FMRI Experimental Design

     Studies of functional magnetic resonance imaging (FMRI) employ various 
types of experiment designs, most of which fall into one of two categories: 
1) block design, or 2) event-related design.  In this 'How-To', block design and 
various forms of event-related design will be presented, along with the 
advantages and disadvantages of using each paradigm.  Below is an outline of the
experiment designs that will be discussed, along with information regarding 
randomization of stimulus trials and fixed versus stochastic or "jittered" 
inter-stimulus intervals (ISI) in rapid event-related design:

	1) Block Design

	2) Event-Related Design (ER-FMRI)
	     A. "Slow" Event-Related Design
	     B. "Rapid" Event-Related Design

		1. Fixed ISI and NON-RANDOMIZED stimulus presentation

		2. Fixed ISI and RANDOMIZED stimulus presentation

		3. "Jittered" ISI and RANDOMIZED stimulus presentation



     Block design was the first type of experimental paradigm to be used in FMRI
     research, as well as the first to involve more complex statistical analysis.
     It is still the most commonly used experimental paradigm in FMRI studies.

     The block design consists of several discrete epochs of on-off periods, 
     with the "on" representing a period of stimulus presentations, and the 
     "off" referring to a state of rest or baseline.  Although blocks may range 
     in duration from 16 seconds to a minute or more (average is about 20-30 
     seconds), they all share the same basic on-off pattern.  These on-off 
     states are alternated throughout the experiment to ensure that signal 
     variation from small changes in scanner sensitivity, subject movement, or 
     attention shifts have a similar effect on the signal responses associated 
     with each of the different states.  Below is an example of a block design 
     with two experimental conditions: pictures (red blocks) and words (green 
     blocks), along with their resulting hemodynamic response functions.

     Once an experiment has been run and the data have been collected, the 
     appropriate statistical analysis must be implemented.  With block design, 
     individual trials are not compared.  Rather, the underlying hemodynamic 
     responses acquired during one blocked condition are compared to the 
     signals acquired from baseline, or from other blocks involving different 
     task conditions (e.g., "picture" blocks versus "word" blocks).  As such, 
     regions of signal activity that change between one condition and another 
     can be identified with considerable statistical power.

     Advantages of Block Design:

	* A simple block design is adequate for many types of experiments,
 	  especially in early, exploratory stages of research projects.

	* Block designs allow for considerable experimental flexibility, 
	  allowing parametric designs and multi-factorial designs to be 

	* Block design can be especially advantageous and a good starting point 
	  for newcomers to the field of FMRI research.

	* Block design is statistically powerful and straightforward to analyze,
	  as the shape of the response function can be assumed to be simple.

     Disadvantages of Block Design:
	* Block design can be predictable and boring, making it prone to 
	  potential confounds such as rapid habituation, anticipation, set, or 
	  other strategy effects.

	* It may be difficult to control a specific cognitive state for the 
	  relatively long periods of each block.  A 'rest' state is rarely true 
	  rest, as the mind may wander in a subject who is not engaged in a 
	  specific task.

	* Information regarding activation response time courses cannot be 
	  obtained with block design because individual responses are lost 
	  within the block.

	* The high predictability of block design makes it inappropriate for 
	  certain cognitive tasks, such as an 'oddball' paradigm where a 
	  reaction to an unexpected stimulus is probed.

	* The BOLD signal may not remain constant across the epoch of interest. 
 	  Within a block, the underlying hemodynamic response can change from 
	  the first trial in the block to the last trial within the block.  This
	  result may be a consequence of anticipatory effects.

	* Block design is not feasible for certain patient populations.  For 
	  instance, hallucinatory schizophrenics who often display irregular or 
	  uncontrollable behaviors cannot be forced into a block design.



     Event-related designs associate brain processes with discrete (rather than 
     blocked) events, which may occur at any point in the scanning session.  
     That is, different trials or stimuli are presented in arbitrary sequences. 
     This type of design mimics the format of a behavioral study more closely 
     than block design.  With behavioral studies, stimulus events such as 
     pictures or words are presented one at a time, usually in a randomized 
     fashion, and separated by an inter-stimulus interval of a specified length.
     Advantages of ER-FMRI:

	* This type of design allows for stimulus events from various experiment
	  conditions (e.g., conditions A, B, and C) to be presented randomly in 
	  one run.  This type of scenario is not possible with block design.

