SPM How-tos SPM预处理及统计分析指南

Source: SPM How-tos

NOTE: The manual is detailed but too long to edit. Just pasted here for further consulting.

---

This web-page contains excerpts from the SPM user-group's burster. SPM users around the globe run into the same problems as you do- here are there questions, with answers from the experts! You can search the SPM burster archives (at the official SPM website) for a particular keyword, or peruse this web-page, which has the same items, only organized topically. You will find that this web-page adheres to the Socratic method, but you don't have to do all that walking to and fro. There are several general categories:

Menu Items: Helpful hints about how to use the various menus, hidden menu features, and various display options. Model Items: Examples of many common models, considerations for selecting a particular model, etc. This section very likely contains an example similar to what you want to do. Analysis Items: Discussion and examples of how to use and apply various aspects of analysis within SPM. Particularly helpful for beginners are the discussions about Contrasts and how they work. Concepts: Discussions of any of the above areas that are more theoretical and of a more general nature than many of the answers to specific user's problems. Data Acquisition Pointers: Hints about acquiring your data so it is useful in subsequent analysis steps. Discussion of drop-out artifact in fMRI. Click on an item in the outline below to go to a particular topic, or just dig in and start reading! (Items without hyperlinks in the outline are still pending...) Note that most of the topics (but not the individual discussions) are repeated in the several major categories, so if you want to find out about e.g. Contrasts, you can look in the "Menu Items" for how to enter a contrast, in "Analysis Items" for what contrast may be appropriate to a particular study, and in "Concepts" for a more detailed look at how contrasts work under the hood.

Many of the answers/responses have been edited slightly for brevity's sake. If you feel there was too much editing, you can easily search the archive and obtain the full text, since the responder and date are included after most of the entries.

Return to LfAN SPM resources page.

Outline:

1. Interface items:

Selecting Image Data fMRI design with >16 sessions: picking scans Plotting variables: fMRI time-courses. Plot 'contrast of parameter estimates' Plot parametric responses Values from plots Masking Segmentation Contrast Manager View a large number of sessions Conjunction Analysis Pixel Coordinates for voxels in a cluster Global normalization mean global estimates threshold for Images Displaying t-statistic images Saving Printing print results to a text-file Img Calc hints Memory (RAM) management hints

2. Model Items (Creating a design matrix)

1 subj compared to controls 1 group, 2 conditions 1 group, 2 conditions, 1 covariate 1 group, 2 conditions, 1 covariate, 2 nuisances 1 group, 2 conditions, 3 levels/condition 1 group, 4+ conditions 1 group, multi-factor design 2 groups, 2 conditions 2 groups, 2 conditions, 1 covariate 3 groups (2 patient, 1 control), 2 conditions Event-related

3. Analysis Items

Realignment (a.k.a. Motion Correction) Spatial Normalization Talairach Coordinates Smoothing Covariates Contrasts F-contrasts explained Time x Condition A > (B & C) Conjunctions General With Small Volume Correction (SVC) Differences between SPM96, SPM99 Orthogonal Contrasts Random Effects General 2 groups, 2 conditions 3 conditions (A, B, Rest) 2 conditions, 1 covariate, 1 nuisance 4 conditions, 4 matched rest Variable event-related fMRI HRF (Hemodynamic Response Function) Extract fMRI time-course Slice Timing Interslice gap Small Volume Correction Hemisphere (L/R) effects Image Scaling Output Files

4. Concepts

Spatial Normalization Smoothing Smoothness estimates Gaussian Field Theory Fixed vs. Random Effects Contrasts Confounds Eigenimages Power Analysis fMRI Time Modeling

5. Data Acquisition Pointers

fMRI Susceptibility Artifacts

1. Interface Items

Selecting Image Data

fMRI design with >16 sessions: picking scans

We are attempting an fmri analysis, using a design matrix with 16 sessions. When we run the estimation model, instead of asking us which scans are required for each individual session, it asks which scans are required for all sessions. Choosing all scans at the same time does not work. When we use 12 sessions or lower, we are asked which scans are required for each session (1 through 12), and the analysis runs correctly. Does anyone know why can we do this with 12 sessions and not 16?

This is my fault. The idea was that some designs could be viewed as a series of short sessions (e.g. burst-mode or sparse sampling). In this context it would be easier to select all sessions at once. The limit is 16 sessions. To change this, say to 32, change line 236 in spm_fmri_spm_ui.m from

if nsess < 16 to if nsess < 32 % get filenames %--------------------------------------------------------------- nsess = length(xX.iB); nscan = zeros(1,nsess); for i = 1:nsess nscan(i) = length(find(xX.X(:,xX.iB(i)))); end P = []; ****if nsess < 16 **** for i = 1:nsess str = sprintf('select scans for session %0.0f',i); if isempty(BCH) q = spm_get(Inf,'.img',str); else q = sf_bch_get_q(i); end %- P = strvcat(P,q); end else [Who/when ???}

Plotting variables

Plot fMRI time-course

Y is the fitted response. y is the adjusted data (adjusted for confounds and bandpass filtering) Raw data has to be read from the Y.mad file.

For the plot option "fitted and adjusted responses", Y and y refer to the whole timeseries. For the plot option "event- or epoch-related responses", Y and y refer to peristimulus time (effectively "averaging across trials"). The PSTH option is just a way of binning the y values into separate peristimulus timebins, allowing calculation of mean and standard errors within each timebin. There is no easy way of collapsing y across trial-types (eg plotting a contrast of event-related data), other than collapsing across trial-types in the original model (ie having one column only).

Ran into a problem when I tried to plot fitted and adjusted responses against time a couple times. Despite high significance, SPM99 indicated that no raw data had been saved at that voxel and gave me the option of moving to the closest point which had data; when I chose this option, it went to another cluster entirely. In one case, this was for the local maximum of the most significant cluster (better than 0.000 uncorrected for the voxel and for the cluster). The entire area was grayed in on the glass brain. The default for the statistics was set at 0.001 so it seemed like there should not have been a problem. Am I doing something wrong or is this a bug? For which voxels is SPM99 supposed to be saving raw data for? This apparent paradox is due to the fact that the p value for a particular T-contrast may be more significant than the default F-contrast used to decide whether to save data in Y.mad. All I can suggest is that you reduce you default threshold further. Note that you can still plot fitted responses and standard error for every voxel (but not the actual residuals thenselves unless the data are saved in Y .mad). To do this simply decline the option to 'jump'. [Karl Friston 3 Jul 2000] Plot 'contrast of parameter estimates'

17. I have some PET data I'm trying to interpret, but I'm not sure about what is plotted in the 'contrast of parameter estimates' when they are plotted for each condition for an effect of interest in SPM99.

18. This bar-plot shows the mean-corrected parameter estimates of all effects of interest. The red lines are the standard errors of the parameter estimates.

19. This plot seems to be representing the 'size' of the effect on interest at a given maxima - my question is how a negative value in this plot should be interpreted. Is it a 'deactivation'?

20. Because of the mean correction, the bar-plot shows the deviations of the parameter of interest estimates from their mean. Therefore a negative value does not necessarily mean that the parameter estimate is negative, it is just lower than the mean of the parameter of interest estimates. Note that the (non mean-corrected) parameter estimates of a given voxel are stored in the workspace variable 'beta', when you plot them. By typing beta in your matlab window, you can display them.

21. Or, asked another way, what does the 0 effect size mean in these plots?

22. It means that this parameter estimate is equal to the mean of all parameter of interest estimates. As a special case of only one parameter of interest, it would mean that this parameter is zero.

I guess that the typical use of this plot is to easily assess the relative sizes of the parameter estimates for a given voxel. You could also use this plot to extract the vector of parameter estimates (and other variables like the standard errors of the parameter estimates, the fitted and the adjusted data stored in 'SE', 'Y' and 'y') from SPM99. [Stefan Kiebel, 20 Jun 2000]

17. Can anyone tell me what exactly is being plotted when I choose "contrast of parameter estimates" for my plot.

18. This plot shows one or more linear combinations of the parameter of interest estimates, where the linear combinations are contrasts. In the case, when you specify 'effects of interest', there is one (mean corrected) contrast for each parameter such that each grey bar shows the relative height of each estimated parameter of interest. In the case that you specify one of your own contrasts, the single bar shows the estimated parameters of interest weighted by the contrast. In both cases, the red line denotes the standard error SE of the weighted parameter estimates. The range of the red line is [-SE SE]. If you like to read some informative matlab code, the possibly most exact description can be found in spm_graph.m, lines 231 - 251.

19. How does this relate to the fitted response?

20. Let the general linear model used in the analysis be Y = X * \beta + \epsilon where Y are the functional observations, X is a design matrix, \beta the parameter vector and \epsilon the error of the model. Let b be the estimated parameters. The design matrix X can be subdivided into X = [X_1 | X_2], where X_1 denotes the covariates of interest and X_2 the covariates of no interest. Equally, b = [b_1 b_2]. Let c be the contrast(s) you choose for the plot, where c is a matrix with onecontrast per column. Then c' * b is the height of the grey bar(s) plotted. Note that c is zero for all estimated parameters of no interest b_2. The fitted (and corrected for confounds) data is then given by X_1 b_1. To make it complete, the adjusted data (and corrected for confounds) is given by X_1 * b_1 + R, where R are the residuals R = Y - Xb.

