Re: [HCP-Users] Basis Function Selection

[Burgess, Gregory]

Hi Michael,


Again, I would not use a simple double-gamma in the context of a mixed design. It can cause problems with the estimation of the _sustained_ effect. Look at the Petersen paper, section labeled "Limitations", paragraph starting
 "Another situation where BOLD signal can be misapplied is when an assumed shape is used to model the transient events.".


--Greg


Greg Burgess, Ph.D.
Staff Scientist, Human Connectome Project
Washington University School of Medicine
Department of Neuroscience
Phone: 314-362-7864
Email: gburg...@wustl.edu

On Sep 23, 2016, at 8:03 AM, Michael Dreyfuss  wrote:




Thank you,

Do you have any information on how to implement this? The fsl wiki has been down for over a day...


Also, for blocks within a mixed design, do you agree that the basis function is less critical and so something like double gamma is fine? Or what do you recommend for that?


Thank you,
Michael


On Fri, Sep 23, 2016 at 3:39 AM, Stephen Smith 
 wrote:


Hi - in general it's better to use smooth basis functions (eg as used by default in FLOBS) than "square" bases like default FIR bases.  
There's some literature on that but I can't quite remember now who published that - but it seems logical given that you're fitting smooth data - assuming (eg) jittering of events relative to TRs.
Cheers.











On 22 Sep 2016, at 21:44, Michael Dreyfuss  wrote:


OK, thank you. I've read both these papers but it's been a while - do you mind if I ask where specifically you see that this misatribution is problematic for events in mixed designs rather than events in an event only design?


I would be interested in using the FIR function. I'm assuming that using a fixed function is not problematic for modeling the blocks though, right? 


Unfortunately the fsl wiki appears to be down now, so I can't see how to implement FLOBS. for FIR, I'm assuming I can just modify the fsf files before feat_model.

Thank you,
Michael



On Thu, Sep 22, 2016 at 3:35 PM, Burgess, Gregory 
 wrote:

If you’re referring to what is sometimes called a “state-item” design (cf. 
http://www.nil.wustl.edu/labs/schlaggar/Publications_files/MIxedBlockPaper_Final.pdf), you should not use a canonical / assumed response shape. That’s because the variance that is not captured by your assumed HRF can be misattributed to your state
 / sustained regressor.

For these designs, your event-related effects should be modeled with a basis set that will capture varying response shapes (e.g., FIR or FLOBS) to ensure that you do not misattribute poorly-modeled activation to the sustained regressor. I don’t know much about
 the inverse logit basis set, but you might consider looking at it too (Lindquist et al. 2009). An advantage of the FIR basis set is that you can easily look for interactions with “time” to test if the response shape varies between regions or individuals.

Lindquist, M. A., Meng Loh, J., Atlas, L. Y., & Wager, T. D. (2009). Modeling the hemodynamic response function in fMRI: efficiency, bias and mis-modeling. NeuroImage, 45(1 Suppl), S187–98.

http://doi.org/10.1016/j.neuroimage.2008.10.065

--Greg


Greg Burgess, Ph.D.
Staff Scientist, Human Connectome Project
Washington University School of Medicine
Department of Psychiatry
Phone: 314-362-7864
Email: gburg...@wustl.edu



> On Sep 22, 2016, at 2:05 PM, Michael Dreyfuss  wrote:
>
> Thank you both.
>
> This is for our task which is actually a mixed design. I'm not too concerned about the blocks because like you say the main goal is estimating amplitude there. For the jittered events, however, I would want more flexibility in the basis function because like
 you said the HRF could have quite different shapes in different regions and different individuals. Regardless, the activation patterns I'm seeing seem reasonable. I'm just wondering if the double gamma is also better fitted to visual cortex and so activation
 there is more detectable than in other regions, and if so maybe activity in other regions would be better detected using a more flexible basis function like FLOBS of FIR. I think your explanation about proximity to the head coil may be a big part of that too,
 though, so I'm reluctant to assume there is a problem with using double gamma (and there is a cost to estimating the basis function everywhere too).
>
> I will continue to look into these other options...
>
> Thanks again,
> Michael
>
> On Thu, Sep 22, 2016 at 2:31 PM, Burgess, Gregory  wrote:
> Hi Michael,
>
> A few things:
> 1) Matt’s point about the increased activation estimates in visual cortex is a good one. There is increased signal in occipital cortex in functional connectivity analyses that do not assume a response shape. In part, this may 

Re: [HCP-Users] Basis Function Selection

[Michael Dreyfuss]

Thank you,

Do you have any information on how to implement this? The fsl wiki has been
down for over a day...