	* By detecting signals to individual trial events, ER-FMRI can parallel 
	  behavioral studies by examining responses to individual trials rather 
	  than blocks of trials.

	* Event-related paradigms allow for greater flexibility and 
	  randomization than block design, leading to more clever and less
	  predictable experiments.

	* Unlike block design, ER-FMRI allows the experimenter to estimate the      
	  hemodynamic response function from a single event type. The 
 	  hemodynamic response can be identified by averaging data acquired 
	  after many discrete events.  This approach is more powerful than 
	  block design because it allows considerable flexibility for 
	  determining, for example, responses to novel or aperiodically 
	  presented stimuli, or exploring changes over time.

	* Post-hoc sorting of stimulus trials can be done with ER-FMRI (e.g., 
	  correct vs. incorrect responses, aware vs. unaware, remembered vs. 
	  forgotten items, quick vs. slow response times, etc.).

     Disadvantages of ER-FMRI:
	* The event-related design requires a greater understanding and grasp of
	  functional MRI because the design and statistical measures that follow 
	  are more complex than those of block design.  As a result, a newcomer 
	  to the world of FMRI may initially encounter some difficulty when
	  attempting to design an event-related study.

	* One major disadvantage of event-related design involves the signal-to-
	  noise (SNR) ratio.  The timing of single events results in a lower SNR
	  for event-related FMRI.  Specifically, for block design, the percent 
	  signal change may be in the range of 3% to 5% while for event-related 
	  design, it may be less than 1%.  To compensate for this loss in 
	  statistical power, the number of trials should be increased by 
	  approximately 50 to 100 trials per condition.  The result, however, is
	  longer scanning runs (more on this issue in the sections that follow).

     The inter-stimulus interval between stimulus trials can vary, and this time
     interval determines whether the event-related design is identified as 
     'SLOW' or 'RAPID.'  An explanation of slow and rapid event-related design 
     is provided in the sections below.

     Within the scanner, a patient's exposure to a stimulus event may result in 
     a significant increase in brain activation, which is correlated with 
     localized changes in blood flow, oxygenation, and volume.  These local 
     increases in blood flow and microvascular oxygenation take some time to 
     occur.  The result is a delay in onset of the BOLD signal, which evolves 
     over an extended period of time, even for brief neuronal events.  In fact, 
     the "plateau" of the hemodynamic response may not occur until 6-9 seconds 
     after the stimulus onset.  The result will be a hemodynamic response 
     function that is spread out, usually far beyond the stimulus duration.  
     This phenomenon is known as "dispersion."  On average, one should expect 
     the BOLD signal to rise and fall within 12-20 seconds.  When implementing 
     an event-related design, one must consider this post-stimulus delay of the 
     BOLD signal.  Stimulus trials spaced too close together will result in an 
     overlapping of their respective hemodynamic response functions, causing
     them to become "tangled" or convolved.  When this happens, more sophisticated
     statistical measures are required to deconvolve the data.

     With SLOW event-related design (a.k.a. 'widely spaced' or 'simple' event-
     related design), the individual stimulus trials are spaced far apart in 
     time to prevent overlap of their hemodynamic functions.  In other words, 
     the hemodynamic response that results from a single trial is allowed to 
     rise and fall completely before the next trial begins.  Below is an
     illustration of a slow event-related design, along with the resulting
     hemodynamic response functions.


     Advantages of slow event-related design:	

	* Since there is no overlap of the hemodynamic responses, slow event-
	  related paradigms do not require deconvolution analysis and are 
	  therefore fairly easy to analyze statistically.

     Disadvantages of slow event-related design:

	* The long rest periods between stimulus presentations mitigate 
	  habituation, expectation, and boredom, which can taint the experiment 
	  with anticipatory effects.

	* This type of design tends to be extremely time inefficient.  Since 
	  scanner time is limited, it is wasteful to spend so much time waiting 
	  for the hemodynamic response to return to baseline.  

	* In addition to being wasteful, a disproportionate amount of time at 
	  baseline results in the collection of less non-baseline data.  Since 
	  FMRI is a measurement of differences in response signal, it is perhaps
	  more efficient to get half of one's data at or near the baseline state,
	  and half at the non-baseline state.  If too much time is spent at 
	  baseline, the result is a good estimate of baseline ('small sigma' in 
	  statistical terms), but a bad estimate of the activation ('large 
	  Sigma').  If too much time is spent in activation, then the reverse is 
	  true.  However, an equal number of baseline and active trials will 
	  increase the likelihood of sucessfully detecting a statistically 
	  significant difference in response signal between baseline and
	  non-baseline states.