In other words, the relationship between the contrast of parameter estimates and the fitted response is given by the parameter estimates. In one case you weight the parameter of interests by a contrast vector and in the other case, you project the estimated parameters (of interest) back into the time domain. [Stefan Kiebel, 21 Jun 2000]

Plot parametric responses

I want to plot parametric responses (time x condition effects) using the same scaling (e.g. from -1 to 2 with 0.5 steps) on the z-axes for different subjects . In the interactive windows "attrib" (plot controls) I can change only the x-axes (Xlim, peristimulus time) and the y-axes (YLim, time) but not the z-axes (responses at XYZ). Has someone a modified matlab script (I think spm_results_ui.m and spm_graph.m) to do this ?

One could of course modify spm_results_ui.m, but I think the shortcut for you is to change the ZLim (or any other property of the plot) directly from within matlab. To make the figure the current axes, click on your plot and then type: set(gca, 'ZLim', [-1 2]); and set(gca, 'ZTick', [-1:0.5:2]); [Stefan Kiebel 17 July 2000]

Values from plots

Is there a way to use the matlab window to obtain the values used by SPM to generate plots (contrast of parameter estimates)? I am interested in obtaining the plot values and the standard deviation.

Yes, during each plot in SPM several values are stored in workspace variables. When you plot the parameter estimates or a contrast of these, SPM writes the variables beta (vector of parameter estimates) and SE (standard error) to the workspace. If you look at spm_graph.m lines 241 - 251, you can e.g. see how SPM99 generates the bar plot based on beta and SE. [Stefan Kiebel 27 July 2000]

Back to Outline.

Segmentation

To customize the segmentation:

One of the steps is to explore [spm_sn3d.m] file. And then insert the following line at the beginning of the main routine (way below where the %'d lines end, and after the definition of global values, eg, line number 296) in spm_sn3d.m sptl_CO=0; This will direct you, when you're running normalization, to choose all the options currently available in SPM normalization. Read the descriptions in the comment lines. [Jae S. Lee 21 Jun 2000] Back to Outline

Contrast Manager

View more than 51 sessions

we have constructed a design matrix with 60 sessions. When we explore the design we are able to view only 51 sessions. Is it possible to check the other 9 sessions? Which is the matlab routine where the max number of session are specified?

Actually, it is only partially a SPM issue. Your 60 sessions are still there, the limitation is due to the inability of matlab5.3.1 to display menus with more than 51 entries on your screen. To see the other sessions as well, you could type the following in matlab after starting spm and cd to your analysis directory:

load SPM_fMRIDesMtx.mat spm_fMRI_design_show(xX,Sess,60,1)

This would show you trial 1 of session 60. Change the last two arguments to see the other sessions and trials. [Stefan Kiebel 05 Jul 2000]

I also have a programing question. When attempting to plot "contasts of parameter estimates" I am not able to view or choose from all contrasts. I have a data set with about 50 contrasts and I am only able to choose from those that fit on the screen. If I type the contrast number, SPM only allows me to enter 1-9. Is there any way to plot the data for contrasts that do not fit in the window? Yes, there is a way around... It involves some typing:

1. Change line 213 in spm_graph.m from Ic = spm_input('Which contrast?','!+1','m',); to Ic = spm_input('Which contrast?','+1','m',);

2. Before plotting type in the matlab window global CMDLINE CMDLINE = 1

The first action makes sure that you can get into command line mode and the second actually activates the command line mode. [Stefen Kiebel, 14 July 2000]

Back to Outline

Masking

Is it possible to instruct spm99 to search all voxels within a given mask image rather than all above a fixed or a %mean threshold?

Yes, with SPM99 it's possible to use several masking options.

To recap, there are 3 sorts of masks used in SPM99:

1. an analysis threshold
2. implicit masking
3. explicit masking

1: One can set this threshold for each image to -Inf to switch off this threshold. 2: If the image allows this, NaN at a voxel position masks this voxel from the statistics, otherwise the mask value is zero (and the user can choose, whether implicit masking should be used at all). 3: Use mask image file(s), where NaN (when image format allows this) or a non-positive value masks a voxel.

On top of this, SPM automatically removes any voxels with constant values over time.

So what you want is an analysis, where one only applies an explicit mask.

In SPM99 for PET, you can do this by going for the Full Monty and choosing -Inf for the implicit mask and no 0-thresholding. Specify one or more mask images. (You could also define a new model structure, controlling the way SPM for PET asks questions).

With fMRI data/models, SPM99 is fully capable of doing explicit masking, but the user interface for fMRI doesn't ask for it. One way to do this type of masking anyway is to specify your model, choose 'estimate later' and modify (in matlab) the resulting SPMcfg.mat file. (see spm_spm.m lines 27 - 39 and 688 - 713). Load the SPMcfg.mat file, set the xM.TH values all to -Inf, set xM.I to 0 (in case that you have an image format not allowing NaN). Set xM.VM to a vector of structures, where each structure element is the output of spm_vol. For instance: xM.VM = spm_vol('Maskimage'); Finally, save by save SPMcfg xM -append

If so, does the program define a voxel to be used as one which has nonzero value in the named mask image?

Not nonzero, but any positive value and unequal NaN. Note that you can specify more than one mask image, where the resulting mask is then the intersection of all mask images. [Stefan Kiebel 27 Jun 2000]

Do I have to mask this contrast by another contrast (e.g. main effect) and how can I specify the masking contrast? You do not have to but if you wanted t; use a 2nd-level model with (Ae-Ce) in one column and (Be-Ce) in another (plus the constant term). Then mask [1 -1 0] with [1 1 1]. The latter is the main effect of Factor 1. [Karl Friston 18 July 2000]

[also see "Model Items: 3-factor design"]

For those of you wanting to specify explicit masking at the SPM (PET/SPECT) model setup stage, here's a recipe to do it without having to resort to the "Full Monty" design: Start SPM99 and paste the following into the MatLab command window:

%-Choose design class D = spm_spm_ui(char(spm_input('Select design class...','+1','m',...% {'Basic stats','Standard PET designs','SPM96 PET designs'},...% {'DesDefs_Stats','DesDefs_PET','DesDefs_PET96'},2)));

%-Choose design from previously specified class D = D(spm_input('Select design type...','+1','m','))

%-Turn on explicit masking option D.M_.X = Inf

%-Pass this design definition to SPM (PET/SPECT) spm_spm_ui('cfg',D)

[Andrew Holmes 20 July 2000]

It appears masking is a binary operation-- does this mean the mask specified must be in a bitmapped {0,1} format, or just that it is treated that way? The latter. The mask can have any numbers. If the mask image format (e.g. 'float') supports NaN, NaN is the masking value, otherwise it is 0. [Stefan Kiebel 21 July 2000]

With respect to estimating a model. I would like to potentially do an apriori mask of my collected brain. I could go in and just change all of my img files and mask explicity each one (ie zero out the non-interesting portions), however, any hints on where in the estimation code I would insert a masking to zero out the portions of the brain that I am not interested in estimating. That is, if we could we would have acquired a smaller region of volume during the scanning, but I can affect this by just masking my data before estimation.

Absolutely, if you want to assess the number of voxels above a given threshold, you can count these in the t-images. With respect to your question about the masking to effectively constrain the analysis to a ROI, you could look at http://www.mailbase.ac.uk/lists/spm/2000-06/0196.html http://www.mailbase.ac.uk/lists/spm/2000-07/0205.html which might provide a solution, how to implement your explicit masking easily (without changing each image, but just constraining the analysis to a subset of voxels). If you do an explicit masking, a script to counting voxels above threshold in a ROI wouldn't be necessary, because then you could use the cluster sizes as computed by SPM. You could also try to use a mask-image to apply the SVC.

Mask part of the brain

for the analysis of SPECT perfusion data, I would like to "crop" my images prior to statistical analysis

• that is, remove non-brain counts [scalp, sinuses, muscles]

from reading about "Mask object" in spm_sn3d, I gather spm will not do this during this step. True?

if not, is there a function available to do so?

Yes, there is a function to do exactly what you want. During the statistical analysis set-up, you can specify an explicit masking. To get to this and related masking options, you have to choose Full Monty as your design option. Then you can specify a mask-image, which could be in your case e.g. a normalized cropped image, where NaN (or 0) would mean to exclude this voxel from the analysis. You find a more detailed documentation about this type of masking in the SPM-help for PET-models. [Stefan Kiebel 19 Jul 2000] Back to Outline

Conjunction Analysis

Conjunctions are specified by holding down the'control' key during contrast selection.

