Presentation is loading. Please wait.

Presentation is loading. Please wait.

: To Block or Not to Block?

Similar presentations


Presentation on theme: ": To Block or Not to Block?"— Presentation transcript:

1 : To Block or Not to Block?
Basic fMRI Design : To Block or Not to Block? Tor D. Wager Columbia University With help from:

2 What we want from a design
Power: Can I detect results? Interpretability: Can I relate brain data to specific psychological events? Memory retrieval and comparison processes associated with recognition Robust to errors in model specification Timing, hemodynamic response (HRF) shape Experimental (A) Control (B)

3 Types of designs: Blocked and event-related
Block design: Similar events are grouped Event-related design: Events are mixed If these were reaction time (RT) data, they would be equally statistically good. (The only concerns would be psychological). But for fMRI designs, they are very different!

4 Power in linear models estimate standard error
Power: probability of detecting a true effect Function of statistical, psychological, and physical constraints Powerful designs maximize test statistics where true effects are present: estimate standard error Psychological: Maximize true effect Physical: Reduce error by reducing noise Statistical: minimize estimation uncertainty due to design; maximize efficiency See work by Dale, Henson, Liu, Friston, others

5 Design Efficiency To minimize standard error: Reduce noise
Make design more efficient Efficient designs: Maximize variance of predictors Minimize covariance among predictors In fMRI designs: Formula is more complex, principle is the same Factor in contrasts, filtering and autocorrelation What is it? A-optimality

6 Maximize variance of predictors
Maximize variation along x-axis of the plots below (Rescaling doesn’t count) predicted data Balanced: Max. x variance predicted data No x variance predicted data Low x variance Principles: Keep equal numbers in high and low predicted groups Concentrate on extremes

7 Minimize covariance among predictors
Avoid confounds and partial confounds Fully confounded Fame Sex Y N F M Partial (low power) Fame Sex Y N F M Orthogonal Fame Y N Sex F M

8 fMRI Data Hemodynamic response (HRF) is delayed and prolonged in time
Stimuli are often convolved with an HRF Data is typically autocorrelated: low-frequency drift Data and predictors are low-pass filtered These two kinds of filtering operations make order and timing of stimuli critically important

9 Efficient (powerful) fMRI designs
Depends on what you’re trying to measure Main effect (Task vs. implicit baseline) Difference between conditions Assumed HRF shape or estimates of shape Can test different designs with efficiency metric First discuss assumed shapes, then talk about estimating shape

10 Convolution refresher
Time (s) A B C D Indicator functions (onsets) Assumed HRF (Basis function) Design Matrix (XT) A B C D Time (s) Time Design Matrix (X) A B C D Convolving with this HRF blurs high frequencies Changes variance & covariance of predictors Assumed “neural” event duration+shape, HRF shape, linear summation of responses to events

11 Blocked vs. ER power Consider Famous (red) vs. NonFamous (green)
predictors Event indicators Block Event-related

12 Blocked vs. ER power: higher contrast variance = higher power
Red-Green: [1 -1] contrast HIGH POWER Red+Green [1 1] “contrast” LOW POWER Block Red-Green: OK POWER Event-related Red+Green OK POWER

13 Summary so far Block designs are efficient…if
Balance time on expt’l and control conditions, including rest

14 fMRI designs: Block length matters
Rise and fall: High predictor variance means efficient design 2 sec blocks: Too fast! Signal is blurred away by convolution 50 sec blocks: About the same

15 High-pass filtering Reduces noise Reduces efficiency
Removing low frequencies from design and data Often done using nuisance covariates Tradeoff in precision: Reduces noise Reduces efficiency

16 High-pass filtering fMRI Noise: Time domain Frequency domain
HP filter 1/60 = .016 Hz Unfiltered Filtered

17 Efficiency in block designs
18 s blocks, 80 s filter Filtering reduces efficiency… (But you’re removing noise, too) Best design depends on noise structure (and HRF) Low autocorrelation: 18 s block, 80 s HP filter Higher + autocorrelation: Shorter blocks (12 s) and 60 s filter

18 Summary so far Block designs are efficient…if
Balance time on expt’l and control conditions, including rest Choose appropriate block length and filtering options

19 HRFs vary across regions
Checkerboard, n = 10 Thermal pain, n = 23 HRF shape depends on: vasculature time course of neural activity Stimulus On Aversive picture, n = 30 Aversive anticipation See Schacter et al. Aguirre et al.