	* Compared to block design, the signal-to-noise ratio (SNR) is lost by 
	  approximately 33% in slow ER-FMRI.  As mentioned before, one way to 
	  compensate for this loss in statistical power is to increase the 
	  number of trials per condition for event-related averaging.  
	  Unfortunately, the additional trials will increase the experiment 
	  duration, thus taking up more scanner time.  Logistically, this option
	  may not be feasible when employing a slow ER design.  


     Rapid event-related design is similar to slow event-related design with the
     exception that it takes individual stimulus events and spaces them at close
     intervals.  For instance, the ISI may be set to as little as two seconds. 
     The result is a significant overlap of hemodynamic response functions that 
     must later be disentangled to determine the effect of each stimulus 
     condition on brain activation.
    Advantages of rapid event-related design:

	* The shorter resting gaps between events leads (hopefully) to a 
	  decrease in subject boredom.  Thus, rapid ER design is much more 
	  resistant to habituation, set, and expectation than slow ER paradigms.

	* Rapid stimulus presentation makes is possible to adequately squeeze in 
	  more stimulus trials per run, thus improving statistical power by 
	  increasing the number of responses to be averaged per unit of time.

     Disadvantages of rapid event-related design:

	* The signal-to-noise ratio loss is even greater for rapid event-related
	  design than it is for slow ER-FMRI.  Specifically, SNR loss is 
	  approximately 17% more for rapid ER than for slow ER, and 50% more for
	  rapid ER compared to block design. 

	* The decreased inter-stimulus interval results in hemodynamic responses
	  that overlap substantially.  Assuming linearity, the overlapping 
	  hemodynamic responses often found in rapid designs must be separated 
	  by a statistical process known as deconvolution.  In simpler terms, 
	  each individual hemodynamic response function must be disentangled so 
	  that the effect of each stimulus condition (say, conditions A, B, and 
	  C) can be differentiated and measured.  This requires greater 
	  statistical savvy and know-how on the part of the experimenter.  In 
	  addition, this overlap problem can only be resolved if the 
	  experimental design is properly randomized. (It should become quite 
	  clear in the next section why randomization of stimuli is so essential
	  in rapid event-related FMRI).

     Now that the basics of rapid event-related design have been covered, the 
     issues of proper randomization of stimuli and fixed versus "jittered" 
     inter-stimulus intervals will be discussed.  

     1. Rapid ER-FMRI: 
	Fixed ISI + NON-randomized stimulus presentation = BAD DESIGN

	     When using an event-related design, it is important to remember
	that the stimuli must be properly counterbalanced to ensure that each 
	trial type is preceded and followed by each trial type equally often.  
	If this does not happen, the result can be detrimental when it comes 
	time to run the statistics on the data.  Primarily, a short, fixed ISI 
	paired with a sequential ordering of the stimulus events can lead to a 
	problem known as "multicollinearity" or "identification problem".  The 
	problem of multicollinearity is illustrated below in Figure 3:

	As Figure 3 demonstrates, it is problematic to combine a short, fixed 
	ISI with a stimulus presentation that always orders the trials in the 
	same exact manner (e.g., A, rest, B, rest, C, rest, A, rest, B, rest,
	C, rest...).  In this example, the responses always overlap in the same
	way (A followed by B followed by C).  As a result, there is much
	ambiguity as to the source of the observed response.  Is the observed
	sum of the hemodynamic response functions due to stimulus A alone?  Is
	it due to the combined contributions of stimuli A and B?  A and C?  A,
	B, and C?  It is impossible to answer this question.  It is also
	important to note that multicollinearity is not a problem due to the
	limitations of statistical programs that calculate the deconvolution of
	time-series datasets (e.g., AFNI 3dDeconvolve).  Rather, the limitation
	is mathematical in nature.  In such a case, it is mathematically
	impossible to determine the contribution of each individual stimulus
	trial to the sum of the hemodynamic responses.

	Fortunately, the resolution to this dilemma can be simple.  When
	implementing a rapid ER design, it is important to randomize the 
	stimulus presentations so that every trial is preceded and followed by 
	every other trial type an equal number of times.  The 'AFNI_howto' 
	section of this HowTo provides a script that does this very thing using 
	the AFNI program 'RSFgen' (i.e., Random Stimulus Functions generator).

