Hi Hamed,
There are important design considerations that are specific to
multivariate analyses (decoding or similarity). We are currently
writing up a paper on this.
(1) Optimizing efficiency will have to incorporate not only the
contrast of interest, but also equal estimability of all conditions.
The estimability of the regressors plays a large role here and you
want that to be equal in all conditions you want to compare. If there
are differences in the estimability, you will have some patterns that
are noisier than others. If you classify a noisy pattern vs. a stable
one, then even with a linear classifier you will be able to do so
above chance, even if the centroids are on average the same. If you
compare the similarity between two stable patterns it will be higher
than the similarity between two unstable patterns or the similarity
between a stable and an unstable pattern. Jitter plays a role, because
it affects estimability at the trial level. That means single trial
analyses are not ideal when using jitter, but sometimes that is
unavoidable. As long as everything is balanced, this should only add
noise to your final estimation and the (open) question becomes what
the ideal trade-off is between better average estimability and larger
trial-by-trial variability in estimability. If you combine multiple
trials into one regressor, then make sure you balance the jitter
across trial types and also get roughly the same estimability.
(2) When controlling for potential confounds and using runwise
estimates, make sure you balance them within run. The story is a
little more complicated than this and relates to the fact that we are
using different, incompatible statistical approaches: classical design
principles for balancing confounds, but cross-validation for
estimating informational content.
Irrespective of the above, I would not jitter the delay period, but
balance the design. I'm also wondering whether it makes sense to
assume a canonical HRF for the delay period, or whether it would be
best to use FIR. We've worked quite successfully with FIR (Christophel
et al., 2012; 2014).
Hope that helps!
Martin
2015-11-04 15:56 GMT+01:00 hamed nili <[log in to unmask]>:
> thanks,
> I am thinking of something like:
> 750 ms :sample
> 2-4 seconds : (jittered) delay period
> 750 ms : test
> 2-4 seconds: (jittered ITI)
>
> Please note that I am not going to model 3 event types, I'll have two event
> types: sample and test.
> Do these numbers sound reasonable?
>
> cheers
> hamed
>
>
> On Wed, Nov 4, 2015 at 3:03 AM, MCLAREN, Donald <[log in to unmask]>
> wrote:
>>
>> I was assuming that the design was something along the lines of:
>> <1 second sample
>> 3-4+second delay
>> 1-2 second test
>>
>> These should be separable because their shapes are quite different. As the
>> times of each phase shrink, it becomes harder to separate.
>>
>> Additionally, you could test the efficiency of the design by building the
>> design matrix and estimating its efficiency to see if the trials can be
>> separated with the pre-specified number of trials.
>>
>>
>> Best Regards,
>> Donald McLaren, PhD
>>
>>
>> On Tue, Nov 3, 2015 at 9:53 PM, <[log in to unmask]> wrote:
>>>
>>> As an additional note, you might want to try using M-sequence to optimize
>>> your estimation efficiency since you are planning to do multivariate
>>> analysis.
>>>
>>> Also, I am not so sure that it is a good idea to model the three event
>>> types (sample, delay, test) as suggested by Donald, because their onsets are
>>> pretty close in time (within the same trial). In that case, I would expect
>>> strong correlation between the regressors in your GLM, which is extremely
>>> bad for your parameter estimation. Perhaps Donald has some further insights
>>> regarding this issue?
>>>
>>>
>>> Best,
>>>
>>> Ce
>>>
>>>
>>>
>>> ----- 原始邮件 -----
>>> 发件人:"MCLAREN, Donald" <[log in to unmask]>
>>> 收件人:[log in to unmask]
>>> 主题:Re: [SPM] jittering question
>>> 日期:2015年11月04日 10点35分
>>>
>>> Use optseq2 to jitter the ITI between trials. Do not jitter the time of
>>> the delay period unless you want to change the difficulty of each trial.
>>> This won't be ideal, but you could use the possible designs to build the
>>> correct design matrices containing the three event types (sample, delay,
>>> test) and compute the efficiency of them. Additionally, if you test it this
>>> way, you can modulate correct/incorrect trials to see which jittering leads
>>> to the best efficiency over a range of accuracy and trial orders.
>>>
>>> As the delay period will likely be at least a couple of seconds, the
>>> sample and test will be separable events.
>>>
>>> Best Regards,
>>> Donald McLaren, PhD
>>>
>>>
>>> On Tue, Nov 3, 2015 at 7:12 AM, hamed nili <[log in to unmask]> wrote:
>>>
>>> Dear all,
>>>
>>>
>>>
>>> I guess this might be a rather boring, old question about design (!) but
>>> I am looking for a short recommendation.
>>>
>>> The question is about jittering the ISI:
>>>
>>> I want to design a simple visual experiment (delayed match-to-sample
>>> task) and plan to do both univaraite and multivariate analysis on the data.
>>>
>>>
>>> There are two time windows that I am considering jittering of their
>>> length: the delay period (between sample and test) and the ITI.
>>>
>>> Having read some articles and posts on design power and efficiency, am I
>>> right that if I am happy with a canonical HRF, I could go for the following
>>> setting:
>>>
>>> - jittering the length of the delay period (to be able to estimate
>>> responses to both sample and test stimuli, modelling them as separate
>>> events)
>>> - jittering the ITI in an independent manner from the delay
>>> jittering (I first thought of jittering the ITI in a way that the sum of the
>>> delay and the ITI are constant, giving maximum power in estimating the
>>> response to the sample, but then thought this might make the sample special
>>> consequently).
>>>
>>> Thanks,
>>>
>>> Hamed
>>>
>>>
>>
>
|