20 How robust are blocked and ER designs to variation in HRF?
Depends on how HRF is modeled ER designs can flexibly estimate HRF shapes, making them robust First look at power when the predicted and actual HRFs don’t match Then look at basis functions, which allow flexible estimation

21 HRF mismatch in blocked and ER designs
What happens when the true HRF does not match the assumed one? Simple case: mis-specification of onsets

22 HRF mismatch in blocked and ER designs

23 Basis sets Image of predictors Data & Fitted Canonical Single HRF
HRF + derivatives Finite Impulse Response (FIR) Time (s)

24 Comparing efficiency for different design types
Block best for detection M-sequence best for shape (Buracas et al.) Event-related designs so-so on both Optimized designs good tradeoff Greve, OptSeq Wager, Genetic Algorithm Liu Block, 16 s on/off Theoretical limit Optimized (GA) Contrast detection power Event-related m-sequences HRF shape estimation power

25 Summary so far Block designs are efficient…if
Balance time on expt’l and control conditions, including rest Choose appropriate block length and filtering options Get the HRF shape right for the brain area of interest Event related designs are good for… Flexible modeling of HRF shapes (both + and - for power) Mixed/hybrid designs are good for… Balancing power and shape estimation

26 Interpretability Block designs do not inform about whether activity is related to specific psychological events Case study: Face recognition; compare Famous-Nonfamous 3 s picture viewing Recog: 250 ms “I used to wear Batman underoos…” “What was he in?” “What a cool movie!” “Get back on task”

27 Interpretability Differences between famous and nonfamous faces in any of these processes could show up in block contrast More face viewing time does not equal more power! Need time on task of interest Recog: 250 ms “I used to wear Batman underoos…” “What was he in?” “Get back on task” “What a cool movie!”

28 Interpretability Blocking trials may change subjects’ strategy
Stroop task: “name the color of the print” Control block: green yellow red blue Experimental: red blue green yellow Block: more word reading in control Go-no go task: “press fast, but withhold if X” N P R S X X X X Many tasks cannot be blocked

29 Interpretability Block design power and interpretability depends on actual time on task (duty cycle) Subjects should be doing the mental operation you want them to during the whole block Case study: Error detection / correction Researchers develop interference task with up to 25% errors Researchers want power: “Let’s use a block design” 20 s blocks, 6 trials per block Compare high error blocks (25%) vs. low (5%)

30 Interpretability: Case study
Block predictions would look like this: But the true response in an error-selective brain region might look more like this: The block design has high efficiency, but low power because it makes inaccurate predictions (r = 0.42)

31 Summary Block designs are efficient…if
Balance time on expt’l and control conditions, including rest Choose appropriate block length and filtering options Get the HRF shape right for the brain area of interest Task performance & strategy are not context-sensitive Event related designs are good for… Flexible modeling of HRF shapes (both + and - for power) Better identification of activity related to specific events When unpredictability is important Mixed designs are good for… Balancing power and shape estimation

32 ER design practical advice
What is the best inter-trial interval (ITI)? “Jitter” (include rest events) or no jitter? 1% rest events 41% rest events 81% rest events ISI in sec Efficiency of contrast [1 -1] Assuming Linear Responses 2 4 6 10 14 Randomize order, present as fast as you can, no rest

33 fMRI data nonlinear at short ISIs
Meizin et al. (2000): 10% nonlinear saturation at 5 s ISI Design 1,2,5,6,10 or 11 visual flashes, 1s ISI, then 30s rest Note actual vs. predicted relative magnitude Wager et al., 2005

34 ER design practical advice
Nonlinear response model For A - B About 5 s, on average, between reps of the same event, no rest 1% rest events 41% rest events 81% rest events I like at least 4 s between events; conservative on linearity Efficiency of contrast [1 -1] For A + B About s, on average, between events (I.e., use jitter) 2 4 6 10 14 ISI in sec Wager & Nichols, 2003

35 Download the Genetic Algorithm toolbox at:
Thank you! Download the Genetic Algorithm toolbox at:

36 Design efficiency in fMRI
Formula is more complex, principle is the same Factor in contrasts, filtering and autocorrelation Define: filtering matrix = K, autocorrelation matrix = V Matrix whose rows contain a set of contrasts = C filtered design matrix = Z Z-: pseudoinverse of Z = inv(Z’Z)Z’ K * X = Z Not equal to power, but can be converted to power given effect sizes See Friston et al., 2000; Zarahn, 2001

37 Nonlinearity in BOLD signal

38 Pros and cons of blocking
+ High power, if parameters chosen correctly + Simple to implement + Relatively robust to changes in HRF shape - Predictable events may change task strategy and activity patterns - Cannot infer activity related to specific psychological events - Power limited if Ss are not doing cognitive operation of interest throughout blocks


Download ppt ": To Block or Not to Block?"

Similar presentations


Ads by Google