Model-Based Neuroimaging with the General Linear Model

77 concepts73 papers1 proposal

When we record brain activity using functional neuroimaging techniques like fMRI or fNIRS, we do not get a direct readout of thoughts or computations. Instead, we obtain a massive, noisy time series of blood oxygenation levels (the BOLD signal) across tens of thousands of spatial units called voxels. The fundamental challenge of cognitive neuroimaging is mapping these continuous, physical signals back to discrete experimental events or internal cognitive states.

The primary tool for solving this mapping problem is the General Linear Model (GLM), implemented within a framework known as Statistical Parametric Mapping (SPM). By treating the brain's response as a linear combination of predictable signals, the GLM allows researchers to localize where specific computations occur.

The Voxel-Wise Regression Framework

The core intuition behind the GLM in neuroimaging is to treat every single voxel in the brain as an independent linear regression problem. As Hongtu Zhu et al. explain in Statistical Learning Methods for Neuroimaging Data Analysis with Applications, this massively univariate approach fits a separate model to the time series of each spatial unit to identify significant regional effects.

For a single voxel, the observed BOLD signal over time ($Y$) is modeled as:

$$Y = X\beta + \epsilon$$

Where:

  • $Y$ is a vector representing the voxel's blood oxygenation level at each time point.
  • $X$ is the design matrix, containing the hypothesized timelines of our experimental conditions (the regressors).
  • $\beta$ is a vector of weights (beta values) that quantify how much the voxel's signal tracks each regressor.
  • $\epsilon$ represents residual noise.

By solving for $\beta$ using ordinary least squares, we can determine whether a voxel significantly increases its activity during a specific task condition.

Modeling the Hemodynamic Response

We cannot simply insert a binary timeline (e.g., $1$ when a stimulus is on, $0$ when it is off) directly into the design matrix. Neural activity triggers a metabolic demand that causes blood flow to change slowly and smoothly over several seconds—a process described by the Hemodynamic Response Function (HRF).

To create realistic predictors for the design matrix, the raw experimental timeline must be mathematically convolved with a canonical HRF. Vince D. Calhoun et al. demonstrate in fMRI Activation in a Visual-Perception Task: Network of Areas Detected Using the General Linear Model and Independent Components Analysis that this convolution transforms discrete task events into a smooth curve that matches the physical latency of blood flow. The design matrix also includes nuisance regressors—such as head motion parameters or respiration rates—to ensure that the estimated beta weights reflect true neural tracking rather than motion artifacts.

Task Timeline:   [__|-|___|-|___]  (Discrete events)
                       *
Canonical HRF:   [  _/\_        ]  (Delayed blood flow response)
                       =
Regressor (X):   [___/‾\_/‾\____]  (Convolved predictor for GLM)

Isolating Cognitive Functions via Contrasts

Once the beta weights are estimated for every voxel, researchers use statistical contrasts to isolate specific cognitive processes. Because the brain is never truly "off," simply looking at absolute activation during a task is uninformative. Instead, we must subtract the activation of a control condition from an experimental condition.

For example, to isolate the cognitive effort required to understand speech in difficult environments, Patrick C.M. Wong et al. in Neural bases of listening to speech in noise applied a contrast that directly subtracted the BOLD response of listening to speech in quiet from speech in noise. Similarly, Matthew M. Botvinick et al. in Viewing facial expressions of pain engages cortical areas involved in the direct experience of pain designed explicit contrasts to identify brain regions that are mutually active during both the direct experience of pain and the passive observation of someone else's pain.

For more complex experimental setups, factorial designs allow researchers to separate overlapping cognitive variables. In Human hippocampal processing of environmental novelty during spatial navigation, Raphael Kaplan et al. utilized a factorial GLM to successfully disentangle the neural responses to environmental novelty (the spatial layout) versus object novelty (items within that layout). The resulting t- or F-statistics across all voxels are assembled into a 3D map—the Statistical Parametric Map—highlighting localized clusters of significant activation.

The Computational Leap: Model-Based fMRI

Traditional GLM analysis relies on categorical regressors (e.g., "Condition A" vs. "Condition B"). However, many cognitive processes are dynamic and hidden; for instance, when a person learns from rewards, their internal sense of value updates on a trial-by-trial basis.

This limitation motivated the development of model-based neuroimaging and computational parametric modulation. Instead of using raw task events as regressors, researchers first fit a mathematical model (such as a reinforcement learning algorithm) to the participant's behavioral choices. The model uncovers latent variables—such as expected values, prediction errors, or model entropy—that change continuously throughout the task. These trial-by-trial computational values are then convolved with the HRF and inserted as continuous regressors in the GLM design matrix.

Behavioral Choices ---> Fit Computational Model ---> Latent Variables (e.g., Prediction Error)
                                                            |
                                                   Convolve with HRF
                                                            |
                                                            v
                                                GLM Parametric Regressor

The authors of Different Cortico-Basal Ganglia Loops Specialize in Reward Prediction at Different Time Scales leveraged this approach by using reinforcement learning value functions as parametric regressors. This allowed them to discover that parallel circuits within the human basal ganglia specialize in predicting rewards across distinct, topographically organized timescales.

Similarly, Aaron M. Bornstein et al. in Cortical and Hippocampal Correlates of Deliberation during Model-Based Decisions for Rewards in Humans used computational parameters to track how the brain simulates potential futures during deliberation, mapping specific algorithmic processes directly to the hippocampus. To isolate the neural mechanics of decision-making, Tommy C. Blanchard in Pure correlates of exploration and exploitation in the human brain utilized computational regressors representing belief updates to cleanly separate the neural correlates of information-gathering (exploration) from reward-seeking (exploitation).

Parameter Robustness and Identifiability

A common concern in model-based neuroimaging is whether the resulting brain maps are overly sensitive to the exact parameters (such as the learning rate $\alpha$) extracted from behavior. If the behavioral parameter estimation is slightly inaccurate, does the subsequent GLM analysis fail?

In Is Model Fitting Necessary for Model-Based fMRI?, Yael Niv demonstrates that model-based fMRI is surprisingly robust to parameter misspecification. Because the latent variables generated by slightly different learning rates remain highly correlated over time, the convolved regressors look remarkably similar. Consequently, the GLM can still successfully identify the brain regions tracking those computational variables even if the underlying behavioral model is not perfectly tuned.

However, Yael Niv notes a critical trade-off: while this robustness makes model-based fMRI highly accessible and forgiving of noisy behavioral data, it also means that fMRI signals are often unable to adjudicate between competing, closely related computational theories. The strength of the GLM lies in localizing where a specific class of mathematical computation occurs, rather than fine-tuning the exact parameters of the algorithm itself.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_statistical_parametric_mapping_0359d13d,
  title = {statistical parametric mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/statistical-parametric-mapping/rdr_0359d13d57b743cc8e277c4102654020},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_statistical_parametric_mapping_0359d13d_v1,
  title = {statistical parametric mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/statistical-parametric-mapping/rdr_0359d13d57b743cc8e277c4102654020/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake