fmri experimental design - mitweb.mit.edu/hst.583/www/course2001/lectures/hst583_lect...because fmri...

FMRI Experimental Design

Lila DavachiDepartment of Brain & Cognitive Sciences, M.I.T.

Because fMRI BOLD data is not anBecause fMRI BOLD data is not anabsolute measure of neuronal activity,absolute measure of neuronal activity,

all study designs must provide theall study designs must provide theopportunity to statistically contrast theopportunity to statistically contrast the

neuronal activity of interest with aneuronal activity of interest with asuitable rest or background conditionsuitable rest or background condition..

Thus, study design is Thus, study design is of paramount importance.of paramount importance.

Why?Why?HypothesisHypothesis

How?How?fMRI Study DesignfMRI Study Design

Where?Where?NeuroanatomyNeuroanatomy

What?What?BehaviorBehavior

Thanks to Chantal Stern

43 * 7 = ?

Key Points

• What can fMRI tell you?

• Always comparing across conditions

• Characteristics of the hemodynamic response(HRF) and how this affected the sequentialdevelopment of fMRI paradigms and influencesstudy design

• Sense of important design issues

What (good) is fMRI?

What it can tell you:

• Relative local “neural” activity (LFP’s ?)• NOT absolute neural activity• NOT excitation vs inhibition• NOT about necessity of a given region for a task• NOT fine-grained temporal information

Key Points





Subtraction ParadigmDonder’s method:

Ex: How to measure time of a mental transformation?

A random series of A’s and B’s presented and the subject must:

1. Respond whenever an event occurs (RTi)

2. Respond only to A not to B (RTii)

3. Respond X to A and Y to B (RTiii)

RTi = RT(detect) + RT(response)

RTii = RT(detect) + RT(discrimination) + RT(response)

RTiii = RT(detect) + RT(discrimination) + RT(choice) + RT(response)

THUS, RT(discrimination) = RTii - RTi

RT(choice) = RTiii - RTii

Criticisms of Subtraction Paradigm

1. That we already know what ‘counts’ as a single mentalprocess (i.e. choice is a single mental process?)

2. Assume that adding components does not affect otherprocesses (i.e. assumption of pure insertion)

THUS, one should pick tasks that differ along ONEdimension (either change the task OR the stimuli but notBOTH!)

And a resting baseline is good to include, however, theinterpretation should be taken lightly…(more later)

The loose task comparisonThe loose task comparison

Does Does notnot hold all variables constant BUT: hold all variables constant BUT:

(1) Uses a low level reference task(1) Uses a low level reference task

(2) Allows the data to be examined for predictable(2) Allows the data to be examined for predictablestimulus or response drivenstimulus or response driven activations activations

(3) Allows the more extensive activation pattern to(3) Allows the more extensive activation pattern tobe observedbe observed

The “loose” TaskThe “loose” TaskComparisonComparison

STISTISTIPREPREPRE

DOGDOGDOG

+++ +++ +++ +++

GREGREGRE

STISTISTIPREPREPRE

PAINTPAINTPAINT STRSTRSTR

STISTISTIPREPREPRE

EGGEGGEGGTASK 1TASK 1

TASK 2TASK 2

The tight task comparisonThe tight task comparison

Try to hold all variables constant including:Try to hold all variables constant including:

•• Stimulus display (nominally or statistically) Stimulus display (nominally or statistically)

•• Response and response selection characteristics Response and response selection characteristics

•• Performance level- especially if comparing cohorts Performance level- especially if comparing cohorts

•• Eye movements Eye movements

•• Emotional state (minimize anxiety and boredom) Emotional state (minimize anxiety and boredom)

The “tight” TaskThe “tight” TaskComparisonComparison

STISTISTIPREPREPRE

DOGDOGDOG

+++ +++ +++ +++

GREGREGRE

STISTISTIPREPREPRE

PAINTPAINTPAINT STRSTRSTR

STISTISTIPREPREPRE

EGGEGGEGGTASK 1TASK 1

TASK 3TASK 3

TASK 2TASK 2

22 mmiinnuuss 11TASK 1TASK 1 TASK 2TASK 2

BRAIN AREAS THAT DIFFERBRAIN AREAS THAT DIFFER

ALL ACTIVE BRAINALL ACTIVE BRAINAREASAREAS

Thanks to Randy Buckner

Example...