Pixel Coordinates

Get pixel coordinates for all voxels within an activated cluster

One easy way would be to position the cursor on the cluster you're interested in (after displaying the results using the 'results' button), and paste the following lines from spm_list.m at the matlab prompt:

[xyzmm,i] = spm_XYZreg('NearestXYZ',... spm_results_ui('GetCoords'),SPM.XYZmm); spm_results_ui('SetCoords',SPM.XYZmm(:,i)); A = spm_clusters(SPM.XYZ); j = find(A == A(i)); XYZ = SPM.XYZ(:,j); XYZmm = SPM.XYZmm(:,j);

The last two variables - XYZ and XYZmm - would contain the pixel and the mm coordinates of all voxels in the current cluster. (Check the cursor to see where it is after pasting the above, it may jump a bit, moving to nearest suprathreshold voxel.) [Kalina Christoff 25 Jun 2000]

You could also use spm_regions in 'results' (VOI)

help spm_regions

VOI time-series extraction of adjusted data (local eigenimage analysis) FORMAT [Y xY] = spm_regions(SPM,VOL,xX,xCon,xSDM,hReg);

SPM - structure containing SPM, distribution & filtering detals VOL - structure containing details of volume analysed xX - Design Matrix structure xSDM - structure containing contents of SPM.mat file xCon - Contrast definitions structure (see spm_FcUtil.m for details) hReg - Handle of results section XYZ registry (see spm_results_ui.m)

Y - first eigenvariate of VOI xY - structure with: xY.name - name of VOI xY.y - voxel-wise data (filtered and adjusted) xY.u - first eigenvariate xY.v - first eigenimage xY.s - eigenimages *** xY.XYZmm - Co-ordinates of voxels used within VOI *** xY.xyz - centre of VOI (mm) xY.radius - radius of VOI (mm) xY.dstr - description of filtering & adjustment applied

Y and xY are also saved in VOI_*.mat in the SPM working directory [Karl Friston 26 Jun 2000] Back to Outline

Global Normalization

See mean global estimates for individual raw scans.

load SPMcfg.mat plot(xGX.rg)

Change the threshold for global normalization.

If you want to try different thresholds, then you need to modify line 57 of spm_global.c, and then recompile. The modification would involve something like changing from: s1/=(8.0m); to: s1/=(4.0m); Back to Outline

Images

Displaying t-statistic images

| 1) This may be a really idiotic question, but how does one view the | uncorrected t-statistic images? I'm assuming that viewing the t-statistic | images for a given contrast using the default values: "corrected height | threshold = no", "threshold = 0.001", and "extent threshold | = 0" still applies a correction that is based on the the | smoothness estimates and consequently the number of resels.

This displays the raw uncorrected t statistics that are more significant than p<0.001. There is no correction for the number of resels when you dont specify a corrected height threshold.

Another way of displaying the statistic images is to use or . [John Ashburner 21 Jun 2000]

Saving Images

I'm performing a manual rotation and I don't know how to save the rotated image. Use the display button. Your image will come up in the graphics window. Use the gray boxes to the left and below the image to alter the orientation, then, when you are happy with the result, press the reorient images button in the same window. spmget will launch. Select the images you want to be rotated (the image you have been working on +/- any others), and the changes to the orientation will be written out in a *mat file. [Alex Leff 19 July 2000] Back to Outline.

Printing

Print results to a text-file

A right click in the background of an SPM results table brings up a context menu including options to "Print Text Table" and "Extract Table Data Structure". The first prints the table as plain text in the Matlab command window, the second returns the table data structure to the base matlab workspace (as 'ans'). See the help for spm_list.m for further details (also available from the table context menu as "help").

Img Calc Hints

| I'd like to create, for each individual subject, a subtraction image that | reflects %change in normalized rCBF. Thus, instead of t-values, the pixel | values of this image would be numbers reflecting change above or below | average whole brain. In my particular case, I have two baselines and two | activations, so I'd like to create the percent change subtraction image of: | (i1+i3)/2 - (i2+i4)/2. | | Is there a way to easily accomplish this in SPM? As far as I can tell, | proportional scaling only comes as part of a process that produces a | statistical parametric map image, and I don't see anything in the image | calculator that would enable me to perform this step separately (i.e., take | an image, normalize each pixel by whole brain average, and then do the | subtractions).

In Matlab, you can obtain the "globals" for each image by: V = spm_vol(spm_get(4,'*.img')); gl1 = spm_global(V(1)) gl2 = spm_global(V(2)) gl3 = spm_global(V(3)) gl4 = spm_global(V(4))

Then these can be plugged into the ImCalc expression by: (i1/gl1+i3/gl3)/2 - (i2/gl2+i4/gl4)/2

I think you actually need to enter the values of the globals rather than the variable names. [John Ashburner 11 Aug 2000]

Back to Outline.

Memory (RAM) management hints

If you have problems with SPM halting, and perhaps with your Matlab session also quitting, type the following before entering Matlab: unlimit stacksize N.B. this only works on a Unix machine.

Back to Outline.

2. Model Items

1 subject compared to controls

| 1. How can SPM best be used to compare a single subject to a group of | controls in order to establish the pattern of regional abnormalities? I | have tried using the two sample t-test, two groups, one scan per subject | model, with success, but was wondering if anyone had ideas about other | approaches using the software.

This is probably the best approach, but it may be worth also modelling confounding effects such as age or possibly nonlinear age effects (by also including age^{2 and age}3).

Depending how many controls you have, you may also wish to try a non-parametric analysis using SNPM. [John Ashburner 13 July 2000] Back to Outline.

1 group, 2 conditions

number of conditions or trials : 1 (is this correct? Should I enter "2"?)

Yes. With one condition alternating with rest it is appropriate to model the rest implicitly by specifying just the active condition onsets. To use 2 conditions would not be wrong, but is redundant.

Results button -> I set default value for mask, threshold and so on. I set t-contrast "1 -1" or "-1 1", is it correct? I want to z-score, which is (mean(rest)-mean(activation))/SE, but the different options give different z-scores.

This is what you are doing wrong I think. You specified one condition so have two columns in the resulting design matrix. One represents the boxcar (activation vs rest), the other is a constant term modeling the mean activity over all conditions. Your t-contrasts are comparing these two regressors, which will give weird results.

What you should do is use contrasts [1] or [-1] to see areas where activation>rest, or rest>activation respectively. If you had used two regressors to model activation and rest separately then the corresponding contrasts would be [1 -1] and [-1 1]. [Geraint Rees 25 July 2000] Back to Outline.

1 group, 2 conditions, 1 covariate

PET/SPECT models: Multi-subject, conditions and covariates

| I'm trying to do simple correlations with SPM99..will someone please | help me, this should be very simple. | | I have 2 PET scans per subject, one at baseline and one on drug. I | have 2 clinical rating scores, one at baseline and one after drug. | I want to look at increases in GMR after drug correlated with | increases in the clinical rating. I also want to look at negative | correlations. What model should I use and how do I define the | contrasts??

PET/SPECT models: Multi-subject, conditions and covariates. For each subject, enter the two scans as baseline and then drug. One covariate, values are the clinical rating scores in the order you selected the scans, i.e. baseline score for subject 1, drug score for subject 1, baseline score for subject 2, drug score for subject 2, &c. No interactions for the covariate. No covariate centering. No nuisance variables. I'd use proportional scaling global normalisation, if any. (You could use "straight" Ancova (with grand mean scaling by subject), but SPM99 as only offers you AnCova by subject, which here would leave you with more parameters than images, and a completely unestimable model).

Your model (at the voxel level) is:

[1] Y_iq = A_q + C * s_iq + B_i + error

...where: Y_iq is the baseline (q=1) / drug (q=2) scan on subject i (i=1,...,n) A_q is the baseline / drug effect s_iq is the clinical rating score C is the slope parameter for the clinical rating score B_i is the subject effect

...so the design matrix has: 2 columns indicating baseline / drug 1 column of the covariate n columns indicating the subject You will have n-1 degrees of freedom. Taking model [1] and subtracting for q=2 from q=1, you get the equivalent model:

[2] (Y_i2 - Y_i1) = D + C(s_i2-s_i1) + error

...where D = (A_2 - A_1), the difference in the baseline & drug main effects. (Note that this only works when there are only two conditions and one scan per condition per subject!) I.e. a simple regression of the difference in voxel value baseline to drug on the difference in clinical scores, exactly what you want.

---

Entering [0 0 1] (or [0 0 -1] as an F-contrast will test the null hypothesis that there is no covariate effect (after accounting for common effects across subjects), against the alternative that there is an effect (either positive or negative. I.e., the SPM will pick out areas where the difference baseline to drug is correlated with the difference in clinical scores.

[0 0 +1] and [0 0 -1] as t-contrasts will test against one sided alternatives, being a positive & negative correlation (respectively) of baseline to drug scan differences with difference in clinical scores. Since you're interested in both, you should interpret each at a halved significance level (double the p-values). This will give you the same inference as the SPM (which is the square of the SPM's), but with the advantage of separating +ve & -ve correlations in the glass brain for you.

---

Incidentally, the variance term here incorporates both within and between subject variability, and inference extends to the (hypothetical) population from which you (randomly!) sampled your subjects from. [Andrew Holmes, when ???]

Given 2 conditions, 1 scan/condition, 1 covariate obtained at each scan, mean-centered covariate with proportional global scaling. A condition & covariate design with a contrast 0 0 1 is equivalent to correlation between the change in covariate and the change in the scans.

