Presentation is loading. Please wait.

Presentation is loading. Please wait.

Efficiency – practical Get better fMRI results Dummy-in-chief Joel Winston Design matrix and.

Similar presentations


Presentation on theme: "Efficiency – practical Get better fMRI results Dummy-in-chief Joel Winston Design matrix and."— Presentation transcript:

1

2 Efficiency – practical Get better fMRI results Dummy-in-chief Joel Winston Design matrix and

3 Experimental design & efficiency Getting the “right” results for a given amount of scanner time requires maximising your efficiency in detecting the experimental effect Because of the temporal smoothing that the HRF applies in translating neural responses to BOLD signal, we know something a priori about how to maximise an experimental effect As Paul has shown (?), mathematically the block design turns out to be highly efficient, essentially by maximising the experimental variance within a time frame that escapes two filters: HRF (low pass) and SPM (high pass)

4 Temporal filtering and neuroimaging As mentioned, there are two components to temporal filtering routinely applied to fMRI data, one by the brain, the other by us… The brain’s temporal filter is the Haemodynamic Response Function (HRF), whose form we all know and love:

5 HRFHRF – power spectrum What this means in reality is that the HRF acts as a low pass filter on our recording of the brain’s activity Peak at ~0.04Hz => Max sensitivity for designs with on-off cycles of 0.04 -1 = 25s

6 Why the high-pass filter? We routinely apply a high pass filter in SPM The reason for this is simply because we can For the most part, we use SPM to analyse designed experiments where we have some control over the interesting parameters, and little control over disinteresting ones Disinteresting parameters are often slow-moving things, like scanner drift, physiological noise, and occur outside the temporal space of designed experiments So we get rid of these by high-pass filtering the data but not so severely that we lose our experimental effects…

7 How can I check that I’m not losing anything interesting by high pass filtering?

8 The importance of being event-related When is it necessary/advantageous to use event-related designs? 1.Trials that by definition can’t be blocked e.g. oddballs 2.Post-hoc classification e.g. classification by memory, parametric scores, subjective perception 3.Randomise trial order Where phasic/tonic effects might be dissociable Where anticipation/predictability might be a problem

9 OK, so you’ve persuaded me that I have to use event-related fMRI for my experiment (the neural correlates of doughnut eating…) How do I make the most of an event-related paradigm? Two things that we’ll talk about: 1.Spacing of events 2.Sequences of events

10 The spacing of events Simulations show that efficiency to detect differential effects between event types increases with shorter SOAs:

11 But I’m also interested in detecting main effects relative to baseline (“evoked responses”)… Consider including null events as an extra event type:

12 So the bottom line is… …pack it in!!

13 But my events have to be spaced out! Then you might want to consider not randomising event orders, but having them alternate or nearly alternate (permuted designs):

14 Planning in advance… One of the best ways to increase the efficiency of event-related designs is to ensure mini-runs of same stimuli… …and one way of ensure mini-runs is to modulate the probability of different event- types over experimental time

15 Stochastic designs Essentially a stochastic design defines a (variable) probability of a given event type at each SOA min Stochastic designs can be stationary or dynamic One incarnation of dynamic stochastic designs (implemented in SPM99) is to modulate the underlying probability of events at each SOA by a sine wave:

16 How will this translate into an event train? (Not that sort of train, dummies)

17 This sort of train:

18 A practical example Faces vs scrambled faces SOA was fixed at 2.97s TR was 2.5s Three runs of 128 scans: –Blocked faces and scrambled faces –Fully randomised stimulus order –Modulated probability of face/scrambled face Task was detection of very infrequent (1%!) targets (chairs)

19 The design matrix F S CF S CF S C F = faces S = scrambled faces C = chairs M =movement parameters M MM Blocked Fully randomised Dynamic stochastic

20 Calculated efficiency for the 3 sessions

21 Superficial comparison between sessions Fully randomised

22 Superficial comparison between sessions Blocked Randomised Dynamic stochastic

23 Right anterior fusiform (36,-24,-30) Results – Interaction of efficiency type and faces vs scrambled faces Blocked Dynamic stochastic Randomised Differential effect (faces > scrambled faces) Left STS (-57,-33,6) Posterior cingulate (-9,-54,42) Right anterior fusiform Visual cortex (-12,-78,-6) Blocked Dynamic stochastic Randomised

24 Results – Chairs vs other visual stimuli Occipital pole Anterior cingulate “Chair” area Dorsal occipital pole

25 Results – Chairs vs other visual stimuli Raw time series from anterior cingulate: 5 10 15 200

26 …and another thing:

27 Take home messages Efficiency can be estimated before you do your study to allow a comparison between different designs However, on any one implementation, a given design may prove less successful in detecting effects than another, less efficient design Psychological validity is an important design constraint


Download ppt "Efficiency – practical Get better fMRI results Dummy-in-chief Joel Winston Design matrix and."

Similar presentations


Ads by Google