Dear all,
I've spent some time for adjusting and testing my data according to your
suggestions (thanks to Peter and Brian!), in order to try to improve the
explained variance of my models (spm_dcm_fmri_check).
I used SPM12 for all processing steps.
The suggestions were as follows:
a) Concatenating my multiple sessions data (sessions 1-6) into one large
session. In contrast to my single session data of session 1 ("SS"), I
will refer to this as concatenated data of sessions 1-6 ("CS").
b) Switching from states per region: "one" to states per region: "two"
in DCM-specification.
c) Including more events as driving input in my event-related design,
i.e. adding more pertubation to my system. As I had already included all
possible events as driving input in my original SS-data, I could not
test for this.
So I have four combinations: "SS/one", "SS/two", "CS/one", "CS/two".
For one of my subjects I have estimated 15 DCMs under each of these
combinations (= 60 DCMs) to compare their mean percentage of explained
variance ("MPEV").
These are my results:
- SS/one: MPEV = 1.13 %
- SS/two: MPEV = 0.47 %
- CS/one: MPEV = 1.00 %
- CS/two: MPEV = 0.40 %
Interpretation is quite simple:
a) A slight drop of MPEV when concatenating into one large session
(-0.13% and -0.07%).
b) A even bigger drop of MPEV when switching states per region to two
(-0.66% and -0.60%).
Unfortunately both suggestions did not help improving the explained
variance of my models. Maybe this is due to the already very low MEV my
DCMs, I think that in other cases these suggestions are working very
well. I hope that other people from the SPM community may benefit from
this. Please feel free to comment!
----------------------
Just to make sure that I've made no mistake (from a technical point of
view), I will give a brief report of what I've done, when concatenating
my sessions and specifying my DCMs. Note that I have about 30-45 s of
additional baseline scanning after the last event at the end of my
sessions. This should protect against at least some of the usual
problems with concatenating multiple sessions.
I have sessions 1-6 with 225 scans each and my TR is 3 s. So I have 1350
scans and the duration of each session is 675 s. This is without dummy
scans.
Preprocessing:
1. I have realigned and resliced all 1350 scans to the first slice. I
used the default values of realignment and reslicing. I used my already
preprocessed data (swua4D.nii) of session 1-6, which I've used
beforehand for my SS-data. I know that this is the "quick and dirty"
way, but I did not want to go through all the preprocessing steps again.
Specification of my GLM:
2. I have included the resulting files from a), keeping all other
parameters unchanged
- Scans: rswua4D.nii (1350)
- Multiple regressors: rp_rswua4D.txt
3. I have added 5 new regressors indicating sessions 1-5.
4. I have included my new onsets and durations of my factors for all
sessions. The new onsets for session 2-6 were calculated by adding 675 s
to the values of the 2nd session, adding 1350 s to the values of the 3rd
session and so on.
Contrasts:
4. I kept my design-matrix and contrasts unchanged and I was using the
same contrasts as in SS. I did not add zeros for the additional 5
regressors (see 3.).
Extraction of VOIs:
5. I used the same procedure as before (see older posts).
Specification of my DCMs:
6. As mentioned before, I set up simple DCMs with only two regions
connected by reciprocal intrinsic connections
7. My factors had different numbers of onsets (~500, ~200, ~100) and
time of duration (0.1s up to 9s).
8. I only specified DCMs with at least one driving input to my models in
order to prevent from the "stochastic" DCM calculation routine (see
older posts).
9. I chose the default of VOI-timings [s] of "1.5 1.5" [in SPM8 I used
the default of "3 3"]. Besides that I used the same parameters as before.
10. For the 15 DCMs under each condition I systematically varied regions
(VOIs) and my factors for driving and modulatory input in the same fashion.
Best,
Eric
---
Hi Eric,
in our case we had one session only but with 3 trial types- one of these
was more frequent (ie, around 100 trials), and the other two less
frequent (around 30).
We were interested in how a parametric modulator on the frequent trial
type modulated the connectivity, so initially we used the trial onsets
as direct input and its parametric modulator as modulatory input. But it
helped on the fitting also to include those 2 other trial types as
direct inputs, so the system was more frequently perturbed. So you could
try to increase the number of different trial types as inputs.