     2. Rapid ER-FMRI: 
	Fixed ISI + RANDOMIZED stimulus presentation = BETTER DESIGN

	     Figure 3 illustrates the detrimental effects of ignoring 
	randomization in rapid event-related design.  To successfully deconvolve
	the overlapping HRF's, a rapid ER design should include every possible 
	combination of trial sequences.  Since responses sum in an approximately
	linear fashion, the responses to rapidly presented stimuli can be 
	extracted from the data if the stimulus presentations are randomly 
	varied.  Figure 4 provides an example of such a design and the resulting

	In the above figure, the effort was made to randomize the stimulus 
	trials in a way that ensured successful deconvolution of the overlapping
	HRF's.  By properly counterbalancing the trials, one can now 
	mathematically determine the contribution of each stimulus condition on 
	the observed sum of the hemodynamic responses.

	Hopefully, these examples demonstrate how taking the time to carefully 
	design and execute an experiment is well worth the effort.  The end 
	result will be the collection of 'deconvolvable' data, which can be 
	properly analyzed and understood. 

     3. Rapid ER-FMRI with "jittered" ISI

	     Up to this point, this HowTo has illustrated examples of FMRI 
	designs involving inter-stimulus intervals that are "fixed."  In other 
	words, the ISI remains constant throughout the experiment.  However, 
	just as stimulus presentations can be randomized, so can the ISI.  With 
	a "jittered" or stochastic stimulus timing, the inter-stimulus interval 
	is randomized throughout the experiment.  The result is a varied onset 
	of successive stimulus events, with randomly intervening rest intervals.
	Below is an example of a rapid event-related design with a jittered 
	inter-stimulus interval:

	A differential ISI results in an even more differential HRF overlap,
	further reducing the probability that an experiment design will confront
	multicollinearity problems.  However, these types of designs will be more
	challenging to analyze.
	In some cases, differential ISI's are necessary because the timing of
	the stimulus presentations is determined by the subject.  Many
	behavioral studies implement this type of "self-paced" timing.  In other
	cases, jittered stimulus timing is incorporated into the experiment 
	design because it introduces more overall randomness to the study.  This
	can be a good thing considering subjects are inquisitive creatures who 
	are constantly attempting to "figure out" the experiment.  More 
	randomness and unpredictability significantly decreases anticipatory 


     While block design provides a fairly simple and straightforward approach to
     creating an experimental paradigm, the habituation and possible hemodynamic
     lag that afflict a design of this sort is problematic.  Furthermore, block
     design may not be feasible for certain types of cognitive tasks or patient 
     populations.  On the other hand, event-related design is more flexible and
     can introduce randomization of stimulus trials into the experiment, thus 
     mimicking behavioral studies more closely.  The downside is that analysis 
     of the data may be more complicated, particularly if the hemodynamic 
     response functions overlap significantly as in the case of rapid event-
     related design.  In addition, the signal-to-noise ratio decreases 
     dramatically with event-related design.  

     When implementing an event-related design, one must decide between a fixed or 
     "jittered" inter-stimulus interval, and appreciate the benefits of proper 
     counterbalancing of stimulus events.  One must also choose between a "slow"
     or "rapid" ISI, and consider the pros and cons of each.  While slow ER 
     design avoids the overlap of hemodynamic response functions, the longer ISI
     mitigates habituation and results in the collection of less non-baseline 
     data.  Conversely, rapid ER design is less boring and more non-baseline 
     data can be obtained, but the substantial overlap of hemodynamic response 
     functions reduces the SNR ratio and requires more complex deconvolution 

     Given the advantages and disadvantages of each research paradigm, many 
     researchers are opting for the best of both worlds by creating experiments 
     that incorporate both block and event-related design.  Ultimately, it is 
     up to the user to decide which particular design (or combination) is most 
     suitable for their experimental needs.  Whichever paradigm one decides to 
     implement, it is important that the experiment design be thought out and 
     planned well in advance.  One should consider which experimental paradigm 
     will be best for the cognitive phenomenon being examined and the patient 
     population being tested.  Likewise, one should consider which statistical 
     measures are appropriate for the data being gathered.  By doing so, the 
     scanning process, data collection, and statistical analysis will go much