 |
 |
Hi Doug; I can't answer all the issues with this questions, but I wanted to point out that sometime after SPM2 was released the basis function code was updated so that the HRF and TD were orthogonalized to eachother. This may have affected the sensitivity, although I would have expected the other way. In spm5 they are also orthogonal. Darren > -----Original Message----- > From: SPM (Statistical Parametric Mapping) > [mailto:[log in to unmask]] On Behalf Of Douglas Burman > Sent: Tuesday, January 23, 2007 8:49 AM > To: [log in to unmask] > Subject: Re: [SPM] Using temporal derivative > > Vince, > > I see from the NeuroImage article that you tested the > effectiveness of including the temporal derivative in t-tests > with SPM99. > > In my informal tests, including the derivative in the t-test > did indeed improve sensitivity using SPM99, but not for SPM2. > I checked this after one of the other "major players" in SPM > wrote to the list saying that including the derivative could > actually decrease the responsivity. > > I haven't seen anything posted that would explain this > difference in SPM99 and SPM2, but the difference in my > informal tests on the same dataset was marked -- sensitivity > improved when including the derivative in SPM99, sensitivity > decreased when including it in SPM2. > > > Doug > > > ==============Original message text=============== On Tue, 23 > Jan 2007 2:09:45 am CST Vince Calhoun wrote: > > No, we compute a single image quantifying the amplitude of > the HRF (including both derivative and non-derivative terms) > a nd compute a t-test at the second level. An f-test as you > describe will pick up differences due to either latency, > amplitude or some combination. Our approach is designed to > measure only HRF amplitude differences while removing latency > bias. Sort of the complement to computing a latency image > using, e.g. Rik Henson's approach. > > Regards, > > Vince > > > _____ > > From: SPM (Statistical Parametric Mapping) > [mailto:[log in to unmask]] On Behalf Of Cédric Lemogne > Sent: Tuesday, January 23, 2007 1:03 AM > To: [log in to unmask] > Subject: Re: [SPM] Using temporal derivative > > > > Dear Vince, > > Thank you for the article that I will read carefully but > please allow me one more question. > > If you carry the temporal derivative to the second level, I > presume that you perform a F-contrast including both HRF and > its derivative, right ? > > But what about the contrats at the first level ? For my part, > I aimed to perfo rm t-contrasts separately for HRF and its > derivative at the first level, and then include both in a > F-contrast at the second level. Does it sound correct to you ? > > Best regards, > > Cédric > > > > > > > > Message du 23/01/07 08:51 > > De : "Vince Calhoun" > > A : "'Cédric Lemogne'" , [log in to unmask] Copie à : > > Objet : RE: [SPM] Using temporal derivative > > > > > Dear Cedric, > You might find the following paper helpful which discusses computing a > corrected amplitude estimate from the derivative and non-derivative > parameters for use in a subsequent second-level analysis. In > it we also > shown that, for second level analysis, the inclusion of a temporal > derivative in the model does little to help latency-induced > bias unless you > actually take this parameter to the second level. > > V. D. Calhoun, M. Stevens, G. D. Pearlson, and K. A. Kiehl, > "FMRI analysis > with the general linear model: Re moval of latency-induced > amplitude bias by > incorporation of hemodynamic derivative terms," NeuroImage, > vol. 22, pp. > 252-257, 2004. > > Regards, > > Vince > > > > > _____ > > From: SPM (Statistical Parametric Mapping) > [mailto:[log in to unmask]] On > Behalf Of Cédric Lemogne > > Sent: Tuesday, January 23, 2007 12:36 AM > > To: [log in to unmask] > > Subject: Re: [SPM] Using temporal derivative > > > > > > > Hello Ged, hello Douglas, > > > Yes, I did mean variability in the latency of the HRF's onset > > > Many thanks for your responses ! > > > Best regards, > > > Cédric > > > > > > > > > > > > > Message du 22/01/07 18:04 > > > De : "Douglas Burman" > > > A : [log in to unmask] > > > Copie à : > > > Objet : Re: [SPM] Using temporal derivative > > > > > > Greetings, Ged. > > > > > > In theory, I agree with your statistical approach, and I've seen a > number > > > of messages that indicate this is often done. In > practice, I only use > the > > > HRF-convolved regressor for my analyses. > > > > > > I came to this conclusion after going through a series of > analyses, > using > > > t-tests and F contrasts with and without the temporal > derivative in both > > > > SPM99 and SPM2. (I'm only recently beginning to try SPM5, > so I have > > > nothing useful to say about it.) I'll spare you the > details -- although > I > > > will mention that SPM99 and SPM2 differed in the > independence of the HRF > > > > and TD regressors when using a t-test. (SPM99 showed > improved activation > > > > when including both in a contrast, whereas SPM2 gave substantially > better > > > results when the TD regressor was omitted.) From my empirical > observations > > > on a limited number of analyses, I concluded that using the HRF > regressor > > > only for analyses was preferable, with my tests showing > no improvement > when > > > using the F contrast. > > > > > > I can think of theoretical reasons why this might occur, > although none > that > > > would suggest this must be so. Conceptually, I think of > the temporal > > > derivative as soaking up some variability that shows > little consistent > > > relationship to the effect of interest, since this > variability is not a > > > direct effect of the experimental variable. As such, I > often include the > > > > derivative in the model, but don't use it for analyses. > > > > > > To summarize -- I have no problem with someone's use of > the F contrast > to > > > include both the HRF and TD regressors, but personally I > prefer to limit > my > > > analyses to the HRF. > > > > > > Doug Burman > > > > > > At 03:44 PM 1/2 2/2007 +0000, Ged Ridgway wrote: > > > >Hi Doug, > > > > > > > >>I don't think the temporal derivative has anything to > do with the peak > > > > >>latency -- only with the latency for the onset of the > hemodynamic > response. > > > > > > > >Sorry, "peak latency" was a sloppy quote of mine from > Cedric's original > > > > >phrase "effect of the HRF peak latency variation", I > think we both > meant > > > >variability in the latency of the HRF's peak/onset/curve > rather than > > > >anything to do with peak as in maximum (latency). > > > > > > > >I guess the important question is: do you agree with my stats? > > > > > > > >>>I think that a t-contrast on just > > > >>>the HRF-convolved regressor has a null-hypothesis which still > contains > > > >>>the TD-convolved regressor without the HRF-convolved > one. Since this > > > >>>null wouldn't make physical sense, I think a more > correct thing to > test > > > >>>would be an F-contrast with rows including ones over > both the HRF- > and > > > >>>TD-convolved regressors. > > > > > > > >Cheers, > > > >Ged. > > > > > > Dept. of Communication Sciences & Disorders > > > Northwestern University > > > 2240 Campus Drive > > > Frances Searle Building, Room 2-356 > > > Evanston, IL 60208 > > > phone 847-467-1549 > > > fax 847-491-4975 > > > email: [log in to unmask] > > > > > > > > ===========End of original message text=========== > > > > > > >