlation between the trait of interest and FC strength for every connection in the brain, separately for the low-motion and high-motion halves.
• Compute Motion Impact Score: The motion impact score is the difference between the trait-FC effect in the high-motion half and the low-motion half. A positive score aligned with the trait-FC effect suggests motion overestimation; a negative score suggests underestimation.
• Statistical Testing: Use permutation testing (e.g., shuffling the high/low motion labels) to determine if the motion impact score is statistically significant.

Protocol 2: Cross-Dataset Validation with Connectome-Based Predictive Modeling (CPM)

This protocol outlines how to build a predictive model in one dataset and test its generalizability in another [59].

Methodology:

• Feature Selection (in Dataset 1):
  • For each brain connection, calculate its correlation with the behavioral measure (e.g., attention score) across participants in Dataset 1.
  • Retain connections that show a significant positive correlation (this forms the "positive network") and those that show a significant negative correlation (this forms the "negative network").
• Model Building (in Dataset 1):
  • For each participant in Dataset 1, create a summary score by summing the strength of all connections in the positive network and another by summing the strength of all connections in the negative network.
  • Train a linear model (e.g., using linear regression) to predict behavior from these two summary scores.
• Model Application (in Dataset 2):
  • For each participant in the independent Dataset 2, extract the strength of the same positive and negative networks identified in Step 1.
  • Apply the linear model from Step 2 to these summary scores to generate predicted behavior scores for Dataset 2.
• Generalizability Assessment:
  • Calculate the correlation between the model-predicted scores and the actual observed scores in Dataset 2. A significant correlation indicates successful generalizability.

Workflow Diagrams

Generalizability Assessment Workflow

Motion Impact Score with SHAMAN

Table: Key Resources for Generalizable Connectome Research

Resource Name	Function & Purpose	Key Considerations
Human Connectome Project (HCP) [63]	Provides high-quality, publicly available neuroimaging and behavioral data from a large cohort of healthy adults. Serves as a benchmark for model development and a source for pre-trained models.	Data may not be representative of clinical populations or standard acquisition protocols.
ABCD-BIDS Pipeline [19]	A standardized denoising algorithm for fMRI data, including global signal regression, respiratory filtering, and motion parameter regression. Helps ensure consistent preprocessing.	Even after this denoising, residual motion artifact can persist and impact trait-FC associations.
Connectome-Based Predictive Modeling (CPM) [59]	A machine learning method to build predictive models of behavior from functional connectivity data. Designed to be interpretable and to identify relevant networks.	Models can be sensitive to data quality and motion; cross-dataset validation is essential.
SHAMAN (Split Half Analysis of Motion Associated Networks) [19] [21]	A method to assign a trait-specific "motion impact score" to determine if a brain-behavior association is spuriously influenced by head motion.	Critical for validating that a finding is not a motion artifact, especially for motion-correlated traits.
Prospective Motion Correction (PmC) [60]	A hardware/software solution that tracks head position in real-time and adjusts the scanner to correct for motion during data acquisition.	Most effective at reducing spin-history effects and distortions that retrospective correction cannot address.
ComBat	A statistical harmonization tool used to remove site-specific or scanner-specific effects from neuroimaging data, improving the ability to pool data from multiple sources.	Helps mitigate technical variability when applying models across different datasets.

In-scanner head motion is a major source of artifact in neuroimaging data, systematically biasing functional connectivity measures and potentially leading to spurious brain-behavior relationships. Research shows that even after applying standard denoising algorithms, residual motion artifact continues to impact findings. A recent analysis of 45 traits from n=7,270 participants in the Adolescent Brain Cognitive Development (ABCD) Study found that after standard denoising, 42% of traits had significant motion overestimation scores and 38% had significant underestimation scores [19] [21].

This framework provides standardized methods for reporting motion impact to enhance research reproducibility and prevent false positive associations in motion-related research, particularly important when studying populations with motion-correlated traits such as psychiatric disorders.