Interested in semantic processing and how it affectsmemory...

Parameters to specify in any experiment

1. Subjects: normal vs special populations

2. What part of brain look at? How many slices canyou have for your TR?

3. Choosing your TR: How often can you take a full setof pictures

4. What coil will you use?surface coils: higher SNR, only partial coveragehead coils: lower SNR, complete coverage

5. Toggle many times between conditions within a scan

6. Run as many scans as possible within a subject

Key Points





Visual Stimulation - 2 sec Flashes

0 10 20 30 40 50 60 70 80 90 100 110 120

TIME (seconds)

Per

cent

Sig

nal C

hang

e

0.0

1.0

2.0

3.0

4.0

-1.0

[Blamire, Ogawa et al., PNAS, 1992]

Visual Cortex During Brief Visual Stimulation

0.0

1.0

2.0

5 0 5 10 15

34

100

1000msec

TIME (seconds)

msec

msec

[Courtesy of Robert Savoy, Kathleen O’Craven MGH-NMR Center]

Blocked design fMRI

BLOCKED:

HORSE

(abstract or concrete?)

love

(upper or lowercase?)

“Blocked” fMRI:Memory Paradigm

STIPRE

HORSE

+ +

0 8 74 118 140 184 206 250

GRE

STIPRE

LOVE STR

STIPRE

HUMOR

TIME (SEC)52

+ +

STR

STIPRE

CHAIR

+

[Wagner et al., OHBM, 1998]

Typical Blocked-Design Response

-0.25

0

0.25

0.50

Per

cent

Sig

nal C

hang

e

0 20 40 60 80 100 120 140 160 180 200 220 240

Time

HORSE LOVE HUMOR CHAIR

13 SlicesPer

Brain Image

Thanks to Robert Savoy

80 brain imagesper 4 minute

run


For purposes of illustration…….


Examine thedata fromone slice

of the brainas a function

of time


Are these voxellevels statistically

different from each other?

Combinethese

Combinethese


Combinethese

Combinethese

Consider EACH voxelacross all time points


Typical Blocked-Design Response

-0.25

0

0.25

0.50

Per

cent

Sig

nal C

hang

e

0 20 40 60 80 100 120 140 160 180 200 220 240

Time

HORSE LOVE HUMOR CHAIR

Event-Related fMRI

BLOCKED:

SPACED EVENT-RELATED:

16 sec

“Spaced Event-Related” fMRI:Language Paradigm

[Buckner, Bandettini et al., PNAS, 1996]

TIME (seconds)

0 32 64 96

COU PRE TRA AFT STA PEL DRI

“Single-Trial” Response Across a Run“Single-Trial” Response Across a Run

TIME (SEC)TIME (SEC)

00 3232 6464 9696 128128 160160 192192 228228 256256

375375

376376

377377

378378

379379

380380

381381

ME

AN

MR

SIG

NA

LM

EA

N M

R S

IGN

AL

“Event-Related” Selectively Averaged Response

0

.5

0 1 2 3 4 5 6 7 8 9 101112131415

TIME (seconds)

PE

RC

EN

TS

IGN

AL

CH

AN

GE

375

376

377

378

379

380

381

ME

AN

MR

SIG

NA

L

374

375

376

377

378

379

0 2 4 6 8 10 12 14TIME (sec)

Broca’s Area During Language Paradigm


“Rapid Event-Related” fMRI

BLOCKED:

SPACED EVENT-RELATED:

16 sec

RAPID EVENT-RELATED:2 sec

20 sec0 sec

1 secon

Assessing the Linearity Hypothesis

[Dale and Buckner, Hum. Brain Map., 1997]

Response to Averaged Single Trials

-1

0

1

2

3

4

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19TIME (sec)

