User Tools

Site Tools


biac:courses:advanced_fsl

Advanced Designs in FSL

Unlike the labs and exercises you've done so far, most analyses and designs will involve multiple conditions that must be modeled and contrasted appropriately to examine the cognitive processes you are investigating. In this lab, we will cover many of the things you might do in a real experimental design. You should be able to set up these designs in FEAT, the fMRI analysis tool in FSL. You won't need any real data, but you will need to create some (but not all) timing files (e.g., 3 column files) based on our descriptions.

On the FEAT webpage, there are detailed instructions on how to set up these sorts of models and also other, more complicated, models. We will go through parts of this webpage together and you will do part of it on your own.

Most of the time in fMRI analyses, you will be asking whether one condition is greater than another or vice versa. In FEAT, these are set up as linear contrasts between your regressors. Throughout this laboratory, you should be thinking about how to set up the contrasts for what we are testing. We will do most of this together, but you should get in the habit of thinking about how you can isolate cognitive processes via subtractive techniques like a simple contrast. A simple overview of the first-level stats and contrasts can be found here. If you're interested in a more detailed explanation, you can take a look at the relevant FSL course slides.

First level analyses

This is the part of the analysis that uses timing information from your experimental paradigm. This stage of the analysis is crucially important since the results from these analyses get “carried up” to higher levels of the analysis pipeline.

  1. Opening the FEAT GUI: Within the command terminal (this should not say “xterm” any where on it!), type Feat_gui &. This should open the FEAT GUI, which will automatically be set to do a First level analysis (see pull-down menu in the upper left corner of the FEAT window). You'll have to change this to “Higher level anlaysis” later in the lab.
  2. Within the Data tab, change your TR to 2 seconds and your number of input volumes to 110. You will not load any data.
  3. Go to the Stats tab, and select Full Model Setup.


SCENARIO #1:
You have an experiment with 3 conditions: monetary gains of $5, $10, and $15. These events are randomly interspersed throughout your experiment and there is nothing else that we want to model for this design.

Here are the timing files you would use for this setup:


SCENARIO #2:
Your experiment contains 2 events: 1.) a cue that predicts, with 100% certainty, a reward and 2.) the subsequent reward. Unfortunately, you did not design your experiment very well and these two regressors are highly correlated. Your main aim is to investigate the brain areas that are responsive to receiving a reward, but you are not very interested in the anticipatory effects.

Here are the timing files you would use for this setup:

Modeling multiple conditions

Most experiments involve more than one condition. In both of the scenarios above, we have more than one condition that we might want to model in our design matrix. You have the timing files for both of designs. How do you set up the models? How would you contrast different conditions? Let's work with the data from SCENARIO 1 for this example

  • Within full model setup, you will need to change the number of EVs to match the number of conditions in the scenario you are analyzing.
  • Your input for each EV should be a 3-column file
  • You will want to uncheck “Add Temporal Derivative” for each EV
  • Change the basis function to be the Double-Gamma function for each EV
  • Once you have the EVs loaded, you need to set up your contrasts in the Contrasts and F-tests tab. Unless you're looking at a main effect of a condition, the row of numbers (usually 0s and 1s/-1s) should sum to zero.
    • So, if I want to examine which areas are significantly more active for EV1 vs EV3, would have something that looks like this: [1 0 -1]. Try setting up a few contrasts for this model. How would you model the main effects? How would you see what voxels are more active for the $15 compared to the $5?
  • Once you've set up your entire model and contrasts, click View Design to see what your design looks like
  • Now click on Efficiency to see how “good” your design is. A “bad” design would be one with highly correlated regressors (denoted by light colors in the off-diagonal parts of the correlation matrix) and effects that require over a 3% change in signal to see. This design should look pretty good, but it could've been a little better with improved randomization and ordering of the stimuli.

Using orthogonalizations

Sometimes you might have a design in which your regressors might be correlated (e.g., reward anticipation → reward outcome). Sometimes you can avoid correlated regressors by carefully designing your experiment. However, sometimes this isn't always possible and you may not care about the variability from one condition. If the regressors are correlated, you will not be able to parse the variance appropriately (i.e., variability related to one regressor might be captured by the other regressor or vice versa, leading to misestimations.)

In SCENARIO 2, we have this exact problem. How can we try to ameliorate this problem, especially if we're only interested in the reward outcome component of the task? We can make FEAT force the regressors to be orthogonal, but what do we do if we're just interested in the reward outcome? We can orthogonalize outcome w.r.t. anticipation, or we can orthogonalize anticipation w.r.t. outcome. Which set up should you use if you're interested in reward outcome?

  • Load in the two EVs: anticipation.txt and outcome2.txt
  • Use the same settings as you did earlier and think about how you would set up the contrasts (or really just the main effects for this design).
  • Now look at the model and the efficiency. Notice that the the regressors are highly correlated, which is denoted by the light colors on the off-diagonal elements in the correlation matrix.
  • You need to make the outcome EV orthogonal w.r.t. the anticipation EV. Check the Orthogonalise option in the EV tab that corresponds to your outcome EV. Now make the outcome EV orthogonal w.r.t. the anticipation EV.
  • Now the design has been corrected, take a look at the efficiency and the model. Note how your outcome EV changes a little bit.