Best, Brian
---
Dear Brian,
Thank you for your helpful insight on how you came across the same
problem I am currently struggling with and how you succeeded in
improving your data.
Just one quick question. You say that "increasing the number of
conditions used as direct input" in your case helped to increase the %
of explained variance. Are you referring to Peter's suggestion of
concatenating my data (i.e. increasing the total number of "trials" in
my input factor) or do you suggest to increase the total number of (same
or different) direct input factors to my models?
Your second idea sounds interesting too, I will test it tomorrow and
give feedback about the results.
Best,
Eric
> Hi Eric,
>
> the dynamic expectation maximization is used for inversion of
> stochastic DCM models- I think that when there's no direct input, then
> the spm_dcm_estimate script will use this inversion scheme. This leads
> to much better fits, but it's using noise rather than the paradigm to
> explain a lot of the data. I guess your B-values are very small? That
> also explains the drop in variance explained when you define direct
> inputs, as the estimate-script will use the deterministic scheme instead.
>
> We previously also experienced problems with flat-line fits using
> deterministic DCM on an event-related paradigm - Peter came with some
> helpfull suggestions and 1) increasing the number of conditions used
> as direct input (even though our question was related to the
> modulatory influence of one of the conditions) and 2) using the
> 2-state version of DCM actually did a difference- the average %
> variance explained increased from ca 5 % to 17%.
>
> So these two things may be worthwhile to try?
>
> Best, Brian
>
>
>
>
>
> On 2014-12-04 18:22, Eric Holst wrote:
>> Dear Peter, dear SPM-experts,
>>
>> Leaving the hypotheses of my work aside, I have spent some time
>> testing my VOIs under different conditions by reducing my models to
>> two regions and systematically varying the input of all my factors to
>> regions and/or connections. Let's say my two regions (VOIs) are left
>> and right PPC as mentioned before (I have tested different VOIs as
>> well) with reciprocal intrinsic connections. This is what I've found
>> out:
>>
>> 1.) Whenever I *only* define one ore more factors as modulatory input
>> to one or more connections (without defining the same or a different
>> factor as driving input to any of both regions) the explained variance
>> of my model is ranging from 38-40%. This is independent of the factors
>> or the contrasts I use for extracting my VOIs. The model takes way
>> longer to estimate using a different algorhythm ("Dynamic Expectation
>> Maximisation"). To my knowledge this is not useful, as I have not
>> defined a driving input as pertubation to my system.
>> In other words: it makes no sense setting up a DCM without driving
>> input. But on the other hand I have a high percentage of explained
>> variance.
>>
>> 2.) This high percentage of explained variance drastically drops to
>> 0-3% (mostly 0%) soon as I *additionally* define any factor as driving
>> input to one or more regions. This is again independent of the factors
>> and contrasts I use. But as far as I know this is the right method for
>> setting up DCMs.
>>
>> Maybe you or someone else can explain this finding? I am a little bit
>> confused about it.
>>
>> I am not sure that concatenating my sessions will make a big
>> difference, but I am willing to give it a try. A script for
>> concatenating sessions would be great. Thank you very much Peter for
>> your help!
>>
>> Best,
>> Eric
>>
>>> Dear Eric,
>>> Regarding your DCMs flat-lining. First, you are correct that
>>> extracting the ROIs with SPM12 won't make a difference. And I'll
>>> assume that your contrasts and effects of interest contrast are
>>> correctly specified - always worth checking. (You get a good level
>>> of explained variance in your ROIs, which is a good sign.)
>>>
>>> Model fitting is more challenging with fast event-related designs,
>>> and you do not have many trials per condition. I think concatenating
>>> your sessions could be very important. I appreciate the technical
>>> challenge in doing this - I am planning a feature for SPM to do this
>>> automatically, but it won't be ready for a while. In the meantime, I
>>> think Donald McLaren (CC'd) may have a concatenation script that may
>>> help. Otherwise, you'll need to do this manually with your onsets.
>>> As you say, you'll need to be very careful with dummy scans etc.
>>>
>>> Good luck,
>>> Peter.
