Model-based fMRI

Standard fMRI analysis uses GLMs, so most fMRI analyses are technically "model-based". But that's not what I am going to write about today. Instead, I will be talking about model-based fMRI where one combines computational models of cognition with fMRI to find neural correlates of different cognitive processes. I only hope to give an introduction to the topic. For a more detailed treatment, one can turn to the following two resources:

Daw, Nathaniel D. "Trial-by-trial data analysis using computational models."Decision making, affect, and learning: Attention and performance XXIII 23 (2011): 3-38

and 

O'Doherty, John P., Alan Hampton, and Hackjin Kim. "Model‐Based fMRI and Its Application to Reward Learning and Decision Making." Annals of the New York Academy of Sciences 1104.1 (2007): 35-53.

What is model-based fMRI?

For many years, psychologists have built computational models to study human cognition. These models describe cognitive processes in terms of algorithms (variables and operations). They are then fit to behavioral data collected from real human participants to assess if the cognitive processes posited by the model are indeed descriptive of participants' mental operations. 

Model-based fMRI is an extension of the computational modeling of cognition. Instead of fitting to behavioral data, the computational models are fitted to fMRI data time-series. Finding neural activity that encodes the cognitive variables assumed by the model would provide additional validity of the models, and provide insight into how the cognitive operations assumed in the model might be implemented in the brain.

Why do model-based fMRI?

As I see it, there are two main advantages:

1. Computational models will allow you to track the trial-by-trial dynamics of cognitive/neural processes, so you can see them unfold over time. Both event-related trial-averaging and block designs lose this level of granularity.

2. Model-based fMRI allows you to study how a cognitive process is implemented in the brain, and not just where it is occurring. While localization of function is important, I do think that most neuroscientists would agree that we would prefer if we could understand brain function at the mechanistic level.

A recipe for  standard model-based fMRI

This is taken from O'Doherty et al. (2007), which I recommend to anyone who is interested in using model-based fMRI for their studies.

1. Fit computational model to participants' behavioral data to obtain optimal model parameters that maximize the "fit" between model and data.
2. Using the best-fitting model and model parameters, generate a time-series for the cognitive variable of interest (e.g. value of chosen stimulus, prediction errors, confidence etc etc.). This time series can be thought of as the experimenter's best estimate of what the participant is "thinking" at each time point.
3. Convolve time-series with the hemodynamic response function. The resulting time-series can be thought of what the BOLD activity of a brain area encoding the cognitive variable of interest would look like, given the experimenter's best estimate of the variable.
4. Regress time-series with fMRI data to find voxels that correlate with cognitive variables of interest

I'd be happy to talk about the method. Leave a comment or drop me an email.