Presumably, we could've set up our design better to make the outcome and anticipation less correlated from the outset. This could've been accomplished by jitterring the interval between these events. Try this revised outcome EV and see what you get: :biac:courses:outcome.txt

Using parametric regressors

Sometime we might be interested in linear trends of activation – that is, testing which brain areas respond linearly with the level of a condition (e.g., reward). This sort of pattern is implicit in SCENARIO 1; however, we need change the model and EV files. Remember that the 3rd column in the 3-column files controls the intensity of the the event, so this is the part we can change to look for bigger responses for certain events.

You will make the 3-column files for this, but it shouldn't be too hard. You should open the EV files from the first scenario for reference and so you'll know when each event occurred

  1. We need one regressor that has all the events ($5, $10, and $15) in the same 3-column file. The 3rd column in this file (constant.txt) should have all 1s in it.
  2. We need another regressor that also has all of the events; however, we need change the 3rd column in this file (linear.txt).
    • Events corresponding to $5 should have a 1 in the 3rd column
    • Events corresponding to $10 should have a 2 in the 3rd column
    • Events corresponding to $15 should have a 3 in the 3rd column
  3. Now you have two EVs: a linear term and constant term. These will be partially correlated and you care about the variance related to the linear effect, so you need orthogonalize your linear EV w.r.t. the constant EV.
  4. For the convolution and other options in the model tab, use the same settings as you did earlier. You should also think about how you would set up the contrasts (or really just the main effects for this design).


See this FSL forum post for more details about modeling increasing levels of activation.

Second level analyses

The main interesting thing you might do at this level is combing across different kinds of runs. For example, you might have a set up where the odd-numbered runs are non-social (e.g., played against a computer opponent) and the even-numbered runs are social (e.g., played against a human opponent).

  1. Make sure the pull-down menu in the upper left corner of the FEAT window is set to “Higher level analysis”.
  2. Within the Data tab, change your number of analyses to equal the number of runs you have. You will not load any data.
  3. Go to the Stats tab, and select make sure the analysis type is set to Fixed Effects.
  4. Within the Stats tab, click on Full Model Setup to set up your model.

Collapsing across different types of runs

You have 6 runs in your experiment; however, the subject was doing something fundamentally different in odd vs even runs. Rather than having all 1s in a single column of your design matrix, you now need two columns to pull out these effects. How do you think you would set this up?

  1. You should have 2 EVs: one for each kind of run
  2. Put 1s in the rows of the columns that correspond to the run type. You should have 1s and 0s in each column and nothing else.

Third level analyses

Although you will not be doing third level analyses in this class, there are several analysis principles that can be illustrated when setting up group level analyses. Generally, researchers are interested in the main effects of a condition across their subject sample; however, there is sometimes a need to go past looking at main effects of a particular condition. For example, researchers may want to examine group differences between two populations, or researchers may be interested in how individual differences in some behavioral trait (e.g., sensation seeking) might vary with brain activation (e.g., reward-related activation in the ventral striatum). You will frequently see these techniques applied in papers, and you might one day need to do it yourself.

  1. Make sure the pull-down menu in the upper left corner of the FEAT window is set to “Higher level analysis”.
  2. Within the Data tab, change your number of analyses to equal the number of subjects you have. You will not load any data.
  3. Go to the Stats tab, and select make sure the analysis type is set to Mixed Effects: FLAME1. We used mixed effects instead of fixed effects so that we can generalize our results to the population.
  4. Now click on Full Model Setup to set up the higher level model

Main effects

Setting up main effects is easy. For this, each subject will correspond to one input in your higher level analysis. So, when you're in the full model set up, you will simply create a column of 1s in the design model. You will have to do this for each COPE (Contrast of Parameter Estimates), which will be carried up from your lower-level analyses.

See this page for details.


SCENARIO:
Imagine you have design with 20 subjects and you want to examine the main effect of a particular condition (it doesn't matter what condition). How would you set your model up in FEAT?

Testing whether groups are different

Sometimes you will want to test whether two groups are different. This is relatively easy to set up in FEAT. Let's imaging that our 20 subjects from above also took a sensation-seeking survey (see scores below). We can do a median split based on these scores to divide our sample into two groups: high sensation seeking and low sensation seeking. See this page for details on how to set up a test of two groups. Essentially, rather than having one group and one EV for our model, we will now have two. We can then do contrasts on these EVs.

SUBJ  score
1     1
2     1
3     2
4     4
5     4
6     7
7     8
8     8
9     8
10    9
11    11
12    13
13    13
14    14
15    16
16    17
17    19
18    19
19    20
20    20

Using covariates

When you use covariates in our group level analysis, you are testing whether variability in your data can be “explained” by a regressor that models cross-subjects differences along some trait (e.g., sensation seeking). This is easy to set up in FEAT; however, the main thing you need to remember to do is demean your covariate (i.e., the scores or trait measures should sum to zero across all of your subjects).

Let's take the same 20 subjects from above and, rather than doing a median split based on their sensation-seeking scores, let's add in those scores as a covariate. Remember to demean the scores by subtracting the mean from each individual score.

See this page for further details.

biac/courses/advanced_fsl.txt · Last modified: 2023/02/23 18:43 (external edit)