Quantitative Data on Motion Impact and Correction Efficacy

Table 1: Efficacy of Motion Correction Methods in fNIRS

Correction Method	Mean-Squared Error Reduction	Contrast-to-Noise Ratio Increase	Key Strengths
Spline Interpolation	55% (largest reduction)	Notable improvement	Effective modeling and subtraction of motion periods [64]
Wavelet Analysis	Significant reduction	39% (highest increase)	Effective isolation of motion-related abrupt frequency changes [64]
Principle Component Analysis	Significant reduction	Significant increase	Isolates motion components orthogonal to physiological signals [64]
Kalman Filtering	Significant reduction	Significant increase	Recursive, state-space approach requiring no additional inputs [64]

Table 2: Impact of Motion Censoring on Trait-FC Relationships in fMRI

Processing Stage	Traits with Significant Motion Overestimation	Traits with Significant Motion Underestimation
After standard denoising (no censoring)	42% (19/45 traits)	38% (17/45 traits) [19]
After censoring (FD < 0.2 mm)	2% (1/45 traits)	38% (17/45 traits) [19]
Key Insight	Censoring dramatically reduces false positive inflation	Underestimation effects persist despite aggressive censoring [19]

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Tools for Motion Impact Research

Tool Category	Specific Examples	Function & Application
Motion Tracking Hardware	Accelerometers, Inertial Measurement Units (IMUs), 3D motion capture systems, cameras [14]	Provides reference signal correlated with motion artifacts but not functional response for regression-based correction [64]
Algorithmic Correction Tools	Spline interpolation, wavelet analysis, principle component analysis, Kalman filtering, independent component analysis [64] [14]	Removes motion artifacts from recorded signals based on distinct amplitude and frequency characteristics [64]
Motion Impact Quantification	Split Half Analysis of Motion Associated Networks (SHAMAN), Framewise Displacement (FD) calculation [19]	Assigns motion impact scores to specific trait-FC relationships and distinguishes overestimation vs. underestimation [19]
Advanced Modeling Approaches	3D convolutional neural networks, Fourier domain motion simulation, deep learning frameworks [44]	Retrospective motion correction of structural MRI using simulated artifacts for training [44]

Frequently Asked Questions (FAQs)

What is the difference between motion overestimation and underestimation of trait-FC effects?

Motion overestimation occurs when motion artifact inflates or creates a false positive trait-FC relationship, making effects appear stronger than they truly are. Motion underestimation occurs when motion artifact obscures a genuine trait-FC relationship, making effects appear weaker than they truly are. The SHAMAN method distinguishes these by examining whether the motion impact score aligns with (overestimation) or opposes (underestimation) the direction of the trait-FC effect [19].

Why should I implement additional motion impact analysis when I already regress out motion parameters?

Standard denoising approaches, including motion parameter regression, remove a substantial portion of motion-related variance but leave residual artifact that continues to impact findings. Research shows that even after comprehensive denoising with ABCD-BIDS (including global signal regression, respiratory filtering, and motion parameter regression), 23% of signal variance was still explained by head motion, and significant trait-specific motion impacts remained [19].

Which motion correction method should I use for my fNIRS study?

The optimal method depends on your research goals. If your priority is maximal accuracy in recovering the hemodynamic response, spline interpolation demonstrated the largest reduction in mean-squared error (55%). If maximizing contrast-to-noise ratio is most critical, wavelet analysis produced the highest increase (39%). Each of the four major techniques (PCA, spline, wavelet, Kalman) yields significant improvement over no correction or trial rejection [64].

How does motion censoring affect the detection of true brain-behavior relationships?

Motion censoring involves excluding high-motion fMRI frames from analysis. While censoring at FD < 0.2 mm effectively reduces motion overestimation (from 42% to 2% of traits), it does not decrease motion underestimation, which remained at 38% of traits. This creates a tension between removing enough data to reduce false positives while not excluding so much data as to bias sample distributions or obscure genuine effects [19].

What are the most common causes of motion artifacts in fNIRS recordings?

Motion artifacts in fNIRS primarily result from imperfect contact between optodes and the scalp due to head movements (nodding, shaking, tilting), facial muscle movements (raising eyebrows), body movements (through inertia), and behaviors involving jaw movements (talking, eating). These cause displacement, non-orthogonal contact, or oscillation of the optodes [14].

Experimental Protocols for Motion Impact Assessment

Protocol 1: SHAMAN Motion Impact Analysis

The Split Half Analysis of Motion Associated Networks (SHAMAN) capitalizes on the stability of traits over time by measuring differences in correlation structure between split high- and low-motion halves of each participant's fMRI timeseries [19].

• Data Preparation: For each participant, divide the resting-state fMRI timeseries into high-motion and low-motion halves based on framewise displacement (FD) values.
• Connectivity Calculation: Compute separate functional connectivity matrices for the high-motion and low-motion halves.
• Trait-FC Effect Estimation: Calculate the correlation between the trait and FC for each connection in both the high-motion and low-motion halves.
• Motion Impact Score: For each trait-FC pair, compute the difference in correlation (trait-FC effect) between the high-motion and low-motion halves.
• Statistical Testing: Use permutation testing of the timeseries and non-parametric combining across connections to obtain a p-value for the motion impact score.
• Direction Classification: A motion impact score aligned with the trait-FC effect direction indicates overestimation; a score opposing the trait-FC effect indicates underestimation [19].

Protocol 2: Systematic Evaluation of fNIRS Motion Correction Methods

This protocol follows the approach used in the systematic comparison by Cooper et al. to evaluate different motion correction algorithms [64].

• Dataset Selection: Select resting-state NIRS datasets known to contain motion artifacts, ideally from populations where motion is unavoidable (e.g., clinical populations).
• Motion Artifact Identification: Use established algorithms (e.g., hmrMotionArtifact from HOMER2) to identify periods of motion contamination based on changes in signal amplitude and standard deviation.
• Simulated Activation: Add a known, synthetic hemodynamic response function (HRF) to each dataset at multiple timepoints, simulating functional activation.
• Application of Correction Methods: Apply multiple motion correction techniques (spline, wavelet, PCA, Kalman) to the same artifact-contaminated datasets with simulated signals.
• Recovery Accuracy Assessment: Attempt to recover the average HRF after motion correction and calculate accuracy metrics: mean-squared error (MSE) and contrast-to-noise ratio (CNR) of the recovered HRF compared to the original simulated signal.
• Comparative Analysis: Compare the performance of correction techniques against no correction and against simple trial rejection.

Protocol 3: Deep Learning-Based Structural MRI Motion Correction

This protocol outlines the method for retrospective motion correction of structural MRI using deep learning, as validated by Kaur et al. [44].

• Training Data Generation: Use motion-free T1-weighted structural images and simulate motion artifacts using a Fourier domain motion simulation model.
• Network Training: Train a 3D convolutional neural network (CNN) using the simulated motion-corrupted images as input and motion-free images as targets.
• Model Validation: Quantitatively validate the method on separate test datasets using image quality metrics (peak signal-to-noise ratio, structural similarity index).
• Application to Real Data: Apply the trained model to real motion-affected structural MRI data.
• Downstream Analysis Improvement: Assess improvement in cortical surface reconstruction quality and statistical power in clinical correlations (e.g., cortical thickness analysis in Parkinson's disease) [44].

