Dear Branka
Matthijs Vink has treated an SPM toolbox that might help you. Design
Magic is a program to calculate multicollinearity within an fMRI
design and to create and evaluate a high-pass filter:
http://www.matthijs-vink.com/tools.html
Regards - MFG
>-----Original Message-----
>From: SPM (Statistical Parametric Mapping)
>[mailto:[log in to unmask]] On Behalf Of
>Branka Milivojevic
>Sent: Wednesday, March 22, 2006 3:55 PM
>To: [log in to unmask]
>Subject: Re: [SPM] high pass filter cut-off
>
>Thanks to Eric and Torben for the replies.
>
>I'll take this issue into account next time I
>design an experiment.
>I've already ran the analysis with 1280 s HP
>filter, and as Torben
>pointed out, I loose significance.
>
>I've had another look at the frequency
>histograms in the design matrix,
>and I think 256 would be a better cut off
>than 162, but I will try both.
>
>I am primarily interested in population-level
>statistics, and Eric has
>mentioned that I should choose a filter to
>optimise the variance of the
>estimator of the effect and that there should
>be an assumption about
>temporal autocorrelation and that the
>optimal filter frequency can be
>chosen with software. Is that SPM or
>something else? How would I go
>about choosing an optimal filter this way.
>
>I already correct for serial autocorrelations in
>my analysis.
>
>Thanks again ;-)
>Branka
>
>
>On 22 Mar 2006 at 10:21, Eric Zarahn wrote:
>
>> Branka,
>>
>> One might say there are two
>considerations regarding the
>importance
>> of temporal filtering in fMRI. One is validity
>of within-subject
>> (i.e., time series or first-level) statistics.
>The second is
>> efficiency at the between-subject (i.e.,
>population-level,
>> second-level,or random effects) statistics.
>From this perspective,
>> if one is not performing statistical
>inference at the first-level,
>> then all that will be affected by the choice
>of temporal filter
>> will be sensitivity for effects at the second-
>level. Furthermore,
>> the optimal choice of filter for these two
>considerations might not
>> coincide.
>>
>> If population-level statistics are most
>important, than the filter
>> should be chosen to optimize the variance
>of the estimator of the
>> effect with respect to filter parameters
>(which requires assuming
>> some form for the temporal
>autocorrelation). This can be done with
>> software.
>>
>> Having said that, if first-level inference is
>important, and you do
>> not plan on using software to compute
>optimal filters, then from
>> your design and my fuzzy logical intuition
>(which relies on an
>> internal, prior model of fMRI noise spectra,
>which I would not bet
>> the house on), I'd say that 162 (27*6)
>seconds might be a more
>> reasonable cut-off than 1280 seconds.
>>
>> Eric
>>
>> uoting Branka Milivojevic
><[log in to unmask]>:
>>
>> > Hi all,
>> > I realise that a number of threads on
>this topic are available in
>> > the mailing list archives,
>> > but I am finding those difficult to
>understand.
>> >
>> > I have 6 conditions, each repeated twice
>during a run and the
>> > experimental conditions
>> > alternate with a fixation only condition.
>I've collected 4 runs
>> > of data, alternating the order
>> > of the conditions. There are only 2
>orders: These are as follows:
>> >
>> > Order1:
>> > fix 9 seconds
>> > condition1 18 seconds
>> > fix 9 seconds
>> > condition2 18 seconds
>> > fix 9 seconds
>> > condition3 18 seconds
>> > fix 9 seconds
>> > condition4 18 seconds
>> > fix 9 seconds
>> > condition5 18 seconds
>> > fix 9 seconds
>> > condition6 18 seconds
>> > fix 9 seconds
>> > condition6 18 seconds
>> > fix 9 seconds
>> > condition5 18 seconds
>> > fix 9 seconds
>> > condition4 18 seconds
>> > fix 9 seconds
>> > condition3 18 seconds
>> > fix 9 seconds
>> > condition2 18 seconds
>> > fix 9 seconds
>> > condition1 18 seconds
>> > fix 6 seconds
>> >
>> > Order2
>> > fix 9 seconds
>> > condition6 18 seconds
>> > fix 9 seconds
>> > condition5 18 seconds
>> > fix 9 seconds
>> > condition4 18 seconds
>> > fix 9 seconds
>> > condition3 18 seconds
>> > fix 9 seconds
>> > condition2 18 seconds
>> > fix 9 seconds
>> > condition1 18 seconds
>> > fix 9 seconds
>> > condition1 18 seconds
>> > fix 9 seconds
>> > condition2 18 seconds
>> > fix 9 seconds
>> > condition3 18 seconds
>> > fix 9 seconds
>> > condition4 18 seconds
>> > fix 9 seconds
>> > condition5 18 seconds
>> > fix 9 seconds
>> > condition6 18 seconds
>> > fix 6 seconds
>> >
>> > The TR is 3 seconds, and I am analysing
>the runs as a single
>> > session. I've looked at the
>> > frequency graphs in the design matrix,
>and I estimated the
>> > cut-off frequency to be 1280
>> > seconds. This seems quite long. Is this
>correct? Should I use
>> > another value or perhaps
>> > no high pass filter.
>> >
>> > I would appreciate any help on this
>matter.
>> >
>> > Cheers,
>> > Branka
>> >
>> >
>> > Branka Milivojevic
>> > PhD candidate
>> > Department of Psychology
>> > University of Auckland
>> > Private Bag 92019
>> > Auckland
>> > New Zealand
>> > Ph: (+649) 373 7599 ext. 88636
>> > Fax: (+649) 3737450
>> >
>
>
>Branka Milivojevic
>PhD candidate
>Department of Psychology
>University of Auckland
>Private Bag 92019
>Auckland
>New Zealand
>Ph: (+649) 373 7599 ext. 88636
>Fax: (+649) 3737450
|