Indeed or more precisely the partial correlation between the covariate and scan-by-scan changes having accounted for the condition-specific activations. [Karl Friston 28 Jun 2000]

I have a SPECT study with 34 patients and 2 conditions per patients and 1 covariate. I want to find the regions where there is positive corelation between the rise in blood flow from the first scan to the second scan with the covariate. I centered the covariate around 0 and used a new covariate of +a/2,-a/2 as a new covariiate as recommended by Andrew Holmes.

Could anyone please explain to me what would the difference be in this case, if I use the "Multi subject covariate only" design or a "Multi subject condition and covariate design" and use a [0 0 1] contrast.

If you use 'Multi subject condition and covariate design', the model expresses your assumption that each series of observations in a voxel (over subjects) can be explained by subject effects, condition effects (which are the same for all subjects) and by your covariate.

If you choose 'Multi subject covariate only', you express your belief that there is no need to model a condition effect, but that your covariate alone times the estimated slope is a good explanation for your observations.

So the difference between the two models is that in the first you model some additive condition effect commonly observed over all subjects. [Stefan Kiebel 25 July 2000]

Back to Outline.

1 group, 2 conditions, 1 covariate, 2 nuisances

I will first start with what we have: Within an fmri study, One group Five subjects Two conditions Auditory Monitoring versus its own baseline Working Memory versus its own baseline Two nuisance variables anxiety score (one score per subject) Depressive mood score (one score per subject)

One covariate of interest error score on the working memory task This is what we did Design Description Desgin: Full Monty Global calculation: mean voxel value (within per image fullmean/8 mask) Grand Mean scalingL (implicit in PropSca global normalization) Global normailzation: proportional scaling to 50 Parameters: 2 condition, +1 covariate, +5 block, +2 nuisance 10 total, having 7 degrees of freedom leaving 3 degrees of freedom from 10 images

Is this a valid way of looking at this? We are concerned with the large degrees of freedom that we are using up. Also how would we accurately interpret such a model? Does the statistical map only represent activations that are associated with the covariate of interest after controlling for anxiety and depression scores?

Firstly I assume this is a second level analysis where you have taken 'monitoring' and 'memory' contrasts from the first level. If this is the case you should analyse each contrast separately. Secondly do not model the subject effect: At the seond level this is a subject by contrast interaction and is the error variance used for inference. Thirdly a significant effect due to any of the covariates represents a condition x covariate interaction (i.e. how that covariate affects the activation).

I would use a covariates only single subject design in PET models (for each of the two contrasts from the first level). A second-level contrast testing for the effect of the constant term will tell you about average activation effects. The remaining covariate-specific contrasts will indicate whether or not there is an interaction. [Karl Friston 17 July 2000]

Back to Outline.

1 group, 2 conditions, 3 levels/condition

we're attempting to conduct a parametric analysis. We have two condition A(experimental condition) B(control condition); in the experimental condition the parameter assumes 3 different values.

1. Which is the difference between choosing a polynomial or a linear relationship in the model?

A linear model is simply a polynomial model with 0th and 1st order terms. Any curvilinear relationship between evoked responses and the parameter of interest would require 2nd or higher order terms to be modeled.

2. In the results session how can we specify the contrast for the AB difference? and for the parameter effect on the experimental condition?

Simply test for the [polynomial] coefficients one by one. The 0th order term (e.g. [1 0 0]) models the mean difference between A and B averaged over the three levels. The 1st order coefficient (e.g. [0 1 0]) reflects the linear dependency on the parameter and the 2nd (e.g. [0 0 1]) or higher reflect the nonlinear components. The 0th order term is the conventional box-car or condotion-specific effect modeled in simple, non-paramteric analyses. Note that because you only have 3 levels in condition A a 2nd order model is the highest you would consider (- a parabola can join three points together). [Karl Friston 29 Jun 2000]

1 group, 4+ conditions

I chose the multi-subject: conditions and covariates design, for PET scan. The four scans/conditions (A,B,C,D) were entered in time-order. I want to compare only two conditions (e.g. B and C) in the analysis. Do I have to put the other conditions on 0 while defining contrasts (i.e. 0,-1,1,0), or do I have to make another spm.mat file in which only the two conditions I want to compare are taken and define contrasts as 1,-1? Is there a difference between the two ways?

The first solution is the more appropriate one. One should generally try to specify one design matrix modelling all observations and then use the one estimated parameter set to compute all the contrasts.

The difference between the two solutions is that in the first case you make the assumption that it is valid to use all scans for estimating the variance under the null hypothesis, even if a contrast vector element is zero for the associated basis function/condition. This is a valid assumption for a fixed effects PET analysis. As a result, you have more degrees of freedom in your statistical test at each voxel such that the analysis is more sensitive to the underlying signal. [Stefan Kiebal, when ???]

Further to my question to you earlier this week which was:

17. My paradigm is a block design with 4 different active blocks each followed by it's respective null block, i.e I have 4 different null blocks. How do I go about specifying the design matrix for a second level analysis taking these different null blocks into account? e.g. if my 4 active blocks are: A1 A2 A3 A4 and my 4 null blocks are: N1 N2 N3 N4

If I specify trials 1-8 in the following order: A1 N1 A2 N2 A3 N3 A4 N4

how do I contrast [A1-N1] - [A2-N2]? or vice versa?

Your answer was: A.You would simply specify 8 conditions (A1 - N4) and use the appropriate contrasts. Unfortunately, 'use the appropriate contrasts' is the bit we don't know how to do now that we have so many different nulls.

I have specified the conditions 1-8:A1 N1 A2 N2 A3 N3 A4 N4 for simple contrast A1-N1 I've used:1 -1 0 0 0 0 0 0 & for contrast A2-N2 I've used: 0 0 1 -1 0 0 0 0 How do I specify a 2nd level contrast looking at the activity in A1 minus it's null N1 versus the activity in A2 minus it's null N2, i.e. [A1-N1] - [A2-N2]?

If I use:A1 N1 A2 N2 A3 N3 A4 N4 1 -1 1 -1 0 0 0 0 then surely this is just adding the activity in A1 and A2 and taking away the activity in N1 and N2 which is not what we want to do.

In fact this is [A1-N1] - [A2-N2] and is exactly what you want. I think the confusion may be about the role of the 2nd-level analysis. To perform a second level analysis simply take the above contrast [1 -1 1 -1 0 0 0 0] and create a con???.img for each subject. You then enter these images into a one sample T test under 'Basic Designs' to get the second-level SPM. To do this you have to model all your sujects at the first level and specify your contrasts so that the effect is tested in a subject-specific fashion:

i.e. subject 1 [1 -1 1 -1 0 0 0 0 0 0 0 0 0 0 0 0 ... subject 2 [0 0 0 0 0 0 0 0 1 -1 1 -1 0 0 0 0 ... ... [Karl Friston 19 July 2000]

how do I contrast [A1-N1]-[A2-N2]? or vice versa? i.e perform a 2nd order contrast. The first issue is exactly what question you are asking. [A1-N1] vs [A2-N2] looks like an interaction, and I think that this is what you are after. You can think of it as comparing the 'simple main effect' Ax-Nx in two contexts, x=1 and x=2. Put another way, the interaction is the 'A-specific activity' in context 1 compared with the 'A-specific activity' in context 2, each being compared with its own baseline. Let me know if this is not what you need.

Would the appropriate contrast be A1-N1-A2+A1?

No, it would be A1-N1-A2+N2 (I suspect that this is what you meant and that the A1 on the end is just a typo). Thus with your covariates ordered as specified above, your contrast will be 1 -1 -1 1 0 0 0 0). As Karl pointed out (but for a different contrast), you now need to perform this contrast on each of your subjects, within the 'fixed effects' design matrix:

Subject 1: 1 -1 -1 1 0 0 0 0 0 0 0 0 0 0 0 0 ... Subject 2: 0 0 0 0 0 0 0 0 1 -1 -1 1 0 0 0 0 ... etc.

Each contrast image generated (i.e. one for each subject) gets entered into a one-sample t test in the 'second level' analysis. The question which you now ask of every voxel is whether its value departs significantly from zero (which is its expected value under the null hypothesis).

Incidentally it may be worth just mentioning an alternative (less good) approach, which I suspect that you might have been considering. (You can ignore this bit if you like.) You could specify the simple main effect contrasts A1-N1 and A2-N2, and test for the difference between them. Thus your contrasts would be

Subject 1 (A1-N1): 1 -1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ... Subject 1 (A2-N2): 0 0 1 -1 0 0 0 0 0 0 0 0 0 0 0 0 ... Subject 2 (A1-N1): 0 0 0 0 0 0 0 0 1 -1 0 0 0 0 0 0 ... Subject 2 (A2-N2): 0 0 0 0 0 0 0 0 0 0 0 1 -1 0 0 0 ...

In this case, the second level analysis would test whether the A1-N1 contrasts, as a population, are significantly greater than the A2-N2 contrasts. The reason why this is less good than the first approach outlined above is that it is equivalent to an unpaired t test (in which you just compare an 'A1-N1' population with an 'A2-N2' population) whereas your data are obviously paired (i.e. each A1-N1 estimate goes with the A2-N2 estimate for the same subject).