Also, for blocks within a mixed design, do you agree that the basis
function is less critical and so something like double gamma is fine? Or
what do you recommend for that?

Thank you,
Michael

On Fri, Sep 23, 2016 at 3:39 AM, Stephen Smith  wrote:

> Hi - in general it's better to use smooth basis functions (eg as used by
> default in FLOBS) than "square" bases like default FIR bases.
> There's some literature on that but I can't quite remember now who
> published that - but it seems logical given that you're fitting smooth data
> - assuming (eg) jittering of events relative to TRs.
> Cheers.
>
>
>
>
> On 22 Sep 2016, at 21:44, Michael Dreyfuss  wrote:
>
> OK, thank you. I've read both these papers but it's been a while - do you
> mind if I ask where specifically you see that this misatribution is
> problematic for events in mixed designs rather than events in an event only
> design?
>
> I would be interested in using the FIR function. I'm assuming that using a
> fixed function is not problematic for modeling the blocks though, right?
>
> Unfortunately the fsl wiki appears to be down now, so I can't see how to
> implement FLOBS. for FIR, I'm assuming I can just modify the fsf files
> before feat_model.
>
> Thank you,
> Michael
>
> On Thu, Sep 22, 2016 at 3:35 PM, Burgess, Gregory 
> wrote:
>
>> If you’re referring to what is sometimes called a “state-item” design
>> (cf. http://www.nil.wustl.edu/labs/schlaggar/Publications_files/M
>> IxedBlockPaper_Final.pdf), you should not use a canonical / assumed
>> response shape. That’s because the variance that is not captured by your
>> assumed HRF can be misattributed to your state / sustained regressor.
>>
>> For these designs, your event-related effects should be modeled with a
>> basis set that will capture varying response shapes (e.g., FIR or FLOBS) to
>> ensure that you do not misattribute poorly-modeled activation to the
>> sustained regressor. I don’t know much about the inverse logit basis set,
>> but you might consider looking at it too (Lindquist et al. 2009). An
>> advantage of the FIR basis set is that you can easily look for interactions
>> with “time” to test if the response shape varies between regions or
>> individuals.
>>
>> Lindquist, M. A., Meng Loh, J., Atlas, L. Y., & Wager, T. D. (2009).
>> Modeling the hemodynamic response function in fMRI: efficiency, bias and
>> mis-modeling. NeuroImage, 45(1 Suppl), S187–98.
>> http://doi.org/10.1016/j.neuroimage.2008.10.065
>>
>> --Greg
>>
>> 
>> Greg Burgess, Ph.D.
>> Staff Scientist, Human Connectome Project
>> Washington University School of Medicine
>> Department of Psychiatry
>> Phone: 314-362-7864
>> Email: gburg...@wustl.edu
>>
>> > On Sep 22, 2016, at 2:05 PM, Michael Dreyfuss 
>> wrote:
>> >
>> > Thank you both.
>> >
>> > This is for our task which is actually a mixed design. I'm not too
>> concerned about the blocks because like you say the main goal is estimating
>> amplitude there. For the jittered events, however, I would want more
>> flexibility in the basis function because like you said the HRF could have
>> quite different shapes in different regions and different individuals.
>> Regardless, the activation patterns I'm seeing seem reasonable. I'm just
>> wondering if the double gamma is also better fitted to visual cortex and so
>> activation there is more detectable than in other regions, and if so maybe
>> activity in other regions would be better detected using a more flexible
>> basis function like FLOBS of FIR. I think your explanation about proximity
>> to the head coil may be a big part of that too, though, so I'm reluctant to
>> assume there is a problem with using double gamma (and there is a cost to
>> estimating the basis function everywhere too).
>> >
>> > I will continue to look into these other options...
>> >
>> > Thanks again,
>> > Michael
>> >
>> > On Thu, Sep 22, 2016 at 2:31 PM, Burgess, Gregory <
>> burge...@psychiatry.wustl.edu> wrote:
>> > Hi Michael,
>> >
>> > A few things:
>> > 1) Matt’s point about the increased activation estimates in visual
>> cortex is a good one. There is increased signal in occipital cortex in
>> functional connectivity analyses that do not assume a response shape. In
>> part, this may result from the back of the head being closer to the head
>> coil than other brain regions (because participants are laying down).
>> > 2) To the best of my knowledge, the HCP consortium has not ventured to
>> recommend a single, ideal HRF for use in task fMRI analysis. In fact, I’d
>> wager that most people in the consortium expect the hemodynamic response to
>> vary across brain regions and across people in such a way that there is no
>> single ideal canonical HRF.
>> > 3) We chose the 

Re: [HCP-Users] Basis Function Selection

[Stephen Smith]

Hi - in general it's better to use smooth basis functions (eg as used by 
default in FLOBS) than "square" bases like default FIR bases.  
There's some literature on that but I can't quite remember now who published 
that - but it seems logical given that you're fitting smooth data - assuming 
(eg) jittering of events relative to TRs.
Cheers.




