Lecture 4- analysis and experimental design Flashcards
Typical preprocessing steps for fMRI analysis
Spatial smoothing
Motion correction
Temporal filtering
Statistical maps in fMRI show the probability of…
Type 1 error
To support robust statistical inferences which can be extended to a wider population, multi-subject fMRI experiments typically employ…
random/mixed-effects, within-subjects design
Why are voxel’s activity modelled by combining predicted time series for each distinct class of event (or block) which are fitted to the fMRI data?
to minimise residual variation
If BOLD responses to individual experimental trials of the same task are to be compared (e.g., correct versus incorrect trials) in an fMRI study, it is essential to use:
an event related design
Advantage of block design
High signal to noise
In an fMRI experiment, what is the IV and the DV?
IV= task DV= BOLD signal
What is cognitive subtraction?
Choose 2 tasks which differ in one critical aspect
Then compare BOLD signal in one from the other
How can confounding factors be controlled for?
Randomisation and counterbalancing
Why should fMRI experiments be within-subjects?
To reduce individual variation in BOLD signal and differences in anatomy
What is block design?
Conditions are divided into distinct blocks
Typically several trials within each block (5-30 secs)
IV = constant within a block
What does the transition between blocks represent?
Changes in the level of the IV
Advantage of longer interblock intervals
the hemodynamic response can return to baseline during the control condition = more signal change
Advantage and disadvantage of shorter interblock intervals
Allows for more repetitions
BUT
The sluggish nature of hemodynamic response means less signal change
Disadvantage of block design
Response to individual trials cannot be distinguished
Why is there a high signal:noise ratio in block design?
Because of plateau within a block, there’s excellent power to detect active voxels
What is an event related design?
Each event (trial) is modelled separately
How can responses be estimated for different classes of trial in an event related design?
Carefully spaced and time-locked trials
Disadvantage of event related design
Reduced detection power (smaller signal change)
Advantage of event related design
Data can be analysed according to PS response
Time course of BOLD response can be accurately estimated
How is motion corrected?
Successive images are aligned with one another
What is spatial smoothing?
Blurring by averaging a few voxels together in space = increase signal to noise
What is high-pass temporal filtering?
Gradual changes in signal intensity removed
often caused by breathing and heart beat
In the GLM, what do beta weights tell you?
tells you how much a voxels signal increases in response to particular stimuli/task
BOLD measures the concentration of WHAT?
deoxyhaemaglobin
In fMRI analysis using GLM, what do we KNOW (input) and what do we WANT to know (output)
INPUT: a voxels true BOLD signal and our predictors
OUTPUT: beta weights that combine with the predictors to give us the best approximation of the true signal
What are some predictors in the GLM?
Baseline activity
Response to task condition
Response to control condition
What do we combine with the predictors to best estimate the true signal from a voxel?
beta weights
error signal
What mathematical tranformation do you apply to predicted time-series of activation?
Haemodynamic Response Function (HRF)
In GLM analysis, the aim is to find the __-__ that best appoximate a voxel’s signal time-series.
The best approximation is the one with the ___ ___
Then, a comparison is made between the ___-___ for the task condition to the ___-___ to control condition.
beta-weights
least errors
beta-weights
beta-weights
A voxels BOLD signal at rest is 100.
When the PS is reading scentences, the signal increases to 120.
When reading nonwords, the signal increases to 110.
When approximating the voxel’s BOLD signal using our three predictors, what should the beta weights be for sentences and non words?
20 , 10
the beta tells you by how much the signal increases from baseline!
What are the components of the GLM?
BOLD signal = design matrix (where each column is a predictor)*their beta weights + errors
OR
explained variance/task related activity changes + unexplained variance
What do we give the GLM? What do we know?
The BOLD signal
X/The design matrix (each column is a regressor that we built ourselves)
What does the GLM give us? What do we want to find?
How does it find it?
vector of beta weights (one per regressor in X) that gives the best approximation of the BOLD signal
By minimising the sum of squared errors.
Two reasons why we square the errors?
To get rid of negatives
It gives a big weight to big errors and a smaller weight to small errors (because squaring a big number disproportionately increases it compared to squaring a small number)
A comparison of beta-weights is called a ____.
contrast
What is a beta-weight?
A weight indicating how much a predictor contributes to the true BOLD signal.
e.g. how much the signal changes due to the hypothesised process, holding everything else constant.
The error can be defined as the difference between….?
The difference between the true BOLD signal and the combination of predictors that best approximates it.
These are changes (variations) in the signal we cannot explain with the predictors
Contrast is the difference between…?
The difference between two (groups of) betas.
the contast t-value is the number resulting from dividing the ____ ____ by a measure of ____.
contrast value (e.g. beta for task - beta for control)
DIVIDED BY
Error
What do statistical maps indicate and how do they represent it graphically?
Indicate the probablility of a type-1 error
Display a t/z statistic that is converted to p-value
The lower the probability (p-value) the brighter/hotter the colour
They are overlain on a anatomical image
Threshold is used to exclude higher p-values
What is bonferroni correction?
Takes into account the number of comparisons
What is FWE voxel correction?
Takes into account the number of independent comparisons
What is cluster-size correction?
Takes into account the contiguity of active voxels
What is FDR correction?
Controls for the false-discovery rate
Problem with whole brain, voxel-by-voxel comparison?
Performing more than 50,000 statistical tests so risk of type 1 error!
Advantages of Region-of-Interest analysis?
Avoid multiple comparison problem by restricting analysis to particular region, identified in advance.
Greater statistical power.
Disadvantage of Region-of–Interest analysis?
May miss potentially important regions outside of ROI.
Danger of biased post-hoc selection of ROI
When performing group-level analysis, what two ways can you combine data across subjects?
Fixed effects
Mixed/random effects
What does ‘fixed effects’ analysis entail?
- assumes the effect of the exp. condition is fixed (constant) across subjects
- the data from all PS is combined (as if it came from the same PS) and the stat. tests are performed
What does ‘mixed/random effects’ entail?
- Assume that the experimental condition could have a different effect on different PS
- a stat. test is run on each PS (first level), followed by a (second-level) test to determine whether the first level analysis is consistent across PS
Does modern research utilise ‘fixed effects’ or ‘mixed/random effects’
Mixed/random effects
When performing whole-brain analysis, different PSs’ brains need to be aligned with each other in the same space. What is this called?
Coregistration or Normalisation.