However, if you can do a paired t test, then as I understand it the result should be exactly the same as the first analysis - I've never tried this so I don't know if it is possible within SPM99. [Richard Perry 20 July 2000]

We have conducted a Working Memory study in H2O-PET with the following design:

8 subject 4 Conditions: A: WM 1 (high load) B: WM 1 (low load) C: WM 2 (high load) D: WM 2 (low load) 3 Scans/per condition/subject

Thus we have a total of 96 scans. To look for the main effect of Working memory and for the domain specific effect we have chosen the Multi-subject: cond x Subj interaction & covariates design. We find nice WM main effects and also interesting domain specific effects.

Experimental question: We are interested if there is a correlation between performance (as measured by RT) and WM-specific activation.

What kind of design do we have to choose? As we work mainly with fMRI-studies we first thought of a second level analysis. That is: Formulate subjects specific contrast, e.g. WM 1 high load minus WM 1 low load on the first level, then feed the eight resulting con-images into basic models (simple regression) and chosse the median of the RT for the three WM 1 high load scans as covariates into this modell. Then our analysis should yield regions in which there is a correlation of one WM domain with performance.

Question 1: Is this analysis correct?

Yes it is but it may not be the most sensitive analysis because you are proceeding to a second-level analysis whereas you have scan-specific performance measures.

Question 2: There are several options at the first level in which it is possible to specify scan specific covariates. Is it intelligible to choose one of these models and feed in scan specific RT in order to answer our experimental question? If so, what is the appropriate model? We have tried several but if we enter scans and covariates we use up all our degrees of freedom, e.g. if we use Mulit-subj: covariates only. So something must be wrong.

I would simplify your model and omit subject x condition interactions. You could then enter the condition x performance interaction as a covariate of interest. This is simply the mean corrected performance data multiplied by 1 for high load and -1 for low load. You could do this in a condition-specific fashion for WM 1 and WM 2 using two covariates but ensure the behavioural data are centered within condition before constructing the interaction (here use 0 for the 'other' condition). [Karl Friston, 20 Jun 2000]

I ran SPM99b. We have tested 11 normal young subjects with H2 O15 PET study in five different conditions 1 at rest and 4 with auditory stimulation. The first design matrix to check the differences between stimulation-no stimulation we used is: multisubject conditions & covariates conditions 0 -1 -2 1 2 (rest, cond 1, cond 2 cond 3 and cond 4) Ancova by subject scale subjects grand means to 50 analysis threshold 0.8 mean voxel value (within per image fullmean/8 mask) Is the right approach? Should be -4 1 1 1 1 ?

Yes, the design matrix specification part sounds fine. You should take the contrast [-4 1 1 1] to compare the mean effect of all stimulation conditions vs. rest.

The design matrix we used to look for different activity in one of the four conditions was.: multisubject conditions & covariates conditions 0 1 1 1 -3 (rest, cond 1, cond 2 cond 3 and cond 4 wich is the relevant) Ancova by subject scale subjects grand means to 50 analysis threshold 0.8 mean voxel value (within per image fullmean/8 mask) what is the difference using 0 -1 -1 -1 3, assuming we espected to find greater activity at condition 4?

The difference between [0 1 1 1 -3] and [0 -1 -1 -1 3] is that the first tests at each voxel, whether the mean effect of conditions 1 to 3 is larger than the effect of condition 4 and the second tests, whether the effect of condition 4 is larger than the mean effect of condition 1 to 3.

I was also interested in measuring rCBF in auditory cortex at the different stimulation conditions. Unfortunately the differences between conditions are so subtle that using a T test there is no significance with the actual number of subjects (11).

If you want to see the effect of each condition vs. rest, then you could try the contrasts [-1 1 0 0 0] [-1 0 1 0 0] [-1 0 0 1 0] [-1 0 0 0 1]

If you have a specific hypothesis that e.g. condition 4 should activate more than condition 1, you could try [0 -1 0 0 1]. [Stefan Kiebel 11 July 2000]

Back to Outline.

1 group, multi-factor design

I have a few questions concerning a multi-subjects fMRI experiment. For each subject we acquired 4 separate scan-sessions according to the following experimental design: session 1,2: Ce Cr Ce Cr Ce Cr Ce Cr Ae Ar Ae Ar Ae Ar Ae Ar session 3,4: Ce Cr Ce Cr Ce Cr Ce Cr Be Br Be Br Be Br Be Br where: e = encoding r = retrieval C = baseline A = condition1 B = condition2

Up to now, I have been able to analyze contrasts between conditions belonging to sessions of the same type (e.g. Ae-Ce; Ar-Cr for sessions 1 and 2) both using a fixed-effect and a 2nd-level random effect model.

In addition, however, I would like to compare conditions belonging to sessions of a different type - though I know it would have been much better to include all conditions in one session. E.g. compare Ae-Be. I have tried to do this on a random-effect basis, by computing the simple main effects for each subject (Ae-Ce and Be-Ce) and the entering the corresponding con*.images into a paired t-test (Ae-Ce vs Be-Ce). My questions here are: what am I actually looking by this approach? At interaction effects [of the type (Ae-Ce) - (Be-Ce)]? Yes indeed. You could construe your design as a 3 factor design: Factor 1: C vs. Active (A or B) (2 levels) Factor 2: e vs. r (2 levels) Factor 3: Condition 1 vs. Condition 2 (2 levels) Your effect is a 2 way interaction Factor 1 x Factor 3, under e.

Do I have to mask this contrast by another contrast (e.g. main effect) and how can I specify the masking contrast? You do not have to but if you wanted t; use a 2nd-level model with (Ae-Ce) in one column and (Be-Ce) in another (plus the constant term). Then mask [1 -1 0] with [1 1 1]. The latter is the main effect of Factor 1.

Is there any way to look at direct comparisons (Ae-Be)? Yes; just do a two sample t test on contrasts testing for Ae and Be separately. These will be the same as the beta???.img. [Karl Friston 18 July 2000]

Back to Outline.

2 groups, 2 conditions

We are analyzing PET data of subjects from 2 groups who underwent the same stimulation. Group1 contains 7 subjects, Group2 contains 12 subjects. We can do a single subject analysis and a RFX group evaluation of group 2. For the group evaluation of group 1 the number of 7 subjects might be not enough. (?) We want to do a comparison of the 2 groups. Which model do we have to use for the first and second level analysis?

For the 1st-level analysis use a 'multi-group conditions and covariates' for the 2nd-level analysis compare the two sets of subject-specific contrasts with a two sample t test under 'basic models'. The latter will have reasonable power because you are using all 19 subjects. [Karl Friston 10 Jul 2000]

Now have properly normalized 2 groups with 5 datasets each. Paradigm: RARARAR (A: Activation, R: Rest). Group I: controls, group II patients. Now I want to subtract the group I from group II to see, wether there is any difference between the two groups concerning condition A. How shall I proceed? If you want to assess the differences in activations this constitutes an inference about the group x condition interaction. Having modeled your two conditions in a multi-group PET design you will have four columns (A - group I, B - group I, A - group II and B - group II). The contrast you require is [1 -1 -1 1] for bigger activations in the controls and [-1 1 1 -1] for the converse. [Karl Friston 19 July 2000] Back to Outline.

2 groups, 2 conditions, 1 covariate

Given 2 conditions, 1 scan/condition, 1 covariate obtained at each scan, mean-centered covariate with proportional global scaling. A condition & covariate design with a contrast 0 0 1 is equivalent to correlation between the change in covariate and the change in the scans. Can this approach be generalized to a multi-group design ? If there are 2 groups and 2 conditions, with 1 scan for each subject under each condition, and a single covariate collected during each scan, then would one specify 1 or 2 covariates.

I would specify two, each with a group-centered covariate. This would allow you to look for differences in the partial correlation with the contrast [0 0 0 0 1 -1]. i.e. models group x covariate interactions.

If only one covariate is specified, would one test for a covariate effect in group 1 alone with the contrast 1 0 0 0 1 (Group 1, Group 2, Condition 1, Condition 2, Covariate), and the group x covariate interaction with the contrast 1 -1 0 0 1 ?

No. The interaction is not modeled with only one covariate. This contrast is simply the main effect of group plus the main effect of contrast.

Alternatively would one use 2 covariates ? Each covariate consisting of the mean centered values for a single group with 0 padding for the subjects in the other group, essentially what you get when you specify a group x covariate interation The effect of the covariate for group 1 alone would be tested with the contrast 0 0 0 0 1 0 ( (Group 1, Group 2, Condition 1, Condition 2, Covariate1, Covariate 2) and the interaction across groups tested with the contrast 0 0 0 0 1 -1 ?

Exactly. [Karl Friston 28 Jun 2000]

I have two groups of subjects who, on prescreening, exhibited differential mood responses (one group positive scores, the other negative) to a drug. So with 2 groups (positive and negative responders), 2 conditions (drug and placebo), 1 scan/subject under each condition, and 1 covariate (mood score) collected/scan, how would I go about determing whether rCMglu is affected by:

1. prescreening mood scores both within and across groups

These are the simple and main effects of mood (or group) and would be best addressed with a second-level analysis using the subject-effect parameter estimate images (i.e. averaging over condotions for each subject).

