Let’s see how well you understood GLM.

• What do regressors correspond to in our GLM?
• What do we exactly do with our Hemodynamic response function (HRF) so that it fits our neural data? How does HRF relate to our regressors?
• Do you think HRF is a one-size fits all function? Is HRF representative of the hemodynamic characteristics of all brain regions?
• What is the reason of the post-stimulus undershoot in the HRF? The undershoot corresponds to the part where the values go below the initial level of BOLD response at the end of the cannonical HRF. Google “HRF undershoot” to see a picture of it.
• The HRF is convolved to have an estimate of neural activity. What is meant by “convolution” here? If you do not know, check this video.
• What does the residual error correspond to?
• In fMRI we work with voxels. Do we do GLM on each voxel’s time series seperately, or do we do some kind of averaging for a specific set of voxels?
• In conjuction to the question above, what exactly does the beta value generated from each voxel correspond to, then?
• Can you model two different type of regressors to the same time point in the same GLM?
• Why is stimulus onset asynchrony (jittering) particularly important for event-related designs?
• How do we use realignment movement parameters in our GLM? What is the specific term used for such data?
• How are the beta values generated from our GLM used for when contrasts are calculated?
• How do we generate a contrast vector?
• Why do we subtract values / weights from each other while making a contrast vector? Does directionality matter? (ie. B-A vs. A-B)
• Assume we have regressors A, B,C, D in our GLM. Write a contrast vector (with their corresponding weights) which would generate a ((A+B+C) – D) contrast
• Why do you think we should model events in our experiment as a regressor even though we will not use them in our contrasts?
• Why do you think it is useful to add a seperate regressor for incorrect or unused trials/events?
• Why do you think the HRF is called “sluggish”?
• If you wanted to study the brain activity in a time series (for example, 3-4 TR’s after the onset of your trial) what basis function would you use for your GLM? How is this analysis different than HRF?
• Let’s say you are doing some multivariate analysis (RSA or Machine learning classification) would you do this GLM in smoothened and normalized data, or unsmoothened and normalized data? Why?
• Consider the question above and think of a reason which would justify doing analysis on unnormalized data.
• Considering the correlational nature of fMRI time series data and the sluggish nature of HRF, what are some ways of combatting this correlational nature in event related designs and block designs?
• How do you think you can combat correlations between the durations you choose for stimulus (trial) onset asynchrony? Think of a way that would allow you to investigate whether your TOA’s are inherently correlated.
• What is the difference between modelling your regressor/trial as an event or an epoch?
• When should you model a regressor as an event or an epoch? On what ground do we decide doing this?
• What do T-maps / T-contrasts correspond to?
• Ditto above, but for F-maps and F-tests.
• What kind of a test can we run to understand the statistically significant t-value threshold to check when we are looking at our contrasts under MriCRON?
• Briefly discuss what happens in FDR, FWER and iTT correction.
• Why do you think it is best practice to distribute your events evenly across an fMRI run?
• Why do you think it is important to have a sufficient amount of event repetitions in your GLM?
• How does the question above relate to having longer scans/experiments?
• fMRI data is inherently noisy. What factors do contribute to noise and what can be done both at the acquisition stage and post-acquisition to combat noisy data?
• What are some methods to decrease the effect of noise in your data?
• Distortions in EPI images in the phase-encoding direction is common. Do you know how one can fix distortions in EPI images?
• What are GLM covariates? Why do we use them?
• Why is the average activity of each fMRI run incorporated as a covariate to the GLM?