Workflow Visualization

Motion Impact Assessment Workflow

SHAMAN Method Process

Conclusion

The systematic bias introduced by head motion is not merely a nuisance but a fundamental challenge that can undermine the validity of brain-behavior research and its application in developing biomarkers for drug development. The synthesis of insights from the four intents reveals a clear path forward: researchers must move beyond standard denoising by adopting trait-specific validation frameworks like SHAMAN. Employing rigorous motion censoring thresholds, understanding the asymmetrical effects of mitigation strategies, and comprehensively reporting motion impacts are no longer optional but essential for methodological rigor. Future directions must include the integration of these motion impact assessments as a standard step in the analytical pipeline, the development of real-time correction technologies, and the creation of robust, motion-resistant analytical models. By embracing these practices, the field can significantly enhance the reproducibility and translational potential of neuroimaging, building a more reliable foundation for understanding the brain and developing effective therapeutics.

References

• 1. Behavioral interventions for reducing head motion during ... [https://www.sciencedirect.com/science/article/pii/S1053811918300235]
• 2. Methods to detect, characterize, and remove motion artifact ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811913009117]
• 3. Brain motion networks predict head motion during rest [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2023.1096232/full]
• 4. Head motion: the dirty little secret of neuroimaging in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6306289/]
• 5. Comparison of MRI head motion indicators in ... [https://www.nature.com/articles/s41598-024-79827-9]
• 6. Reducing the effects of motion artifacts in fMRI - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC8934405/]
• 7. Motion Correction Methods for Magnetic Resonance ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7855523/]
• 8. ISMRM24 - Motion Correction: Neuroimaging [https://www.ismrm.org/24/pf/D-12.htm]
• 9. Reducing the Effects of Motion Artifacts in fMRI - PubMed - NIH [https://pubmed.ncbi.nlm.nih.gov/34432631/]
• 10. Rigid real-time prospective motion-corrected three- ... [https://www.sciencedirect.com/science/article/pii/S1053811922003019]
• 11. Motion artifacts in functional near-infrared spectroscopy [https://pmc.ncbi.nlm.nih.gov/articles/PMC3762942/]
• 12. Motion artifacts in functional near-infrared spectroscopy [https://www.sciencedirect.com/science/article/abs/pii/S1053811913004291]
• 13. Removal of high frequency contamination from motion ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7308220/]
• 14. Motion artifacts removal and evaluation techniques for ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2022.878750/full]
• 15. Characterizing motion artifacts in functional near-infrared ... [https://www.sciencedirect.com/science/article/pii/S1746809425007670]
• 16. An evaluation of the efficacy, reliability, and sensitivity ... [https://www.sciencedirect.com/science/article/pii/S1053811917310972]
• 17. Homer2 | fNIRS Systems | NIRS Devices [https://nirx.net/homer2]
• 18. NIRS Toolbox - NIRx Learning Center [https://help.nirx.de/hc/en-us/articles/17400352078108-NIRS-Toolbox]
• 19. Motion impact score for detecting spurious brain-behavior ... [https://www.nature.com/articles/s41467-025-63661-2]
• 20. Motion artifact in studies of functional connectivity [https://pmc.ncbi.nlm.nih.gov/articles/PMC5930165/]
• 21. Motion impact score for detecting spurious brain-behavior ... [https://pubmed.ncbi.nlm.nih.gov/41022827/]
• 22. Motion impact score for detecting spurious brain-behavior ... [https://ohsu.elsevierpure.com/en/publications/motion-impact-score-for-detecting-spurious-brain-behavior-associa/]
• 23. Risk factors for developmental vulnerability [https://pmc.ncbi.nlm.nih.gov/articles/PMC10624014/]
• 24. Marker-Less Video Analysis of Infant Movements for Early ... [https://www.mdpi.com/2075-4418/15/2/136]
• 25. Evaluation of Denoising Strategies to Address Motion- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5105353/]
• 26. Evaluation of denoising strategies for task‐based ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7978116/]
• 27. fMRI denoising pipeline [https://web.conn-toolbox.org/fmri-methods/denoising-pipeline]
• 28. Denoising and unringing before eddy [https://community.mrtrix.org/t/denoising-and-unringing-before-eddy/2223]
