🧠 neuroscience

Disentangling Brain-Psychopathology Associations: A Systematic Evaluation of Transdiagnostic Latent Factor Models

This study utilizing two large developmental cohorts found that while transdiagnostic latent factor models can distinguish more distinct neural signatures between psychopathology dimensions, they do not systematically improve the reliability or strength of brain-psychopathology associations compared to traditional summary scores, suggesting that fundamental limits in explainable symptom variance may exist regardless of phenotypic modeling approaches.

Original authors: Gell, M., Hoffmann, M. S., Moore, T. M., Nikolaidis, A., Gur, R. C., Salum, G. A., Milham, M. P., Langner, R., Mueller, V. I., Eickhoff, S. B., Satterthwaite, T. D., Tervo-Clemmens, B.

Published 2026-02-16

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Gell, M., Hoffmann, M. S., Moore, T. M., Nikolaidis, A., Gur, R. C., Salum, G. A., Milham, M. P., Langner, R., Mueller, V. I., Eickhoff, S. B., Satterthwaite, T. D., Tervo-Clemmens, B.

Original paper dedicated to the public domain under CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to understand a very complex, noisy radio signal. The signal represents a person's mental health struggles (anxiety, depression, attention issues, etc.). For a long time, scientists have tried to tune into this signal by grouping symptoms into neat categories, like "Anxiety Station" or "Depression Station."

However, real life is messy. People often have symptoms that overlap across these stations. To fix this, some scientists proposed a new way of listening: instead of just tuning into specific stations, they tried to build a "Super-Receiver" that separates the general static (a general tendency to have mental health issues) from the specific songs (unique symptoms like just anxiety or just aggression). They called this the "Bifactor Model."

The big question this paper asked was: Does this fancy "Super-Receiver" actually help us see the connection between the mind and the brain better than the old, simple way of just adding up the symptoms?

Here is the breakdown of what the researchers found, using some everyday analogies:

1. The Setup: Two Ways to Measure the "Noise"

The researchers used data from thousands of teenagers. They looked at their brains using MRI scans (like taking a high-res photo of the brain's wiring) and compared it to how parents described their kids' behavior.

They tested two approaches:

The "Simple Sum" (Summary Scores): Imagine you have a basket of fruit. You just count the total number of apples, oranges, and bananas to get a "Fruit Score." This is the traditional way: adding up all the symptoms to get a total number.
The "Fancy Filter" (Latent Factor Models): Imagine you use a machine to separate the fruit. It isolates the "Fruit-ness" (the general quality of being fruit) from the specific types (just the apple-ness, just the banana-ness). This is the complex statistical model the researchers wanted to test.

2. The Big Surprise: The Fancy Filter Didn't Make the Picture Clearer

The researchers hoped that by using the "Fancy Filter," they would find stronger, clearer links between the brain and the behavior. They thought, "If we clean up the data and remove the noise, the brain patterns should pop out more clearly."

The Result? Not really.

The Analogy: Imagine trying to hear a whisper in a crowded room. The researchers thought that if they used noise-canceling headphones (the fancy model), they would hear the whisper much better. But it turned out that the whisper was just too quiet to begin with. Whether they used the headphones or just cupped their hands (the simple sum), the volume of the whisper (the connection between brain and behavior) was almost exactly the same.
The Finding: The complex models did not predict brain activity any better than the simple "add-up-the-symptoms" scores. In fact, for some specific symptoms, the fancy models were actually less reliable over time than the simple sums.

3. The "Ceiling" Effect: Why Didn't It Work?

Why didn't the fancy math help? The authors suggest there is a glass ceiling.

The Analogy: Imagine you are trying to predict the weather by looking at a single cloud. No matter how perfectly you measure that cloud (even if you use a super-advanced satellite), you can only predict so much about the storm because the cloud itself doesn't hold all the information about the storm.
The Reality: The brain features they measured (like the thickness of the brain or how parts talk to each other) simply don't contain enough information to explain complex human behavior. The "signal" from the brain is just too weak to be boosted by better math. The problem isn't the math; it's that the brain data itself has a limit on how much it can tell us about mental health.

