basics of experimental design for fmri last update: november 2008
TRANSCRIPT
“Attending a poster session at a recent meeting, I was reminded of the old adage ‘To the man who has only a hammer, the whole world looks like a nail.’ In this case, however, instead of a hammer we had a magnetic resonance imaging (MRI) machine and instead of nails we had a study. Many of the studies summarized in the posters did not seem to be designed to answer questions about the functioning of the brain; neither did they seem to bear on specific questions about the roles of particular brain regions. Rather, they could best be described as ‘exploratory’. People were asked to engage in some task while the activity in their brains was monitored, and this activity was then interpreted post hoc.”
-- Stephen M. Kosslyn (1999). If neuroimaging is the answer, what is the question? Phil Trans R Soc Lond B, 354, 1283-1294.
Brains Needed
"...the single most critical piece of equipment is still the researcher's own brain. All the equipment in the world will not help us if we do not know how to use it properly, which requires more than just knowing how to operate it. Aristotle would not necessarily have been more profound had he owned a laptop and known how to program. What is badly needed now, with all these scanners whirring away, is an understanding of exactly what we are observing, and seeing, and measuring, and wondering about."
-- Endel Tulving, interview in Cognitive Neuroscience (2002, Gazzaniga , Ivry & Mangun, Eds., NY: Norton, p. 323)
Localization
Localization for localization’s sake has some value– e.g., presurgical planning
• However, it is not especially interesting to the cognitive neuroscientist in and of itself
• Popularity of brain imaging results suggests people are inherent dualists
moving bodiessocial cognition
faces objectsstatic bodies
grasping
motion perception
motion near head
orientation selectivitymemory
scenes
motorcontrol
reaching and pointing
touch
retinotopic visual maps eyemovements
executive control
The Brain After fMRI (Incomplete)
Useful Types of Imaging Studies
• Testing of theories and models• Comparing stimuli or tasks within a region• Comparing stimuli or tasks across a network• Examining coding within areas
– fMRI adapation– Multi-voxel pattern analysis
• Correlations between brain and behavior• Evaluation of the role group differences, experience
and even genetics• Comparisons between species• Exploration of specialized human functions
– e.g., language, tool use, mathematics
• Derivation of general organizational principles
So you want to do an fMRI study?
CONCLUSION: Unless you are Bill Gates, a thought experiment is much more efficient!
Average cost of performing a thought experiment:
Your Salary
Average cost of performing an fMRI experiment in 1998:
Thought Experiments• What do you hope to find? • What would that tell you about the cognitive process involved? • Would it add anything to what is already known from other techniques? • Could the same question be asked more easily & cheaply with other techniques?
• What would be the alternative outcomes (and/or null hypothesis)? • Or is there not really any plausible alternative (in which case the experiment may not be worth doing)? • If the alternative outcome occurred, would the study still be interesting? • If the alternative outcome is not interesting, is the hoped-for outcome likely enough to justify the attempt? • What would the “headline” be if it worked? Is it sexy enough to warrant the time, funding and effort? • “Ideas are cheap.” -- Jody’s former supervisor, Jane Raymond
• Good experimenters generate many ideas and ensure that only the fittest survive
• What are the possible confounds?• Can you control for those confounds?
• Has the experiment already been done? “A year of research can save you an hour on PubMed!”
Three Stages of an ExperimentSledgehammer Approach• brute force experiment• powerful stimulus• don’t try to control for everything• run a couple of subjects -- see if it looks promising• if it doesn’t look great, tweak the stimulus or task• try to be a subject yourself so you can notice any problems with stimuli or subject strategies
Real Experiment• at some point, you have to stop changing things and collect enough subjects run with the same conditions to publish it• incorporate appropriate control conditions• there is some debate on how many subjects you need
• some psychophysical studies test two or three subjects• many studies test 6-10 subjects• random effects analysis requires at least 10 subjects
• can run all subjects in one or two days • pro: minimize setup and variability• con: “bad magnet day” means a lot of wasted time
Whipped Cream• after the real experiment works, then think about a “whipped cream” version• going straight to whipped cream is a huge endeavor, especially if you’re new to imaging
Mental Chronometry
• use reaction times to infer cognitive processes
• fundamental tool for behavioral experiments in cognitive science
F. C. DondersDutch physiologist
1818-1889
Classic Example
DetectStimulus
PressButton
DetectStimulus
PressButton
Discriminate Color
DetectStimulus
PressButton
Discriminate Color
ChooseButton
Time
T3: Choice Reaction Time• Hit left button when light is green and right button when light is red
T1: Simple Reaction Time• Hit button when you see a light
T2: Discrimination Reaction Time• Hit button when light is green but not red
Subtraction Logic(A + B) - A = B
DetectStimulus
PressButtonT1
DetectStimulus
PressButton
Discriminate ColorT2
-
Discriminate Color
=
Subtraction Logic(A + B) - A = B
DetectStimulus
PressButton
Discriminate ColorT2
-
=
DetectStimulus
PressButton
Discriminate Color
ChooseButtonT3
ChooseButton
Limitations of Subtraction Logic
Assumption of pure insertion• You can insert a component process into a task without
disrupting the other components• Widely criticized
Top Ten Things Sex and Brain Imaging Have in Common
10. It's not how big the region is, it's what you do with it.
9. Both involve heavy PETting.
8. It's important to select regions of interest.
7. Experts agree that timing is critical.
6. Both require correction for motion.
5. Experimentation is everything.
4. You often can't get access when you need it.
3. You always hope for multiple activations.
2. Both make a lot of noise.
1. Both are better when the assumption of pure insertion is met.
Source: students in the Dartmouth McPew Summer Institute
Now you should get this joke!
Subtraction Logic: Brain Imaging Example
Hypothesis (circa early 1990s): Some areas of the brain are specialized for perceiving objects
Simplest design: Compare pictures of objects vs. a control stimulus that is not an object
minus = object perception
seeingpictures
like
seeingpictures
like
Malach et al., 1995, PNAS
Other Differences
• Is subtraction logic valid here?• What else could differ between objects and textures?
Objects > Textures• object shapes• irregular shapes• familiarity
– namability
• visual features (e.g., brightness, contrast, etc.)• actability• attention-grabbing
Other SubtractionsLateral Occipital Complex
Visual Cortex (V1)
Malach et al., 1995, PNAS
>
>
>
Grill-Spector et al., 1998, Neuron
Kourtzi & Kanwisher, 2000, J Neurosci
Dealing with Attentional ConfoundsfMRI data seem highly susceptible to the amount of attention drawn to the stimulus or devoted to the task.
Add an attentional requirement to all stimuli or tasks.
How can you ensure that activation is not simply due to an attentional confound?
Time
Example: Add a “one back” task• subject must hit a button whenever a stimulus repeats• the repetition detection is much harder for the scrambled shapes • any activation for the intact shapes cannot be due only to attention
Other common confounds that reviewers love to hate:• eye movements• motor movements
Change only one thing between conditions!
As in Donders’ method, in functional imaging studies, two paired conditions should differ by the inclusion/exclusion of a single mental process
How do we control the mental operations that subjects carry out in the scanner?
i) Manipulate the stimulus• works best for automatic mental processes
ii) Manipulate the task• works best for controlled mental processes
DON’T DO BOTH AT ONCE!!!
Source: Nancy Kanwisher
Beware the “Brain Localizer”• Can have multiple comparisons/baselines• Most common baseline = rest• In some fields the baseline may be straightforward
– For example, in vision studies, the baseline is often fixation on a point on an otherwise blank screen
• Be careful that you don’t try to subtract too much
Reaching – rest• = visual stimulus• + localization of stimulus• + arm movement• + somatosensory feedback• + response planning• + …
What are people doing during rest?
What are people really doing during rest?• Daydreaming, thinking• Remembering, imagining• Attending to bodily sensations
– “I really have to pee!”, “My back hurts”, “Get me outta here!”
• Getting drowsy
Problems with a Rest Baseline?
• For some tasks (e.g., memory studies), rest is a poor, uncontrolled baseline– memory structures (e.g., medial temporal
lobes) may be DEactivated in a task compared to rest
• To get a non-memory baseline, some memory researchers put a low-memory task in the baseline condition– e.g., hearing numbers and categorizing
them as even or odd
ParahippocampalCortex
Stark et al., 2001, PNAS
Is concurrent behavioral data necessary?“Ideally, a concurrent, observable and measureable behavioral response, such as a yes or no bar-press response, measuring accuracy or reaction time, should verify task performance.”-- Mark Cohen & Susan Bookheimer, TINS, 1994
“I wonder whether PET research so far has taken the methods of experimental psychology too seriously. In standard psychology we need to have the subject do some task with an externalizable yes-or-no answer so that we have some reaction times and error rates to analyze – those are our only data. But with neuroimaging you’re looking at the brain directly so you literally don’t need the button press… I wonder whether we can be more clever in figuring out how to get subjects to think certain kinds of thoughts silently, without forcing them to do some arbitrary classification task as well. I suspect that when you have people do some artificial task and look at their brains, the strongest activity you’ll see is in the parts of the brain that are responsible for doing artificial tasks.