> On 22 Sep 2016, at 21:44, Michael Dreyfuss  wrote:
> 
> OK, thank you. I've read both these papers but it's been a while - do you 
> mind if I ask where specifically you see that this misatribution is 
> problematic for events in mixed designs rather than events in an event only 
> design?
> 
> I would be interested in using the FIR function. I'm assuming that using a 
> fixed function is not problematic for modeling the blocks though, right? 
> 
> Unfortunately the fsl wiki appears to be down now, so I can't see how to 
> implement FLOBS. for FIR, I'm assuming I can just modify the fsf files before 
> feat_model.
> 
> Thank you,
> Michael
> 
> On Thu, Sep 22, 2016 at 3:35 PM, Burgess, Gregory  > wrote:
> If you’re referring to what is sometimes called a “state-item” design (cf. 
> http://www.nil.wustl.edu/labs/schlaggar/Publications_files/MIxedBlockPaper_Final.pdf
>  
> ),
>  you should not use a canonical / assumed response shape. That’s because the 
> variance that is not captured by your assumed HRF can be misattributed to 
> your state / sustained regressor.
> 
> For these designs, your event-related effects should be modeled with a basis 
> set that will capture varying response shapes (e.g., FIR or FLOBS) to ensure 
> that you do not misattribute poorly-modeled activation to the sustained 
> regressor. I don’t know much about the inverse logit basis set, but you might 
> consider looking at it too (Lindquist et al. 2009). An advantage of the FIR 
> basis set is that you can easily look for interactions with “time” to test if 
> the response shape varies between regions or individuals.
> 
> Lindquist, M. A., Meng Loh, J., Atlas, L. Y., & Wager, T. D. (2009). Modeling 
> the hemodynamic response function in fMRI: efficiency, bias and mis-modeling. 
> NeuroImage, 45(1 Suppl), S187–98. 
> http://doi.org/10.1016/j.neuroimage.2008.10.065 
> 
> 
> --Greg
> 
> 
> Greg Burgess, Ph.D.
> Staff Scientist, Human Connectome Project
> Washington University School of Medicine
> Department of Psychiatry
> Phone: 314-362-7864 
> Email: gburg...@wustl.edu 
> 
> > On Sep 22, 2016, at 2:05 PM, Michael Dreyfuss  > > wrote:
> >
> > Thank you both.
> >
> > This is for our task which is actually a mixed design. I'm not too 
> > concerned about the blocks because like you say the main goal is estimating 
> > amplitude there. For the jittered events, however, I would want more 
> > flexibility in the basis function because like you said the HRF could have 
> > quite different shapes in different regions and different individuals. 
> > Regardless, the activation patterns I'm seeing seem reasonable. I'm just 
> > wondering if the double gamma is also better fitted to visual cortex and so 
> > activation there is more detectable than in other regions, and if so maybe 
> > activity in other regions would be better detected using a more flexible 
> > basis function like FLOBS of FIR. I think your explanation about proximity 
> > to the head coil may be a big part of that too, though, so I'm reluctant to 
> > assume there is a problem with using double gamma (and there is a cost to 
> > estimating the basis function everywhere too).
> >
> > I will continue to look into these other options...
> >
> > Thanks again,
> > Michael
> >
> > On Thu, Sep 22, 2016 at 2:31 PM, Burgess, Gregory 
> > > 
> > wrote:
> > Hi Michael,
> >
> > A few things:
> > 1) Matt’s point about the increased activation estimates in visual cortex 
> > is a good one. There is increased signal in occipital cortex in functional 
> > connectivity analyses that do not assume a response shape. In part, this 
> > may result from the back of the head being closer to the head coil than 
> > other brain regions (because participants are laying down).
> > 2) To the best of my knowledge, the HCP consortium has not ventured to 
> > recommend a single, ideal HRF for use in task fMRI analysis. In fact, I’d 
> > wager that most people in the consortium expect the hemodynamic response to 
> > vary across brain regions and across people in such a way that there is no 
> > single ideal canonical HRF.
> > 3) We chose the double-gamma during very early analysis of HCP pilot data. 
> > Using 2.5s TR data, the default 

Re: [HCP-Users] Basis Function Selection

[Michael Dreyfuss]

Thank you both.