4. One Small Win: The "Fancy Filter" Did Separate the Colors

While the fancy models didn't predict better, they did something interesting: they made the different types of symptoms look more distinct from each other.

The Analogy: Imagine you have a bucket of mixed paint (Red, Blue, and Yellow).
- Simple Sum: You just say, "It's a muddy brown bucket."
- Fancy Filter: You manage to separate the paints so the Red looks very red, and the Blue looks very blue.
The Finding: The "Fancy Filter" models showed that the brain patterns for "General Anxiety" looked different from "General Aggression." The simple sums made them all look like a muddy mix. So, while the fancy models didn't predict the future better, they did a better job of showing us that different problems might have slightly different brain signatures.

5. The Bottom Line

The paper concludes that you can't fix a weak signal just by using a better calculator.

Old Belief: "If we just use better statistical models to organize symptoms, we will finally crack the code of the brain."
New Reality: "The code is still hard to crack because the brain data we have right now isn't detailed enough. We need better ways to measure the symptoms themselves (maybe asking better questions, looking at the environment, or tracking development over time) before we can expect the brain scans to make more sense."

In short: Using complex math to sort mental health symptoms is like trying to get a better photo of a blurry object by using a more expensive camera lens. If the object itself is blurry (or the lighting is bad), a better lens won't help. We need to improve the lighting (better symptom assessment) first.

1. Problem Statement

Psychiatric research faces significant challenges due to the high heterogeneity within diagnostic categories and substantial comorbidity between disorders. Traditional categorical diagnoses often fail to capture the underlying structure of psychopathology. To address this, researchers have proposed transdiagnostic latent factor models (e.g., bifactor models and the Hierarchical Taxonomy of Psychopathology, HiTOP) to separate shared variance (general psychopathology or "P-factor") from unique variance (specific symptom domains).

While these models are theoretically superior for improving measurement precision and reliability, it remains unclear whether they translate into stronger or more distinct brain-psychopathology associations compared to standard, simple summary scores (e.g., total problem scores or domain sums). Previous studies have not systematically compared the predictive utility of latent factor scores against summary scores using large-scale neuroimaging data, leaving open the question of whether psychometric refinements alone can overcome the small effect sizes typical in brain-behavior studies.

2. Methodology

Datasets

The study utilized two large, diverse developmental cohorts:

Adolescent Brain Cognitive Development (ABCD) Study: A US-based longitudinal cohort ( $N=10,897$ for phenotypic modeling; $N=6,572$ for neuroimaging analysis).
Brazilian High-Risk Cohort (BHRC): A school-based community cohort enriched with psychiatric symptoms ( $N=771$ ).

Phenotypic Modeling

The Child Behavior Checklist (CBCL) was used as the primary phenotypic measure. The authors compared three scoring approaches:

Summary Scores: Standard, equally weighted sums of items (e.g., Total Problems, Internalizing, Externalizing, Attention).
Correlated Factor Models: 11 different model solutions where specific factors are allowed to correlate.
Bifactor Models: 11 model solutions partitioning variance into a general "P-factor" and orthogonal (uncorrelated) specific factors.

Neuroimaging Data

Structural MRI (sMRI): Cortical thickness derived from T1-weighted images.
Resting-State fMRI (rs-fMRI): Functional connectivity (Pearson correlation) between 360 cortical regions (HCP atlas) and 19 subcortical regions.
Preprocessing: Standardized pipelines (ABCD-BIDS, DCAN BOLD) including motion correction, denoising, and bandpass filtering.