-- Steve Pinker, interview in the Journal of Cognitive Neuroscience, 1994
Source: Nancy Kanwisher
Parameters for Neuroimaging
You decide:• number of slices• slice orientation• slice thickness• in-plane resolution (field of view and matrix size)• volume acquisition time• length of a run• number of runs• duration and sequence of epochs within each run• counterbalancing within or between subjects
Your physicist can help you decide:• pulse sequence (e.g., gradient echo vs. spin echo)• k-space sampling (e.g., echo-planar vs. spiral imaging; single- vs. multi-shot)• TR, TE, flip angle, etc.
Tradeoffs
Number of slices vs. volume acquisition time• the more slices you take, the longer you need to acquire them• e.g., 30 slices in 2 sec vs. 45 slices in 3 sec
“fMRI is like trying to assemble a ship in a bottle – every which way you try to move, you encounter a constraint” -- Mel Goodale
Number of slices vs. in-plane resolution• the higher your in-plane resolution, the fewer slices you can acquire in a constant volume acquisition time• e.g., in 2 sec, 7 slices at 1.5 x 1.5 mm resolution (128 x 128 matrix) vs. 28 slices at 3 mm x 3 mm resolution (64 x 64 matrix)
More Power to Ya!Statistical Power• the probability of rejecting the null hypothesis when it is actually false• “if there’s an effect, how likely are you to find it”?
Effect size• bigger effects, more power
• e.g., LO localizer (intact vs. scrambled objects) -- 1 run is usually enough• looking for activation during imagery of objects might require many more runs
Sample size• larger n, more power
• more subjects• longer runs• more runs per subject
Signal:Noise Ratio• better SNR, more power
• higher magnetic field• multi-channel coils• fewer artifacts (physical noise, physiological noise)
Put your conditions in the same run!
Why?• subjects get drowsy and bored• magnet may have different amounts of noise from one run to another (e.g., spike)• some stats (e.g., z-normalization) may affect stats differently between runs
By this logic, there is higher activation for Places than Faces in the data to the left.Do you agree?
Bottom line: If you want to compare A vs. B, compare A vs. B! Simple, eh?
As far as possible, put the two conditions you want to compare within the same run.
Common flawed logic: Run1: A – baseline Run2: B – baseline
“A – 0 was significant, B – 0 was not, Area X is activated by A more than B”
Faces Places
Error bars = 95% confidence limits
BO
LD
Act
iva
tion
(%
)
Run Duration
How long should a run be?• Short enough that the subject can remain comfortable without moving or swallowing• Long enough that you’re not wasting a lot of time restarting the scanner• My ideal is ~6 ± 2 minutes
Simple Example Experiment: LO Localizer
IntactObjects
ScrambledObjects
BlankScreen
TIME
One volume (12 slices) every 2 seconds for 272 seconds (4 minutes, 32 seconds)
Condition changes every 16 seconds (8 volumes)
Lateral Occipital Complex• responds when subject views objects
(Unit: Volumes)
Options for Block Design Sequences
That design was only one of many possibilities. Let’s consider some of the other options and the pros and cons of each.
Let’s assume we want to have an LO localizerWe need at least two conditions:
but we could consider including a third condition
Let’s assume that in all cases we need 2 sec/volume to cover the range of slices we require
Let’s also assume a total run duration of 136 volumes (x 2 sec = 272 sec = 4 min, 16 sec
We’ll start with 2 condition designs…
Block Design: Short Equal Epochs
Alternation every 4 sec (2 images)• signal amplitude is weakened by HRF because signal doesn’t have enough time to return to baseline• not to far from range of breathing frequency (every 4-10 sec) could lead to respiratory artifacts• if design is a task manipulation, subject is constantly changing tasks, gets confused
HRF-convolved
time course
raw time course
Block Design: Short Unequal Epochs
4 sec stimuli (2 image) with 8 sec (4 image) baseline• we’ve gained back most of the HRF-based amplitude loss but the other problems still remain• now we’re spending most of our time sampling the baseline
HRF-convolved
time course
raw time course
Block Design: Long EpochsThe other extreme…
Alternation Every 68 sec (34 images)• more noise at low frequencies• linear trend confound• subject will get bored• very few repetitions – hard to do eyeball test of significance
HRF-convolved
time course
raw time course
Physiological Noise
Respiration• every 4-10 sec (0.3 Hz)• moving chest distorts susceptibility
Cardiac Cycle• every ~1 sec (0.9 Hz)• pulsing motion, blood changes
Solutions• gating• avoiding paradigms at those frequencies
You want your paradigm frequency to be in a “sweet spot” away from
the noise
Block Design: Medium Epochs
Every 16 sec (8 images)• allows enough time for signal to oscillate fully• not near artifact frequencies• enough repetitions to see cycles by eye• a reasonable time for subjects to keep doing the same thing
HRF-convolved
time course
raw time course
Block Design: Other Niceties
• If you start and end with a baseline condition, you’re less likely to lose information with linear trend removal and you can use the last epoch in an event related average
truncated too soon
Block Design Sequences: Three Conditions
• Suppose you might want to add a third condition to act as a more neutral baseline
• For example, if you wanted to identify visual areas as well as object-selective areas, you could include fixation as the baseline.