2. post drug scan mood scores both within and across groups and also whether:

This is a main effect of 'post scan mood score' within the post drug level of the condition effect. This is best analysed using a condition-centered covariate ('post scan mood score') and testing for a significant regression with 'drug'.

3. mean centering of the covariate is useful/necessary in these cases

Yes. For (1) there is not covariate required.

4. mean should be computed within group or across all subjects

For (2) within condition. There are other centering you could use to look at different interactions. e.g. group-centered would allow you to see if there was any interaction between prescreening and post scan mood. [Karl Friston 28 Jun 2000]

Back to Outline.

3 groups (2 patient, 1 control), 2 conditions (ABABAB)

| I have a data set involving 3 groups of subjects (2 different | patient groups and 1 control group) doing a simple ABABAB style | paradigm. I have made a con* image representing the difference | between A and B for each subject. I can do a within group analysis | by feeding the con images for one group into a 1-sample t-test and I | can compare 2 groups using a 2 sample t-test. What I would like to | do is compare all 3 groups. 3 pairwise comparisons would seem to be | the best I can do in "basic fMRI models". I assume it is more | elegant to mdel all 3 groups in a single analysis. But PET multi- | group models seem to need one image PER CONDITION rather than one | image per subject representing the difference between conditions. Is | there a way round this without generating spm98 style "adjmean" | images for each condition?

Try the "One-way Anova" option in the "Basic models". The default F-contrast produced is the usual analysis of variance F-statistic for any difference (in response, since you're looking at contrast images) between the three groups.

You can conduct follow-up comparisons (comparing pairs of groups), using simple contrasts (like [+1 -1 0] for example). However, you should adjust the significance level at which you examine these follow-up comparisons to take into account the number of comparisons: The simplest method is the Bonferroni method, in which you multiply your p-values by the number of planned follow-up comparisons. Note that although there are three ways to compare pairs of three groups, this corresponds to six contrasts for SPM's in SPM, since SPM's t-contrasts only effect one-sided tests.

Note that Anova assumes that the intra and inter-subject variability expressed in your contrast images is constant across the three groups. If this assumption is met, you'll get slightly more powerful paired comparisons from the Anova model than from simple pairwise comparisons because of the increased degrees of freedom available for variance estimation. In essence, you're comparing two groups with a variance estimator pooled across all three, even the one your contrast appears not to be looking at.

Comparison of the Anova follow-up paired comparisons with the "2 sample t-test" results will give some indication of whether the homoscedasticity assumption is met.

| NB: the number of subjects is not the same in each group if this is | relevant.

That's OK, but note that for the two level contrast image approach to random effects analyses to be valid, the experimental design for each subject (regardless of group) should be such that the design matrix is the same.

[Who / when ???]

Back to Outline.

Event-related model items

Variable epoch lengths

I am having a few problems setting up a model for a study with variable epoch lengths.

I have a resting breathing condition (12 x 30 second periods) and a voluntary hyperventilation condition (12 x 30 second periods)

In the first analysis I set up a model in which I specified two conditions (rest and voluntary), however I have since realised that I maybe should have treated the study as having only one condition - vol, and treated the resting periods as baseline.

It should not make any difference with box-car regressors and one basis fuction per condition.

In the second model I have specified 1 condition (treating the rest as baseline), epoch - fixed response box car, convolved with hrf and temporal derivative. The problem arises when I specify the epoch lengths (some are 5 scans in length instead of 6 because of variability in breathing). After I enter vector of onsets, SPM asks 'variable duration'. At this point I enter yes, then I am prompted by SPM to enter 'duration (scans)'. If I enter the variable epoch lengths at this point, I am also asked at a later point in the model set-up to enter 'Epoch length (scans) for trials'. However, SPM will only accept scalars and not a string of vectors (which specify the variable epoch lengths). I don't understand why it asks me twice to enter epoch length and will only accept scalars on the second prompt (when I set up the first analysis it accepted a string of vectors) Also, should I be specifying each condition in the design matrix or only the vol condition?

Variable length epochs are dealt with using the event-related options. Select event-related, not epoch-related when chosing your basis set. The simpler alternative would be to have two 'vol' conditions (one of 5 scans and one of 6 scans) and simply take to average using contrasts later on. [Karl Friston 17 July 2000]

HRF width

I'm trying to model event-related activity that is due to two stimuli presented sequentially with an SOA of 1.5 seconds. In looking at the raw data, the hemodynamic response to such an event often has a wider peak than does the typical HRF for a single event. I'm wondering whether modelling the data using a dispersion derivative (i.e., hrf + time + dispersion derivatives) will enable SPM to fit a canonical hrf to the data that has the appropriate width. In other words, does the dispersion derivative allow the width of the of the canonical HRF to be adjusted analagous to the way that the temporal derivative allows the onset of the canonical response to be adjusted? Absolutely. [Karl Friston 21 July 2000] Back to Outline.

3. Analysis Items

Realignment (a.k.a. Motion Correction)

| I realigned images [creating mean image only], and tried to run spm_sn3d, | only to be told "not enough overlap". when I viewed the data in Display or | Check Reg, I found that the origin was somewhere far Northeast of the | vertex. I then tried to set the coordinates using Hdr Edit, only to fail. I | think the message below says why: | | At 06:25 AM 03/24/2000 , you wrote: | >Once an image has a .mat file, then SPM99 ignores the origin and | >voxel size information in the headers. The best way of changing the | >origin is via the Display button. It also allows you to re-orient | | so, this means EITHER: a] set the origin in all images PRIOR to realign | using Hdr edit [which can be applied to many images at once] | OR b] once realigned, only Display can be used. [one image at a time] | True? Display can be used to reorient many images at the same time. It is simply a matter of selecting all the images that are in register with the one currently being displayed after you click the "Reorient Images..." button. The reorientation applies a relative transformation to the images, so if the images have been realigned or coregistered, or have different voxel sizes or whatever, they will still be in register with each other after reorienting.

| | To use Hdr edit, I would open an image in display, and set my crosshairs to | the vicinity of the AC. Then, I would enter those voxel coordinates [in mm] | into Hdr_edit, and apply those values to the relevant images. yes? If you were to do it this way, then it should work (providing the images have no .mat files). [John Ashburner 21 July 2000]

|Does one always need to reslice while doing movement correction, or is |coregister alone enough? I'm asking because it appears that one can reslice |later at the spatial normalization stage anyway. Reslicing once only at the normalisation stage will work fine. Note that the smoothness of the resultant data will be slightly less with one reslice than with two. [Geraint Rees 10 Jul 2000]

| Dear SPManagers: I am realigning multiple SPECT scans within | subjects, and am not clear what this option offers me. I have looked | through help files relating to realign, and am not enlightened.

Weighting of the reference image allows the realignment parameters to be estimated from only a specific region of the reference image. For example, there may be artifacts in the images that you do not wish to influence the estimation of the realignment parameters. By giving these regions zero weight in the realignment, they have no influence on the estimated parameters. The weighting procedure uses an image which can contain zeros and ones, or more properly, it can be thought of as containing the reciprocals of the standard deviation at each voxel (unlike weighting in the spatial normalisation where the weight is the reciprocal of the variance - I must fix this).

The function minimised in the realignment is something like: \sum_i (wt_i * ( g_i - s * f_i ))^2 whereas for the spatial normalisation it is more like: \sum_i wt_i * ( g_i - s * f_i )^2

| | I do understand that the defualt for PEt and SPECT is to make a mean | image from a "first-pass" realignment, and tehn realigns all images to | that. Is this other option data-type specific?

It does this whenever you use the PET or SPECT modality. It would also do this for fMRI, but I figured it would slow things down too much if it did two passes. Also, PET and SPECT images are noisier, so realigning to a mean image improves the results more. Ideally, the procedure would be repeated a few times, but again, this would be too slow.

| 1. is there a number which corresponds to hard-coding "Create mean image | only" for spm_realign?

I'm afraid the sptl_CrtWht variable only accepts the two values.

| | 2. what is the range of possible values for regularisation? what numbers | correspond to "medium" and "heavy"?

Any positive value you like. Medium and heavy are given by 0.01 and 0.1 respectively. A value of zero does not regularise the nonlinear part of the spatial normalisation. [John Ashburner 19 July 2000] | I have applied slice timing to my event-related series. For motion correction | (spatial realignment) I would like to use a non-SPM software. This software can | read the aV*.img images but not aV*.mat files. | | My question is: have aV* images been actually written with the slice timing | correction imbedded? That is, if I use them outside SPM without their .mat | files, will slice timing be still taken into account? If not, what should I do?

I think you can quite happily delete these .mat files as they probably don't contain any additional information. Whenever any new images are derived from existing ones, then the positional information is preserved in the new set of images. This positional information is derived from the .mat files if they exist, but otherwise from the voxel-size and origin fields of the .hdr files. If the origin contains [0 0 0], then it defaults to mean the centre of the image, which, depending if you have odd or even image dimensions, is either an integer value or halfway between two voxels. Because the origin field can only store integers, half voxel translations can not be represented in the .hdr files, so this information is written to .mat files. [John Ashburner 25 July 2000]

| I want to coregister a structural (anat.img) and 60 | functional (func.img) images. Shall I: 1. coregister | anat.img with button and 2. coregister | func.img with button or first do 2. and | then 1.. During second coregistration I am asked after | target and object image. What is best to choose for | this?