Analytical Approach

Reliability Assessment: Test-retest reliability (correlation and ICC) and internal consistency ( $\omega$ , $\omega_H$ ) were calculated across baseline and follow-up sessions.
Predictive Modeling: Linear ridge regression (scikit-learn) was used to predict psychopathology scores from whole-brain neuroimaging features.
- Validation: Nested 2-fold cross-validation with matched discovery/replication samples to prevent data leakage.
- Confounds: Sex was regressed out within training folds.
- Performance Metrics: Pearson correlation ( $r$ ) and Coefficient of Determination ( $R^2$ ).
Neural Correlate Comparison: Feature weights were Haufe-transformed to improve interpretability. Similarity between models was assessed via Spearman correlation of feature weight matrices, validated using spin tests (permutation on the cortical surface) to account for spatial autocorrelation.

3. Key Contributions

Systematic Benchmarking: This is one of the first studies to directly compare bifactor, correlated factor, and summary scores regarding their reliability and ability to predict brain structure/function in large-scale samples.
Decoupling Reliability from Prediction: The study challenges the assumption that higher reliability (a psychometric property) automatically leads to stronger brain-behavior associations.
Orthogonalization Insights: It demonstrates that while bifactor models do not improve prediction accuracy, they do improve the separability of neural correlates between constructs due to the orthogonal nature of the factors.

4. Key Results

Reliability

Comparable Reliability: Summary scores and corresponding latent factor scores showed largely comparable test-retest reliability.
Exception: Bifactor-derived specific factors (e.g., specific internalizing or externalizing scores) showed lower reliability than their summary score counterparts or correlated factor scores.
P-Factor: The general P-factor showed high reliability, comparable to the Total Problems summary score.

Prediction Accuracy

No Advantage for Latent Factors: There was no consistent evidence that latent factor models (bifactor or correlated) strengthened the prediction of brain imaging data compared to summary scores.
Performance Ceiling: Prediction accuracies were low but significant for functional connectivity ( $r \approx 0.10$ , $R^2 \approx 0.01$ ) and near-chance for cortical thickness.
Reliability vs. Prediction: While higher reliability generally correlated with better prediction within a specific construct, higher reliability did not translate to better prediction across different constructs (e.g., the highly reliable P-factor was not predicted better than less reliable specific factors).

Neural Correlates and Similarity

High Similarity: The neural signatures (feature weights) for corresponding constructs (e.g., P-factor vs. Total Problems) were highly similar ( $\rho > 0.90$ ). Both were driven by connectivity within and between the Default Mode Network (DMN), Frontoparietal Network (FPN), and Cingulo-Opercular Network (CO).
Bifactor Separability: A key finding was that bifactor models displayed more distinct neural signatures between general, internalizing, and externalizing dimensions compared to summary or correlated factor models. This is attributed to the orthogonalization of factors in bifactor models, which removes shared variance and forces specific factors to rely on unique neural patterns (e.g., bifactor internalizing relied more on DMN-CO connectivity, whereas summary scores relied on DMN-FPN).

5. Significance and Implications

Limits of Psychometric Refinement: The study suggests that simply re-parameterizing symptom inventories (moving from summary scores to latent factors) is insufficient to substantially improve the predictive power of psychiatric neuroimaging. The "ceiling" on prediction accuracy appears to be driven by the fundamental limits of how much brain variance can explain symptom variance, rather than measurement noise.
Reliability $\neq$ Validity: High reliability (precision) does not guarantee high construct validity (alignment with biological reality). A measure can be highly reliable but still fail to capture the specific variance relevant to brain imaging.
Utility of Bifactor Models: While they do not boost prediction magnitude, bifactor models offer discriminant validity by providing more separable neural correlates for specific symptom domains. This may be crucial for understanding distinct neurobiological mechanisms of specific disorders, even if the overall predictive power remains low.
Future Directions: The authors argue that future improvements in brain-psychopathology associations likely require richer phenotypic assessments (incorporating environmental exposures, developmental context, and multi-modal data) rather than just more sophisticated statistical modeling of existing symptom checklists.

Conclusion: The paper concludes that while transdiagnostic latent factor models are valuable for theoretical disentanglement and improving the separability of neural correlates, they do not inherently solve the "small effect size" problem in psychiatric neuroimaging. Improving the quality and breadth of phenotypic data remains the most promising path forward.