• That would allow two subtractions– scrambled - fixation visual areas– intact - scrambled object-selective areas
• Now the options increase.• For simplicity, let’s keep the epoch duration at 16
sec.
Block Design: Repeating Sequence
• We could just order the epochs in a repeating sequence…
• Problem: There might be order effects
• Solution: Counterbalance with another order
Block Design: Random Sequence
• We could make multiple runs with the order of conditions randomized…
Block Design: Regular Baseline
• We could have a fixation baseline between all stimulus conditions (either with regular or random order)
As we will see when we talk about event-related averaging, this regular baseline design is optimal for getting nice average time courses
So What Do We Do?!!!
• Any of these designs should work. Some might work better than others depending on your goals.
• If you only care about the difference between Intact and Scrambled, you’d be best to go with a 16-sec alternating epochs with only those two conditions
• If you are going for three conditions…– putting baselines between all other epochs is great for event-related
averaging BUT it means you’re wasting a lot of your statistical power estimating the baseline
– regular sequences should include counterbalancing– random sequences can be a lot of work to make protocols
But I have 4 conditions to compare!Here are a couple of options.
A. Orderly progression
Pro: SimpleCon: May be some confounds (e.g., linear trend if you predict green&blue > pink&yellow)
B. Random order in each runPro: order effects should average outCon: pain to make various protocols, no possibility to average all data into one time course, many frequencies involved
C. Kanwisher lab clustered design• sets of four main condition epochs separated by baseline epochs• each main condition appears at each location in sequence of four• two counterbalanced orders (1st half of first order same as 2nd half of second order and vice versa) – can even rearrange data from 2nd order to allow averaging with 1st order
Pro: spends most of your n on key conditions, provides more repetitionsCon: not great for event-related averaging because orders are not balanced (e.g., in top order, blue is preceded by the baseline 1X, by green 2X, by yellow 1X and by pink 0X.
As you can imagine, the more conditions you try to shove in a run, the thornier ordering issues are and the fewer n you have for each condition.
My rule of thumb: Never push it beyond 4 main + 1 baseline.
But I have 8 conditions to compare!
• Just don’t.
• In my experience, any block design experiment with more than four conditions becomes unmanageable and incomprehensible
• Event-related designs might still be an option… stay tuned…
Design Types
BlockDesign
Slow ERDesign
RapidCounterbalanced
ER Design
RapidJittered ER
Design
MixedDesign
= null trial (nothing happens)
= trial of one type (e.g., face image)
= trial of another type (e.g., place image)
Default Mode Network
• During resting state scans, there are two networks in which areas are correlated with each other and anticorrelated with areas in the other network
Fox and Raichle, 2007, Nat. Rev. Neurosci.
EXCEPT when the activated region does not fill the voxel (partial voluming)
Voxel size
3 x 3 x 6= 54 mm3
e.g., SNR = 100
3 x 3 x 3= 27 mm3
e.g., SNR = 71
2.1 x 2.1 x 6= 27 mm3
e.g., SNR = 71
isotropic
non-isotropic
non-isotropic
In general, larger voxels buy you more SNR.
Example of a Successful Paper
The most successful empirical fMRI paper(not methods, not meta-analysis)
Context
Neurons in the macaque temporal lobe are tuned to faces
Human patients may lose ability to recognize faces but not other objects
The human fusiform gyrus had been implicated in face processing by earlier PET studies
Background
• Background: Numerous prior studies had demonstrated activation for faces in the fusiform gyrus. However, they had not controlled for other possible differences between faces and control stimuli.