>>>
>>> -----Original Message-----
>>> From: SPM (Statistical Parametric Mapping)
>>> [mailto:[log in to unmask]] On Behalf Of Eric Holst
>>> Sent: 03 December 2014 13:54
>>> To: [log in to unmask]
>>> Subject: Re: [SPM] DCM: Problem with percentage of explained variance
>>>
>>> Dear Peter,
>>>
>>> Thank you very much for helping me again.
>>>
>>> ad 1. I have already estimated the same DCM analysis in SPM12,
>>> unfortunately yielding the same results (explained variance of 0% in
>>> spm_dcm_fmri_check). But I have not tried to extract my VOIs with
>>> SPM12 for the DCM. I am not sure if this has any impact on the priors.
>>>
>>> ad 2. Your assumption is right, I have modeled separate DCMs per
>>> session. Reading the older posts about concatenating sessions
>>> worries me a little bit, as I am not good at MATLAB coding and there
>>> are many pitfalls for calculating my concatenated onsets (dummy
>>> scans at the beginning of each scan, etc.). This is why I am quite
>>> sure, I won't be able to do it manually. Is there an easy way to do
>>> it (maybe using gPPI)? I am very willing to give it a try, if there
>>> is an 'easy' way to do it.
>>>
>>> You are presuming that the problem may result from a small number of
>>> trials. Each of my sessions consists of 40 trials in randomized order
>>> (20 trials for condition 1 and 20 trials for condition 2). Wrong or
>>> missing answers were not further analysed. My subjects had an
>>> average percentage of right answers of about 70-90%, so for each
>>> condition I have 13+ trials. I am not sure, if this is sufficient.
>>>
>>> Please see attached file "DCM_inputs" as a DCM example for one
>>> session of one subject (condition 1). The first input "photic" is a
>>> driving input to bilateral V1 regions, the second input "form" is a
>>> modulatory input to bilateral V1->V4 connections. I am varying the
>>> third input "att_1HEM" as modulatory input to regions/connections in
>>> my models (representing attentional modulation under condition 1).
>>> Note that there is no peak in "att_1HEM" for trials under condition
>>> 2. For these conditions I have a second modulatory input "att_2HEM"
>>> with different models, whereas the other inputs "photic" and "form"
>>> are kept unchanged.
>>> The aim of my study is to find out if the best model under BMS is
>>> same or different for condition 1 vs. condition 2. The duration of
>>> input "form" is only 100 ms, the other durations are always 6+
>>> seconds. I thought that the problem could arise from the short
>>> duration of "form"
>>> in comparison to the other inputs, but when I estimate models
>>> without "form" the explained variance goes up to 1% in some (rare)
>>> cases - which is still not good enough.
>>>
>>> Is there anything more I could try or test? Maybe there is a problem
>>> with my contrasts?
>>>
>>> Best,
>>> Eric
>>>
>>>
>>>
>>>> Hi Eric,
>>>> Your procedure for picking ROIs sounds perfect. Two things come to
>>>> mind:
>>>>
>>>> 1. You mention in your original email that you are using SPM8.
>>>> Could you try exactly the same DCM analysis in SPM12? The priors
>>>> were changed to reduce the chance of flat-lining.
>>>>
>>>> 2. You have 6 sessions, and it doesn't look like you've
>>>> concatenated them into a single session, thus I assume you're
>>>> creating separate models per session? This means in each model, you
>>>> may not have many trials? Try SPM12 first, otherwise I suggest
>>>> concatenating the sessions into one long session, with extra
>>>> regressors to model session effects. See many previous posts on how
>>>> to do this.
>>>>
>>>> Best,
>>>> Peter.
>>>>
>>>> -----Original Message-----
>>>> From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
>>>> On Behalf Of Eric Holst
>>>> Sent: 03 December 2014 00:25
>>>> To: [log in to unmask]
>>>> Subject: Re: [SPM] DCM: Problem with percentage of explained variance
>>>>
>>>> Dear Peter, dear SPM-experts,
>>>>
>>>> This may serve as an example to demonstrate how I have defined and
>>>> extracted my VOIs.
>>>>
>>>> First, I have created masks for my ROIs with the SPM Anatomy
>>>> Toolbox (version 18) in MNI-space.
