Consider a classic problem in cognitive neuroscience: how does the physical structure of the human brain relate to behavior? On one hand, we can extract thousands of features from neuroimaging data, such as cortical thickness or functional connectivity matrices. On the other hand, we can collect extensive behavioral batteries covering memory, impulsivity, and clinical symptoms.
If we try to connect these two domains using classical univariate statistics—such as correlating every single brain region with every single behavioral score independently—we run into a wall. First, we face a massive multiple-testing problem, requiring severe statistical corrections that destroy our power to detect subtle effects. Second, and more fundamentally, univariate methods assume that features act in isolation. In reality, the brain operates as an integrated network, and behaviors co-vary along complex psychological dimensions.
To capture these system-level interactions, researchers turn to multivariate latent variable modeling. Instead of asking how one brain region relates to one symptom, these methods look at entire systems at once, mapping many-to-many relationships between distinct high-dimensional datasets.
The Core Mechanism: Projecting for Alignment
The two primary mathematical workhorses for this alignment are Canonical Correlation Analysis (CCA) and Partial Least Squares (PLS).
Imagine two high-dimensional data matrices: $X$ (e.g., brain features) and $Y$ (e.g., behavioral traits). Each row represents a subject, and each column represents a feature. We want to find a way to compress the columns of $X$ into a single composite score $z_x$, and the columns of $Y$ into a single composite score $z_y$, such that $z_x$ and $z_y$ are as tightly linked as possible across all subjects.
Geometrically, this means finding a weight vector $w$ for $X$ and a weight vector $v$ for $Y$ to create linear combinations:
$$z_x = Xw \quad \text{and} \quad z_y = Yv$$
The latent variables $z_x$ and $z_y$ represent a shared "axis of variation" or a "latent dimension" across the two datasets. The difference between CCA and PLS lies entirely in what objective function they optimize:
-
Canonical Correlation Analysis (CCA) maximizes the correlation between the projected scores: $$\max_{w, v} \text{Corr}(Xw, Yv)$$ As Hongtu Zhu et al. explain in Multivariate Network-Level Approach to Detect Interactions between Large-Scale Functional Systems, CCA captures holistic interactions by analyzing the full internal structure of each set rather than collapsing them into a single representative mean.
-
Partial Least Squares (PLS) maximizes the covariance between the projected scores: $$\max_{w, v} \text{Cov}(Xw, Yv)$$ Because covariance is the product of correlation and the variances of the individual components, PLS balances finding a strong relationship between the datasets with ensuring that the latent variables capture a large amount of variance within each individual dataset. This makes PLS highly robust when dealing with collinearity or noisy features, as noted by the authors of Changes in white matter microstructure in the brain of patients with inflammatory bowel disease are associated with abdominal pain.
Uncovering the Brain-Behavior Connectome
This framework has fundamentally shifted how we understand neuropsychiatric and cognitive phenotypes. Instead of relying on rigid diagnostic categories, multivariate alignment allows researchers to discover continuous, transdiagnostic dimensions of variation.
For instance, Valeria Kebets et al. applied multivariate latent variable modeling in Somatosensory-Motor Dysconnectivity Spans Multiple Transdiagnostic Dimensions of Psychopathology to map whole-brain functional connectivity maps against comprehensive behavioral batteries. They extracted underlying dimensions that span across healthy individuals and patients, revealing that networks like the somatosensory-motor network act as critical hubs linked to general psychopathology and cognitive dysfunction. Similarly, R. Nathan Spreng et al. used latent variable analysis in The default network of the human brain is associated with perceived social isolation to maximize the covariance between the entire connectome and behavioral measures, identifying a distinct global signature in the brain's default network associated with perceived loneliness.
The power of this approach is its ability to handle structural data as well. Tom Fletcher et al. combined high-dimensional "deformation momenta" (the mathematical forces needed to warp a standard brain model into an individual's anatomy) with PLS in Multivariate Statistical Analysis of Deformation Momenta Relating Anatomical Shape to Neuropsychological Measures to isolate specific patterns of structural change—like hippocampal shrinkage—that directly covary with clinical cognitive decline. By looking at the system holistically, Jingwei Li et al. used CCA in Topography and behavioral relevance of the global signal in the human brain to demonstrate that the spatial topography of the fMRI global signal—often dismissed as noise—is tightly aligned with an individual's cognitive abilities and life outcomes.