• Question: Is the processing of faces distinct from the processing of objects when various other confounds are eliminated
Step 1: Face Localizer Scans
Block Design45 faces (30 s) … Fixation (20 s) … 45 objects (30 s) … Fixation (20 s) …
vs.
Kanwisher, McDermott, & Chun, 1997, J. Neurosci.Slide modified from student presentation by Michelle Waese, 2005
Step 2: Identify ROI in Each Individual
Region of Interest = ROIFound in 12 of 15 subjects
Step 3: Test Alternate Hypotheses in ROIs
• Extract activation patterns from individual ROIs– %
• Apply conventional statistics to them
Faces vs. Low-Level Features
• If FFA is truly face selective, it should respond more to faces than scrambled faces with the same low-level visual features (e.g., luminance, contrast). It does.
vs.
vs.
Faces vs. Exemplars of Other Categories
• If FFA is truly face selective, it should respond more to individual faces than individual houses, which are also exemplars within a category of similar objects. It does.
Faces vs. Other Animate Objects
vs.
• If FFA is truly face selective, it should respond more to faces than to hands, which are other body parts that move. It does.
Faces vs. Attention
vs.
• If FFA is truly face selective, it should respond more to faces than to hands even when subjects perform a “1-back task” to maintain attention. It does.
“Press a button whenever you see the same image repeated.”
Conclusions
• A part of the right fusiform gyrus is preferentially active during face viewing even when alternative explanations were ruled out.
• Implications that “faces are special” and not just another type of visual stimulus
Why was this paper so successful?
• “discovered” an area that became hugely studied– not entirely -- numerous other papers had previously
reported activation for faces
• goal was not just localization, but a deeper understanding of brain activation patterns
• theoretically driven – “are faces special”– relates well to other literatures
• unambiguously showed face selectivity– well-thought control experiments
• was the beginning of several controversial debates– category specificity– nature vs. nurture
What have we learned about the face area?
The face area is activated:• when faces are perceived or imagined
correlation between brain and behavior• for stimuli at the fovea
cues to brain organization• by circular patterns
cues/constraints for modelling• in certain areas of the monkey brain
cues to brain evolution• for other categories of objects that subjects
have extensive experience with debate regarding nature/nurture
• to some degree by other categories of objects debate regarding distributed vs. modular coding
in the brain
The fusiform face area may be impaired:• in some but not all patients who have
problems recognizing faces• in people with autism
understanding of brain disorders
Disdaqs• Discarded data acquisitions: trashed volumes at the beginning of a run
before the magnet has reached a steady state• Images are not saved to disk.• Sometimes it can take awhile for the subject to reach a steady state too --
Startle response!
Prepare Well: Subjects
• recruit and screen your subjects well in advance– safety screening
• best to let them read through and self-screen beforehand so you don’t get any embarrassing situations (e.g., discussions about IUDs, pregnancy)
– eye glasses– handedness
• make sure your subjects know how to be good subjects – http://www.ssc.uwo.ca/psychology/culhamlab/Jody_web/Subject_Info/firsttim
e_subjects.htm
• make sure you and the subjects can contact each other in case of problems or delays
• if possible, be a subject yourself to see what the pitfalls and strategies might be
• remember to bring:– subject fees (and receipt book)– consent and screening forms
Dealing with frustration
Sign that used to be at the 1.5 T at MGH
Murphy's law acts with particular vigour in fMR imaging:
Number of pieces of equipment required in an fMRI experiment: ~50
Probability of any one piece of equipment working in a session: 95%
Probability of everything working in a session: 0.95^50 = 7.6%
Solution for a good imaging session =$4 million magnet + $3 roll of duct tape
How NOT to do an imaging experiment
• ask a stupid question– e.g., “I wonder what lights up for daydreaming vs. rest”
• compare poorly-defined conditions that differ in many respects• use a paradigm from another technique (e.g., cognitive psychology)
without optimizing any of the timing for fMRI, e.g., 1 minute epochs• never look at raw data, time courses or individual data, just plunk it all
into one big stat model and look at what comes out• publish a long list of activated foci in every possible comparison• don’t use any statistical corrections• write a long discussion on why your task activates the subcortico-
occipito-parieto-temporo-frontal network
Source: Anonymous, to protect the reputations of the researchers
• Bonus points: if there are areas you don’t want to see, just color them gray so they don’t show up well!