>>>> In my example, this is for posterior parietal cortex of the left
>>>> hemisphere (PPC_L), consisting of Brodman area 5+7.
>>>>
>>>> As mentioned before, I have calculated one SPM per subject
>>>> containing sessions 1-6, so for each subject...
>>>> - SPM => Results
>>>> - t-contrast for "attention"
>>>> - masking (inclusive) => image: ROI_PPC_L (as created with SPM Anatomy
>>>> Toolbox)
>>>> - p value adjustment: none
>>>> - threshold: p = 0.05
>>>> - extended threshold voxels: 0
>>>> - goto: global maximum (which is inside my predefined mask)
>>>> - eigenvariate
>>>> - adjust for: effects of interest (EOI)
>>>> - session: 1-6
>>>> - sphere radius: 6 mm
>>>>
>>>> In this way I have extracted my VOIs for all subjects and sessions
>>>> for PPC_L. For PPC of the right hemisphere and the other regions
>>>> (V1 and V4) I have used different masks and/or contrasts, resulting
>>>> in 36 VOIs per subject (6 regions x 6 sessions). By definition
>>>> these VOIs were based on activations in my SPM analysis, subjects
>>>> without activation in any of these ROIs were not further analysed.
>>>>
>>>> Based on these VOIs I have set up my DCMs. The basic setup of each DCM
>>>> is similar to the one used for the „attention to motion“ paper,
>>>> including 6 regions with reciprocal intrahemispheric connections
>>>> between
>>>> V1-V4 and V4-PPC and reciprocal interhemispheric connections between
>>>> V4-V4 and PPC-PPC. Driving input is allocated to both V1 regions.
>>>> On this basis I have set up multiple DCMs with alternating
>>>> modulatory input of factor "attention" to either regions or
>>>> connections to subsequently perform a BMS analysis.
>>>>
>>>>
>>>> Peter, I hope I have answered your questions. If you have more
>>>> questions, please let me know. If the extraction of my VOIs is
>>>> correct, where else could I dig for my mistake?
>>>>
>>>>
>>>> Thank you for your time!
>>>>
>>>> Best,
>>>> Eric
>>>>
>>>>
>>>>
>>>>> Hi Eric,
>>>>> Sorry you're still having problems with your models. Simplifying
>>>>> the models to 3 regions is a good start. Could you tell us more
>>>>> about how you choose your ROIs? Are they based on activations in
>>>>> your SPM analysis, in the same experimental conditions as you're
>>>>> modelling with the DCM?
>>>>>
>>>>> Best,
>>>>> Peter
>>>>>
>>>>> -----Original Message-----
>>>>> From: SPM (Statistical Parametric Mapping)
>>>>> [mailto:[log in to unmask]] On Behalf Of Eric Holst
>>>>> Sent: 01 December 2014 13:55
>>>>> To: [log in to unmask]
>>>>> Subject: [SPM] DCM: Problem with percentage of explained variance
>>>>>
>>>>> Dear Peter, dear SPM-experts,
>>>>>
>>>>> some time ago I posted a question about very low percentage of
>>>>> explained variance (0%) when checking my DCMs with
>>>>> spm_dcm_fmri_check.
>>>>>
>>>>> (see
>>>>>
>>>>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1411&L=spm&F=&S=&X=
>>>>> 4
>>>>> C60EE6C21C6F1C1B5&Y=eric.holst%40web.de&P=334445)
>>>>>
>>>>> As this could be due to complexity of my DCMs I have reduced my
>>>>> basic model from 6 to 3 regions, but explained variance is still
>>>>> at 0%.
>>>>>
>>>>> I have read about increasing the area of the VOIs (I am using a 6
>>>>> mm sphere), but most of my VOIs already consist of 30+ voxels.
>>>>> [Btw: I smoothed my data in preprocessing.]
>>>>>
>>>>> Now I am wondering about the right way to proceed. What should I
>>>>> do now to improve my models?
>>>>>
>>>>> Any help is appreciated and thanks in advance,
>>>>>
>>>>> Eric Holst
>>>>>
>>>>> Department of Neurology
>>>>> Charité, Berlin, Germany
>