Cross-Domain Versatility
While brain-behavior mapping is a dominant application, the underlying mathematics of CCA and PLS are entirely domain-agnostic. Any problem characterized by two complex, interacting systems can be framed as a latent variable alignment problem.
- Metagenomics: Prianka V. Patel et al. utilized CCA in Analysis of membrane proteins in metagenomics: Networks of correlated environmental features and protein families to discover many-to-many relationships between environmental variables (like pollution or shipping intensity) and microbial membrane protein abundances, showing how ecological features collectively shape biological functions.
- Pharmacology: Suleiman A. Khan et al. applied CCA in Comprehensive data-driven analysis of the impact of chemoinformatic structure on the genome-wide biological response profiles of cancer cells to 1159 drugs to bridge the gap between a drug's 3D chemical structure and the genome-wide biological response it triggers in cancer cells, extracting latent components that map specific chemical features to distinct cellular stress outcomes.
- Speech and Signal Processing: Elizabeth Godoy et al. employed CCA in Canonical Correlation Analysis and Prediction of Perceived Rhythmic Prominences and Pitch Tones in Speech to find the principal dimensions along which acoustic speech features (duration, pitch, intensity) align with human perceptions of prosody. In cellular neuroscience, Evan Archer et al. used CCA in Canonical correlations reveal co-variability between spike trains and local field potentials in area MT) to find canonical variables that maximize the correlation between simultaneously recorded local field potentials (LFPs) and spike trains, proving that these two distinct signal types share a low-dimensional, co-varying information channel.
- Sociology: Tianjun Liu used CCA in Research on Social Network and Farmers' Safety Production Behavior to evaluate how a cluster of informal social network factors collectively influences a cluster of agricultural safety behaviors, capturing complex interdependencies that simple regressions miss.
Overcoming High-Dimensional Overfitting
When applying CCA or PLS to modern datasets, researchers frequently encounter the "curse of dimensionality." If the number of features in $X$ or $Y$ greatly exceeds the number of samples ($p > n$), standard CCA will overfit, finding spurious linear combinations that correlate perfectly in the training data but fail to replicate.
To maintain stability in high-dimensional spaces, Eliana Nicolaisen‐Sobesky et al. utilized regularized CCA in A cross-cohort replicable and heritable latent dimension linking behaviour to multi-featured brain structure, which adds a penalty term to the weights to prevent overfitting and ensure that the discovered latent dimensions are heritable and replicable across independent cohorts. Another strategy is to perform initial dimensionality reduction; Pablo Andrés-Camazón et al. combined principal component analysis (PCA) with CCA in Neurobiology-based cognitive biotypes using multi-scale intrinsic connectivity networks in psychotic disorders to reduce high-dimensional neural connectivity data before mapping it to cognitive performance.
Furthermore, real-world relationships are not always linear. To capture complex, non-linear interactions, researchers have extended this framework using deep learning. Myungjong Kim et al. introduced Deep CCA in Multiview Representation Learning via Deep CCA for Silent Speech Recognition, using deep neural networks to extract highly non-linear features from articulatory movements and acoustic signals before projecting them into a maximally correlated shared space. Finally, when data comes from more than two sources, multi-set canonical correlation analysis can be deployed. For instance, Bhaskar Ray et al. used a multi-set approach in Adolescent brain maturation associated with environmental factors: a multivariate analysis to extract shared variance across multiple independent domains simultaneously, linking environmental stressors, neighborhood demographics, and adolescent brain maturation.
Key Foundational Work
Statistical correlation analysis between lip contour parameters and formant parameters for Mandarin monophthongs (2008) — contributed 1 concept elements to this direction.
Multivariate Statistical Analysis of Deformation Momenta Relating Anatomical Shape to Neuropsychological Measures (2010) — contributed 1 concept elements to this direction.
Multivariate Network-Level Approach to Detect Interactions between Large-Scale Functional Systems (2010) — contributed 1 concept elements to this direction.