• 29. DosenbachGreene/shaman: Split-Half Analysis of Motion ... [https://github.com/DosenbachGreene/shaman]
• 30. How Stable, Really? Traditional and Nonlinear Dynamics ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9265719/]
• 31. Brain Network Reconfiguration and Perceptual Decoupling ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4898667/]
• 32. Motor Function Relating to the Accuracy of Self- ... [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2020.599787/full]
• 33. Motion duration is overestimated behind an occluder in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10184779/]
• 34. Correcting Underestimation and Overestimation in ... [https://www.mdpi.com/2072-4292/14/23/6145]
• 35. Troubleshooting Shaman Resources | Virtuozzo Server Docs [https://www.virtuozzo.com/hybrid-server-docs/7.5/users-guide-troubleshooting-shaman-resources/]
• 36. Balancing Data Quality and Bias: Investigating Functional... | CoLab [https://colab.ws/articles/10.1002%2Fhbm.70094]
• 37. A Set of FMRI Quality Control Tools in AFNI - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC10996659/]
• 38. pyfMRIqc: A Software Package for Raw fMRI Data Quality ... [https://openresearchsoftware.metajnl.com/articles/10.5334/jors.280]
• 39. Guidelines color blind friendly figures [https://www.nki.nl/about-us/responsible-research/guidelines-color-blind-friendly-figures]
• 40. Using FMRIPREP on (somewhat) preprocessed data (ABCD) [https://neurostars.org/t/using-fmriprep-on-somewhat-preprocessed-data-abcd/3397]
• 41. The efficacy of resting-state fMRI denoising pipelines for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12336063/]
• 42. Data processing [https://docs.abcdstudy.org/latest/documentation/imaging/abcc_pipeline.html]
• 43. ABCD-BIDS community collection [https://docs.abcdstudy.org/v/6_0_0/documentation/imaging/abcc_start_page.html]
• 44. Retrospective motion artifact correction of structural MRI ... [https://www.sciencedirect.com/science/article/pii/S1053811921000331]
• 45. The impact of different censoring methods for analyzing ... [https://bmcmedresmethodol.biomedcentral.com/articles/10.1186/s12874-024-02313-3]
• 46. Identifying and characterizing systematic temporally- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5857478/]
• 47. Prospective Motion Correction of fMRI: Improving the quality of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7238750/]
• 48. Motion impact score for detecting spurious brain-behavior ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12479937/]
• 49. Statistical improvements in functional magnetic resonance ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3895106/]
• 50. Spurious but systematic correlations in functional connectivity ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3254728/]
• 51. Potential pitfalls when denoising resting state fMRI data ... [https://www.sciencedirect.com/science/article/pii/S1053811916307480]
• 52. Improving standards in brain-behavior correlation analyses [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2012.00119/full]
• 53. Reaction-time task reliability is more accurately computed ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12000231/]
• 54. Methods to split cognitive task data for estimating split-half ... [https://link.springer.com/article/10.3758/s13423-021-01948-3]
• 55. A spurious correlation between difference scores in ... [https://link.springer.com/article/10.3758/s13428-022-01956-8]
• 56. Cross-validation and permutations in MVPA [https://www.sciencedirect.com/science/article/pii/S1053811921004225]
• 57. Image processing and analysis methods for the Adolescent ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6981278/]
• 58. Overview [https://docs.abcdstudy.org/v/6_0_0/documentation/imaging/]
• 59. A generalizable connectome-based marker of in-scan ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10183743/]
• 60. Prospective Motion Correction in Functional MRI - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC5427003/]
• 61. Evaluating the replicability, specificity, and generalizability ... [https://www.sciencedirect.com/science/article/abs/pii/S1053811917305840]
• 62. On the Definition of Signal-To-Noise Ratio and Contrast- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3819355/]
• 63. HCP Young Adult - Connectome - Publications [https://www.humanconnectome.org/study/hcp-young-adult/publications]
• 64. A Systematic Comparison of Motion Artifact Correction ... - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC3468891/]