This is for our task which is actually a mixed design. I'm not too
concerned about the blocks because like you say the main goal is estimating
amplitude there. For the jittered events, however, I would want more
flexibility in the basis function because like you said the HRF could have
quite different shapes in different regions and different individuals.
Regardless, the activation patterns I'm seeing seem reasonable. I'm just
wondering if the double gamma is also better fitted to visual cortex and so
activation there is more detectable than in other regions, and if so maybe
activity in other regions would be better detected using a more flexible
basis function like FLOBS of FIR. I think your explanation about proximity
to the head coil may be a big part of that too, though, so I'm reluctant to
assume there is a problem with using double gamma (and there is a cost to
estimating the basis function everywhere too).

I will continue to look into these other options...

Thanks again,
Michael

On Thu, Sep 22, 2016 at 2:31 PM, Burgess, Gregory <
burge...@psychiatry.wustl.edu> wrote:

> Hi Michael,
>
> A few things:
> 1) Matt’s point about the increased activation estimates in visual cortex
> is a good one. There is increased signal in occipital cortex in functional
> connectivity analyses that do not assume a response shape. In part, this
> may result from the back of the head being closer to the head coil than
> other brain regions (because participants are laying down).
> 2) To the best of my knowledge, the HCP consortium has not ventured to
> recommend a single, ideal HRF for use in task fMRI analysis. In fact, I’d
> wager that most people in the consortium expect the hemodynamic response to
> vary across brain regions and across people in such a way that there is no
> single ideal canonical HRF.
> 3) We chose the double-gamma during very early analysis of HCP pilot data.
> Using 2.5s TR data, the default double-gamma showed zstat maps with
> slightly higher statistical significance at the group-level than the
> default gamma HRF (in Feat). The double-gamma also seemed to be used more
> widely in the literature, in part due to the commonly observed undershoot
> at the end of the hemodynamic response (see Glover, 1999). We made this
> choice in piloting, and stayed with it for analysis of the Phase II HCP. We
> did not re-evaluate HRFs in the fast TR HCP data.
> 4) In HCP tfMRI, we utilized blocked designs. Blocked designs are good for
> detecting the response, but are not good for estimating the shape of the
> response function. It may follow that differences between canonical HRFs
> will matter less for blocked designs, but I don’t know if anyone has looked
> at that systematically.
> 5) If you’re referring to analysis of your own data using an event-related
> design, your best bet will likely be using a basis set. FSL has FLOBS,
> folks at Wash U tend to use FIR basis sets, but there are others out there
> as well. There are quite a few papers out there to help you choose between
> those basis sets. However, I’m not sure it would make much sense in the
> context of a blocked design.
>
> Hope this all helps!
> --Greg
>
> 
> Greg Burgess, Ph.D.
> Staff Scientist, Human Connectome Project
> Washington University School of Medicine
> Department of Psychiatry
> Phone: 314-362-7864
> Email: gburg...@wustl.edu
>
> > On Sep 22, 2016, at 12:25 PM, Glasser, Matthew 
> wrote:
> >
> > BOLD fluctuations are generally stronger on the occipital cortex
> > (independent of the chosen HRF).  See for example the attached functional
> > CNR map (BOLDVariance / UnstructuredNoiseVariance).
> >
> > Peace,
> >
> > Matt.
> >
> > On 9/21/16, 7:29 PM, "hcp-users-boun...@humanconnectome.org on behalf of
> > Michael Dreyfuss"  > mdreyfus...@gmail.com> wrote:
> >
> >> Hello,
> >>
> >> What kind of basis function are you recommending for tfMRI data?I have
> >> been using double-gamma HRF but I notice that the signal is always
> >> strongest in occipital cortex, so I was wondering if this is not optimal
> >> for other regions. If so, do you have a more customized recommendation
> >> that would better fit HRF functions in other parts of the brain to
> detect
> >> signal there?
> >>
> >> Thank you,
> >> Michael
> >> ___
> >> HCP-Users mailing list
> >> HCP-Users@humanconnectome.org
> >> http://lists.humanconnectome.org/mailman/listinfo/hcp-users
> >
> >
> > 
> > The materials in this message are private and may contain Protected
> Healthcare Information or other information of a sensitive nature. If you
> are not the intended recipient, be advised that any unauthorized use,
> disclosure, copying or the taking of any action in reliance on the contents
> of this information is strictly prohibited. If you have 

Re: [HCP-Users] Basis Function Selection

[Burgess, Gregory]