PE

RC

EN

T M

R S

IGN

AL

Split-half reliability


20 sec0 sec

20 sec0 sec 5 sec

Assessing the Linearity Hypothesis:5 Second ITI


-1

0

1

2

3

4

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19TIME (sec)

PE

RC

EN

T M

R S

IGN

AL

Response to Averaged Double Trials


0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

TIME (SEC)0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

TIME (SEC)

Assessing the Linearity Hypothesis:Separation of Responses


-1

0

1

2

3

4

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19TIME (sec)

Assessing the Linearity Hypothesis:Separation of Responses

FIRST TRIAL

ESTIMATEDSECOND TRIAL


20 sec0 sec

0 sec2 sec 20 sec

0 sec2 sec 20 sec4 sec

Assessing the Linearity Hypothesis:2 Second ITI


-1

0

1

2

3

4

5

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19TIME (SEC)

-1

0

1

2

3

4

5

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19TIME (SEC)

RAW DATA ESTIMATED RESPONSES

Responses to Multiple Rapidly Intermixed Trials

THREE TRIAL

TWOTRIAL

ONETRIAL

FIRST TRIAL

ESTIMATEDSECOND TRIAL

ESTIMATEDTHIRD TRIAL


Structuring Event-Related Trial Presentations

[Burock et al., Neuroreport 1998]

Fixed Interval Presentation

Randomized Presentation

ITI = 16 sec 3 sec 1 sec

Nor

mal

ized

% S

igna

l Cha

nge

8

6

4

2

0

8

6

4

2

00 100 200 300 400

Time (sec)

Variance Associated with Fixed Interval Designs

Seven unknowns, BUTonly three independentequations

INDIVIDUALBOLD

RESPONSE

TIME

h7 + h4 + h1

MEASUREDBOLD

RESPONSE

h6 + h3

h5 + h2

h4 + h1 + h7

h3 + h6 h2 + h5

h1 + h4 + h7

= TR

6 12 18 24 30


Variance Associated with Jittered Designs

INDIVIDUALBOLD

RESPONSE

TIME

Seven unknowns, ANDmore than sevenindependent equations

MEASUREDBOLD

RESPONSE

h7 + h5 + h3

h6 + h4 + h2

h5 + h3 + h1

h4 + h2

h3 + h1 + h7 h2 + h6

h1 + h5

h7 + h4

h6 + h3


Does the neural correlate of priming vary with the lagbetween the first and second episode within asemantic task?

PEACE+

ANVIL

PEACE+

TABLE

+ENERGY

ANVIL~2 min

~25 hr

ABSTRACT or CONCRETE?

Sorting Based onExperimenter Determined Conditions

[Wagner et al., J. Cognitive Neuroscience 2000]

Rest

Rest

Novel

Long Prime

Short Prime

Long Prime

Short Prime

Novel

Rest

Rest

Novel

Long Prime

Short Prime

Long Prime

Short Prime

Novel

Group datatogether

Shorter Lags Yield Greater Neural Priming

Anterior IFG

-0.1

0

0.1

0.2

0.3

% SignalChange

0 2 4 6 8 10 12 14Time (s)

ShortLongNovel

Thanks to Anthony Wagner

TRUTH

HOUSE

PEACE

ANVIL

TAPEREMEMBERED

FORGOTTEN

Sorting Based on Subject Behavior:Subsequent Memory Performance

[Wagner et al., Science 1998]

Neural Regions Predicting Subsequent Memory

ForgottenRemembered

Inferior Prefrontal Gyrus

-1

01234

0 2 4 6 8 10 12 14Time (s)

-101234

0 2 4 6 8 10 12 14Time (s)

B

Left Posterior Parahippocampus

Thanks to Anthony Wagner

Key Points





Critical issues in paradigm designCritical issues in paradigm design

•• Poorly defined Poorly defined neuroanatomical neuroanatomical hypothesishypothesis

•• Poorly controlled baseline Poorly controlled baseline

•• Attentional Attentional effects effects

•• Learning effects Learning effects•• Stimulus habituation or sensitization Stimulus habituation or sensitization

•• System and physiological drift System and physiological drift

Baseline, what is it?