Realign the functional images, possibly creating a mean at the same time. Then coregister the anat.img to the mean (or possibly one of the individual images from the series). To do this, you would have target=mean_func.img, object=anat.img and other=none. [John Ashburner, 13 July 2000]

| In "realign" what does mean "adjust sampling errors" and when I must | use these option? It removes a tiny amount of interpolation error from the data. In reality it should not do much to the data, but it can increase your t statistics slightly. It does slow things down a lot though. Full documentation of what it does can be obtained via the facility in SPM99. [John Ashburner 2 Aug 2000] Back to top.

Spatial Normalization

When you coregister a pair of images, and want to carry an additional image along, then you can specify the additional image as other. In the example below, there were two ways of doing the coregistration:

1. Target: Image 1 Object: Image 2 Other: Image 3 Changes the .mat file of image 2 so that it matches image 1, and applies the same change to image 3.

2. Target: Image 2 Object: Image 1 Other: none Updates the mat file of image 1 so that it matches image 2. Image 2 is already in register with image 1.

Because the voxels of image 2 are aligned with those of image 3, then all you need to do is ensure that image 3 has the same .mat file as image 2. In this case, you can do this by simply copying the .mat file of image 2 to that of image 3. Use the button to make sure everything is OK. [John Ashburner 3 Aug 2000]

If your objective is ultimately to superimpose your functional activations on to your high resolution structural image, then the route I would take would be:

1. Without spatial normalisation: Coregister the structural image to the mean functional. Do the stats on the realigned functional images. Simply display the activations on the structural image (not the resliced version).

This works because the .mat file that is written for the structural image encodes the relative positioning of the structural image relative to the mean functional image. You can check that this has worked with the button, and selecting the unresliced high res structural and the mean (or any of the individual functional images).

2. With spatial normalisation: Coregister the structural image to the mean functional. Estimate spatial normalisation parameters either from the mean functional or the structural image. Apply the spatial normalisation parameters to the functional images to get spatially normalised functional images with whatever resolution you like. Hit the button, select spatial normalisation, then the option for changing the defaults on how the images are written, and specify something like 1x1x1mm resolution. Write the spatially normalised structural image using the new default voxel sizes (selecting the original structural image, rather than the resliced one).

The .mat files created by the coregistration (or realignment for that matter) are incorporated into the affine part of the spatial normalisation procedure. Processing the structural image in this way means that it is not resampled down to the resolution of the functional images at any stage. [John Ashburner 4 Jul 2000] | 1. in the case of PET/SPECT, the "reference image" is the mean image | calculated on the first pass For PET/SPECT realignment, the reference image is the first of the series during the first pass, and the mean of the realigned series for the second pass.

| 2. the default setting is NOT to weight this image This is true.

| 3. one cannot choose another image, since this is data-derived, reflecting | variance -StdDev, really-- at each voxel When weighting the realignment, the user normally specifies their own image. The weighting is not actually derived from the residual variance, although I did think about incorporating this in the realingment model.

| | in my case, I am aligning 2 or 3 images per subject, so I conclude this | option would not help me, as the voxel varainces would not be a useful | index of anything. Do you agree? I don't think you could obtain a useful variance image from 2 or 3 images, although I guess that some optimal variance smoothing could be used in principle.

| | >on the estimated parameters. The weighting procedure uses an image | >which can contain zeros and ones, or more properly, it can be thought of | >as containing the reciprocals of the standard deviation at each voxel | >(unlike weighting in the spatial normalisation where the weight is | >the reciprocal of the variance - I must fix this). | | when you say "fix this", are you referring to Realign or Spatial? fix to | which, var or SD? why ? I was referring to making the weighting of the spatial normalisation consistent with the weighting for the realignment. Currently one uses images that would be proportional to 1/variance whereas the other uses weighting images proportional to 1/sqrt(variance). However, normally the weights are either zero or one, in which case the distinction does not make any difference. [John Ashburner 21 July 2000]

| is it a requirement that all data from multiple subjects hav the same voxel | sizes when running spm_sn3d? They don't all need to have the same voxel sizes. [John Ashburner 26 July 2000]

Is it possible to retrieve SPM-non-linear-transformation information computed after spatial normalization with the T1-template in order to apply it to an other volume in the same coordinate system ?

The record of all the spatial transforms (linear and non-linear) are kept in the *sn3d.mat file produced by the normalization process carried out on your object image. This can be applied to any image that starts off in the same space as the original object image (normalize > write normalized only > subjects (1, presumably) > select the *sn3d.mat file from the smpget window > select image to be transformed).

If you just want the non-linear transforms only, that is possible using commands in the matlab window, but you would have to try someone else who knows a bit more about this. [Alex Leff 27 Jun 2000]

| with 128x128/30 Slices and voxel size 1.95x1.95x4.5 is | it worth to do sinc normalisation or is bilinear quite | enough? Sinc interpolation is generally recommended for interpolation when doing movement correction. If you reslice the images at the realignmnt step, then I don't think you gain much by using sinc interpolation at the normalisation stage. However, if you just estimate the movement parameters at the realignment stage, then these movements are incorporated into the spatial transformations at the normalisation stage, so it is probably better to use sinc interpolation. [John Ashburner 17 July 2000] If its functional data I would have thought bilinear is 'quite enough' [Karl Friston 17 July 2000]

| Similarly, I noticed that the defaults non linear basis function parameters | are 7x8x7. It was set at 4x5x4 in SPM 96. I've read in the spm archives that | 7x8x7 is suitable for T1 MRI images but not for PET images. Shall I go back | to the 4x5x4 value ?

More basis functions generally works better than having fewer when the images can be easily matched to the template. It is sometimes better to use fewer basis functions if the brain images contain lesions, or if the image contrast differs slightly from that of the templates. Alternatively, the amount of regularisation can be varied in order to modify the amount of allowable warping. The main reason for the extra basis functions used in SPM99, is that spatial normalisation in SPM99 tends to be much more stable than the version in SPM96. [John Ashburner, 12 July 2000]

| Can you telle me please the meaning of X ,Y, Z, when i use the mutual | information registration method? | I have used a PET cardiac images and I want to compare the angles that | I've imposed with the results of SPM.

The matrix multiplication displayed after running mutual information coregistration, shows a mapping from voxels in the stationary image, to those in the image that was rotated and translated. This mapping is derived from a series of rotations and translations (and zooms where voxel sizes differ between the images). The X, Y and Z refer to voxel co-ordinates in the stationary image, whereas X1, Y1 and Z1 refer to co-ordinates in the other one. The matrix can be decomposed into a series of translations, rotations, zooms and shears using the spm_imatrix function. For example, if the display says:

X1 = 0.9998X - 0.0175Y + 0Z - 10.0000 Y1 = 0.0174X + 0.9997Y + 0.0175Z + 0 Z1 = -0.0003X - 0.0174Y + 0.9998*Z + 10.0000

Then typing: M = [0.9998 -0.0175 0 -10.0000 0.0174 0.9997 0.0175 0 -0.0003 -0.0174 0.9998 10.0000 0 0 0 1.0000]; spm_imatrix(M)

should produce: ans =

Columns 1 through 7

-10.0000 0 10.0000 0.0175 0 -0.0175 1.0000

Columns 8 through 12

1.0000 1.0000 0.0000 0 0

which describes translations in the x and z directions of -10 and 10 voxels and rotations about the x and z axes (pitch and yaw) of 1 and -1 degrees (0.0175 and -0.0175 radians). The parameters are probably better explained if you type: help spm_matrix [John Ashburner 25 July 2000] |if I do normalisation on anatomical and functional |images, shall I coregister both modalities first or not?

If you want to use the anatomical image to overlay the functional activations, then one way would be to coregister the T1-weighted anatomical to the T2*-weighted EPI; normalise the EPI to the EPI template; then normalise the T1 image using the same parameters.

If you are normalising the T1-weighted and T2*-weighted images separately, to separate templates, then prior coregistration is irrelevant as the normalisation parameters will be determined independently. But this strategy might not be best, unless you have a specific reason. [Geraint Rees 15 July 2000]

| I am trying to normalize the contrast images of individual subjects | to run a random effect analysis. I read some of the messages on the | list about this and the biggest concern it seems to be that with the | normalization procedure one could lose or get strange results on the | outer rim of brain, because of the interpolation of NaN. I tried both | bilinear and nearest neighbour interpolation and, just eyeballing, | couldn't notice any strange border patterns coming up in both cases. | Would you think anyway that with nearest neighbour interpolation | this border effect is negligible (how does NN interp. treat NaN?)? The edge effects involve possibly losing some voxels. For tri-linear interpolation, if any of the 8 closest voxels are NaN, then the interpolated voxel is set to NaN. For nearest neighbour interpolation, if the nearest voxel is NaN, then the voxel is output as NaN. [John Ashburner 2 Aug 2000]

Back to outline.