Hi Michael,

A few things:
1) Matt’s point about the increased activation estimates in visual cortex is a 
good one. There is increased signal in occipital cortex in functional 
connectivity analyses that do not assume a response shape. In part, this may 
result from the back of the head being closer to the head coil than other brain 
regions (because participants are laying down).
2) To the best of my knowledge, the HCP consortium has not ventured to 
recommend a single, ideal HRF for use in task fMRI analysis. In fact, I’d wager 
that most people in the consortium expect the hemodynamic response to vary 
across brain regions and across people in such a way that there is no single 
ideal canonical HRF.
3) We chose the double-gamma during very early analysis of HCP pilot data. 
Using 2.5s TR data, the default double-gamma showed zstat maps with slightly 
higher statistical significance at the group-level than the default gamma HRF 
(in Feat). The double-gamma also seemed to be used more widely in the 
literature, in part due to the commonly observed undershoot at the end of the 
hemodynamic response (see Glover, 1999). We made this choice in piloting, and 
stayed with it for analysis of the Phase II HCP. We did not re-evaluate HRFs in 
the fast TR HCP data.
4) In HCP tfMRI, we utilized blocked designs. Blocked designs are good for 
detecting the response, but are not good for estimating the shape of the 
response function. It may follow that differences between canonical HRFs will 
matter less for blocked designs, but I don’t know if anyone has looked at that 
systematically.
5) If you’re referring to analysis of your own data using an event-related 
design, your best bet will likely be using a basis set. FSL has FLOBS, folks at 
Wash U tend to use FIR basis sets, but there are others out there as well. 
There are quite a few papers out there to help you choose between those basis 
sets. However, I’m not sure it would make much sense in the context of a 
blocked design.

Hope this all helps!
--Greg


Greg Burgess, Ph.D.
Staff Scientist, Human Connectome Project
Washington University School of Medicine
Department of Psychiatry
Phone: 314-362-7864
Email: gburg...@wustl.edu

> On Sep 22, 2016, at 12:25 PM, Glasser, Matthew  wrote:
>
> BOLD fluctuations are generally stronger on the occipital cortex
> (independent of the chosen HRF).  See for example the attached functional
> CNR map (BOLDVariance / UnstructuredNoiseVariance).
>
> Peace,
>
> Matt.
>
> On 9/21/16, 7:29 PM, "hcp-users-boun...@humanconnectome.org on behalf of
> Michael Dreyfuss"  mdreyfus...@gmail.com> wrote:
>
>> Hello,
>>
>> What kind of basis function are you recommending for tfMRI data?I have
>> been using double-gamma HRF but I notice that the signal is always
>> strongest in occipital cortex, so I was wondering if this is not optimal
>> for other regions. If so, do you have a more customized recommendation
>> that would better fit HRF functions in other parts of the brain to detect
>> signal there?
>>
>> Thank you,
>> Michael
>> ___
>> HCP-Users mailing list
>> HCP-Users@humanconnectome.org
>> http://lists.humanconnectome.org/mailman/listinfo/hcp-users
>
>
> 
> The materials in this message are private and may contain Protected 
> Healthcare Information or other information of a sensitive nature. If you are 
> not the intended recipient, be advised that any unauthorized use, disclosure, 
> copying or the taking of any action in reliance on the contents of this 
> information is strictly prohibited. If you have received this email in error, 
> please immediately notify the sender via telephone or return mail.
>
> ___
> HCP-Users mailing list
> HCP-Users@humanconnectome.org
> http://lists.humanconnectome.org/mailman/listinfo/hcp-users
> 



The materials in this message are private and may contain Protected Healthcare 
Information or other information of a sensitive nature. If you are not the 
intended recipient, be advised that any unauthorized use, disclosure, copying 
or the taking of any action in reliance on the contents of this information is 
strictly prohibited. If you have received this email in error, please 
immediately notify the sender via telephone or return mail.

___
HCP-Users mailing list
HCP-Users@humanconnectome.org
http://lists.humanconnectome.org/mailman/listinfo/hcp-users

[HCP-Users] Basis Function Selection

[Michael Dreyfuss]

Hello, 

What kind of basis function are you recommending for tfMRI data?I have been 
using double-gamma HRF but I notice that the signal is always strongest in 
occipital cortex, so I was wondering if this is not optimal for other regions. 
If so, do you have a more customized recommendation that would better fit HRF 
functions in other parts of the brain to detect signal there?

Thank you,
Michael
___
HCP-Users mailing list
HCP-Users@humanconnectome.org
http://lists.humanconnectome.org/mailman/listinfo/hcp-users