Ex: if want to say something about verb generation andcompare it only to reading aloud..

BUT, still do not know if these regions are involved inreading only (thus can include a low level reading condition..)

No inherent “0” baseline for cognition,i.e. what are subjects doing when asked to do nothing?

• Ans: they are doing a lot• how interpret deactivations?

Issues: Generality vs SpecificityHypothesis: Region X is involved in process Y.

Evidence: Region X is activated when subjects do an instanceof process Y

Problem: Without running several further conditions, we can’ttell whether region X might instead be involved in somethingeither more SPECIFIC or more GENERAL than Y.

Example:

Hypothesis Space: Function of Region X

hands

feeteyes

2-toned faces

facesProfileFaces

FacesW/ohair

IDphotos

Any human body part

Anything animate

Thanks to Nancy Kanwisher

Issues: Attentional ConfoundsA given region might respond more strongly in condition A thancondition B simply because A is more interesting/attention-capturing than B.

Solutions:

1. Double Dissociations, i.e. faces versus objects?

2. Test conditions with opposite attentional predictionsi.e. passive viewing vs 1-back task

Thanks to Nancy Kanwisher

Issues: Statistical Significance vs.Theoretical Significance

P levels alone are not sufficientFor example, the FFA may respond significantly more to pineapplesthan watermelons, but the response to pineapples might nonethelessbe much lower than the response to faces.

Solutions:Quantity effect size, e.g. with percent signal change

Provide “benchmark” conditions within the same scan to give thesemagnitudes meaning

Objects Watermelon Pineapple Faces

0.6

0.6

0.7

0.7

0.9

1.8

2.0

2.0Thanks to Nancy Kanwisher

Acknowledgements

Anthony Wagner

Nancy Kanwisher

Randy Buckner

Robert Savoy

Randy Gollub

Chantal Stern

Data AnalysisA. General Issues• Individual vs Group Analyses: brains are very different BUT wantTo make a general claim ANS: do both if can

• Multiple Comparisons….if doing 20,000 T-tests, better not acceptp< .05

B. Methods• Simple comparisons, is X > Y ?, look in each voxel..

• Conjunction Analyses, are any voxels significant for both X>Yand A>B?

• Regression Analyses, obtaining weights for different regressors

• ROI-based Analyses

[Boyton et al., J. Neuroscience 1996]

Hemodynamic Response Summation:Linear Systems Approach

The fMRI response to a stimulus lasting a duration of NT is roughly alinear summation of N temporally shifted responses to a stimulus lastinga duration of T

“Mixed” fMRI

BLOCKED:

RAPID EVENT-RELATED:2 sec

MIXED:

[Chawla et al., Nat. Neuroscience 1999; Donaldson et al., NeuroImage, 2001]

TIME (SEC)

= OLD WORD = NEW WORD

MEMORY

FIXATION

MEMORYA

B

FIXATION FIXATION

Mixed Blocked/Event-relatedDesign

Donaldson et al., NeuroImage, 2001 (see also Chawla et al., Nature Neurosci., 1999)

TIME

TRANSIENT BOLDRESPONSETO EVENTS

MEASUREDBOLD

RESPONSE

SUSTAINED BOLDRESPONSE

TO SET

TIME

“Mixed” fMRI:Trial Separation with Task Blocking

Task 1 Task 2 Task 1

Analysis Strategies:

– Event-related analyses

• Task 1 trials ( ) vs. Task 2 trials ( )

• Trial type A ( ) vs. Trial type B ( )

[Badre et al., in prep.]

“Mixed” fMRI:Trial Separation with Task Blocking

Task 1 Task 2 Task 1

Analysis Strategies:

– Event and State effects• Same event contrasts

• Also model NULL components within blocks to explore “state” effects

CAVEAT: correlation between event and state regressors

vs vs

Task 1 vs. Task 2 Nulls

fmri experimental design - mitweb.mit.edu/hst.583/www/course2001/lectures/hst583_lect...because fmri...

Documents