 |
 |
Thanks Christian, this was very helpful. ________________________________________ From: Christian F. Beckmann [[log in to unmask]] Sent: Thursday, December 10, 2009 5:48 PM To: [log in to unmask] Subject: Re: discontinuities in ppca eigenspectrum That's due to the way that the adjustment to the EIgenspectrum works (adjusted by the expected Eigenspectrum using a Wishart distribution). The adjustment is in essence controlled by a single quantity which is the ratio of effective voxels (total voxels divided by approximate resel size) and number of time points. You have increased the number of raw voxels by a factor of 3, so the adjustment will be different - see the PICA paper for details hth Christian On 10 Dec 2009, at 22:29, Kundu, Prantik (NIH/NIMH) [F] wrote: > I tried something... I concatenated three copies of the (3-slice) dataset in the Z direction to make a new 9-slice dataset, and melodic_PPCA now has the kinds of curves that I'm used to seeing.... > The unique information is still the same, isn't it? but it seems things are different only because there is a larger amount of raw data....why do you think this is happening? Is it in any way related to the dimensionality reduction step? > > -Prantik > ________________________________________ > From: Christian F. Beckmann [[log in to unmask]] > Sent: Thursday, December 10, 2009 5:15 PM > To: [log in to unmask] > Subject: Re: discontinuities in ppca eigenspectrum > > No such thing, all of them are 'right' as estimates under a given estimator. Melodic by default is designed to over-estimate rather than under-estimate - this, if in doubt, is IMHO the better option: if the dim is underestimated some signals might be lost and never be seen while if the dim is being over-estimated then the worst that can happen is that some signals might get split across multiple components and/or there might also some spurious components - but importantly all signals remain being represented in the estimated signal space. > hth > Christian > > > > > On 10 Dec 2009, at 14:33, Kundu, Prantik (NIH/NIMH) [F] wrote: > >> Well, I'm used to seeing smooth curves... >> The dimensionality estimate using either lap or mean would about 50, even the peak for 3 of the estimates is about 30. What would you say is the right estimate? >> ________________________________________ >> From: Christian F. Beckmann [[log in to unmask]] >> Sent: Thursday, December 10, 2009 9:14 AM >> To: [log in to unmask] >> Subject: Re: discontinuities in ppca eigenspectrum >> >> This looks OK to me... not sure what you mean by discontinuities... could it be that you're worries about the 2nd peak (at about 30)? This is induced by the EIgenspectrum (green curve) which shows a very steep drop early on and an unusual (but smaller) drop at the second peak. Dependent on the form of the penality term this second drop is or is not sufficient to estimate a higher dimensionality... >> hth >> Christian > <melodic_PPCA>