Talairach Coordinates

Was wondering if there are any new methods for converting "Talairach" coordinates obtained using spm96 spatial normalization with MNI single brain template to the canonical Talairach's atlas coordinates. I used the transform from this template to the template from spm95 (PET) that Andreas posted but it is not perfect since the template is not matched to Talairach. Since the Talairach Daemon contains a scanned version of the atlas has anybody warped this to the MNI template?

The page:

http://www.mrc-cbu.cam.ac.uk/Imaging/mnispace.html

has a routine on it which does a reasonable job of MNI to Talairach atlas conversion. [Matthew Brett 16 Feb 2000]

I notice that the documentation describes the template images supplied with SPM as, approximate to the space described in the atlas of Talairach and Tournoux (1988). I was just wondering how 'approximate' they are and whether or not the co-ordinates provided by SPM should be reported as being 'Talairach'. For example, we are interested in the fusiform gyrus and it seems to me that the co-ordinates provided by SPM are about 2-4mm inferior compared to the Talairach atlas.

Matthew Brett has posted a helpful discussion of MNI/Talairach differences at http://www.mrc-cbu.cam.ac.uk/Imaging/mnispace.html that I think will answer your queries. The bottom line in terms of reporting is to reference activation loci to the surface anatomy of that individual subject if in doubt. [Geraint Reese 15 Aug 2000]

Back to Outline

Smoothing

Wouldn't it be a good idea to move the smoothing procedure from the "spatial preprocessing" to a step just before the result section. The idea is, that since smoothing (convolution with a kernel) and beta estimation (projection) are both linear operations they commute, and thus the order in which they are applied does not matter. So why not smooth a few beta images and the ResMS.img instead of hundreds raw-images.

I forgot to say that switching the order of smoothing and estimation is a problem in first level analyses because the RPV.img is calculated in the "estimate" section and not in the "results" section. So how about moving the two steps: smoothing and resel per voxel estimation, to the results section?

You are absolutely right about the commutative nature of the estimation and convolution operators yK = XBK + eK ....smooth before B*K = pinv(X)yK

y = X*B + e B = pinv(X)y BK = pinv(X)yK ....smooth after However nonlinearities enter with error vaiance estimation. One would have to smooth the residual images and then compute the sum of squares. This part is not commutative and the number of residual images equals the number of raw images. i.e.diag(K'*e'eK) ~= K'*diag(e'*e) Taking the sum of squared smoothed residuals is not the same as smoothing the sum of sqaured residuals.

The reason one can smooth after a 1st level analysis is because the sum of squared residuals do not enter into second level. [Karl Friston]

One caveat which you may know if you've followed the list, is that the con**.img and beta*.img images have the areas outside the brain values set to NaN. Thus any smoothing operation at this stage will lose the outer layer of voxels. Methods of dealing with that have been discussed previously by John and Russell Poldrack I believe. [Darren Gitelman]

Back to top.

Covariates

I am conducting a research on the response of PTSD and non PTSD patient to the repetition of their traumatic event. We have 2 group one of PTSD patients and one of non PTSD patients. For each patient we have 3 baselines and 4 repetition of the traumatic story.

We want to check the difference in the reaction to the traumatic script between the PTSD and control group.

We use three type of covariates

constant response - 0 0 0 1 1 1 1 rise during the repetition - 0 0 0 1 2 3 4 decrease during the repetition- 0 0 0 3 2 1 0

These covariates are not orthogonal. If I want to check in what region of the brain the response correspond to each of the covariate how should I define my contrasts?

The repetition-dependent increases and decreases are modelling the same effect and you only need to specify one regressor:

main effect of traumatic script - -1 -1 -1 1 1 1 1 script x [linear] time interaction - 0 0 0 -3 -1 1 3

Note that these regressors are orthogonal (and orthogonal to the constant term) and can be tested with contrasts [1 0], [-1 0], [0 1] and [0 -1]. The last two give you increases and decreases respectively. [Karl Friston 13 Mar 2000]

Alternatively would one use 2 covariates ? Each covariate consisting of the mean centered values for a single group with 0 padding for the subjects in the other group, essentially what you get when you specify a group x covariate interaction The effect of the covariate for group 1 alone would be tested with the contrast 0 0 0 0 1 0 ( (Group 1, Group 2, Condition 1, Condition 2, Covariate1, Covariate 2) and the interaction across groups tested with the contrast 0 0 0 0 1 -1 ?

Would the contrast 0 0 0 0 1 1 in the multi-group model be a test of the effect of the covariate collapsed across all subjects disregarding group membership ?

Yes indeed - it would be the main effect of the covariate. [Karl Friston 13 Mar 2000]

Our data is as follows: 11 subjects and 8 scans per subject, which makes 88 scans. At each scan, each subject gave a subjective rating (0 -10). We are trying to do a correlation analysis between subjective ratings and actual voxel values. The problem is that our data is not independent i.e. does not follow the assumption for correlation (regression analysis). Is it possible to analyze this kind of data using correlation analysis? Anyway, we tried the following model: Multi-subject : Covariates only, Covariate interaction with subject. Is this correct?

Yes - this models the main effect of the covariate and the covariate x subject interactions (i.e. computes a regression coefficient for each subject). You can test for the significance of the average regression (or indeed differences among the subjects) with the approproate contrast. [Karl Friston 11 Jul 2000]

I have a few more questions concerning the multisubject correlation analysis. Does "Covariates only: interaction with subject" model take into account the dependence within subject?

I am not sure what 'dependence within subject' means. This design fits a different regression slope for each subject and assumes the error within subject is independently and identically distributed. [Karl Friston 12 Jul 2000]

Given 2 conditions, 1 scan/condition, 1 covariate obtained at each scan, mean-centered covariate with proportional global scaling. A condition & covariate design with a contrast 0 0 1 is equivalent to correlation between the change in covariate and the change in the scans.

Indeed or more precisely the partial correlation between the covariate and scan-by-scan changes having accounted for the condition-specific activations.

Can this approach be generalized to a multi-group design ? If there are 2 groups and 2 conditions, with 1 scan for each subject under each condition, and a single covariate collected during each scan, then would one specify 1 or 2 covariates.

I would specify two, each with a group-centered covariate. This would allow you to look for differences in the partial correlation with the contrast [0 0 0 0 1 -1]. i.e. models group x covariate interactions.

If only one covariate is specified, would one test for a covariate effect in group 1 alone with the contrast 1 0 0 0 1 (Group 1, Group 2, Condition 1, Condition 2, Covariate), and the group x covariate interaction with the contrast 1 -1 0 0 1 ?

No. The interaction is not modeled with only one covariate. This contrast is simply the main effect of group plus the main effect of contrast.

Alternatively would one use 2 covariates ? Each covariate consisting of the mean centered values for a single group with 0 padding for the subjects in the other group, essentially what you get when you specify a group x covariate interation The effect of the covariate for group 1 alone would be tested with the contrast 0 0 0 0 1 0 ( (Group 1, Group 2, Condition 1, Condition 2, Covariate1, Covariate 2) and the interaction across groups tested with the contrast 0 0 0 0 1 -1 ? Exactly. [Karl Friston 28 June 2000] I have a very similar question to that posted by Steven Grant on June 9th. I have two groups of subjects who, on prescreening, exhibited differential mood responses (one group positive scores, the other negative) to a drug. So with 2 groups (positive and negative responders), 2 conditions (drug and placebo), 1 scan/subject under each condition, and 1 covariate (mood score) collected/scan, how would I go about determing whether rCMglu is affected by:

1. prescreening mood scores both within and across groups These are the simple and main effects of mood (or group) and would be best addressed with a second-level analysis using the subject-effect parameter estimate images (i.e. averaging over condotions for each subject).

2. post drug scan mood scores both within and across groups and also whether: This is a main effect of 'post scan mood score' within the post drug level of the condition effect. This is best analysed using a condition-centered covariate ('post scan mood score') and testing for a significant regression with 'drug'.

3. mean centering of the covariate is useful/necessary in these cases Yes. For (1) there is not covariate required.

4. mean should be computed within group or across all subjects For (2) within condition. There are other centering you could use to look at different interactions. e.g. group-centered would allow you to see if there was any interaction between prescreening and post scan mood. [Karl Friston 28 June 2000] When one performs a multisession fMRI analysis with SPM, in either its SPM96 or SPM99 incarnation, my guess is that SPM removes intersession effects, perhaps behind the scenes. If this is so, if each session were a different subject, and if one tried to use the subjects' age (rounded to the nearest year) as a confound in the analysis, would this result in a problem of linear dependence in the columns of the design matrix?

Yes and no. It would result in a linear dependence in the design matrix (there will be a linear combination of "session-regressors" that equals your "age-regressor"). It will not be a problem as long as all you want to do is to use it as a confound (since SPM uses the generalised inverse), but then again it would not do you any good either since the space of the confounds is identical in both cases. You cannot use it as a covariate of interest since it is exactly in your confound space.

In short, the confounds already modelled removes any age effects, and more. [Jesper Anderson]

Dear Jesper: May be I am misunderstanding something here, but I am still slightly puzzled at the apparent impossibility to enter age or IQ as covariates in an SPM fMR analysis. I assume it would be possible to trick the analysis into doing it by treating all data from one group as a single session and entering the age covariate for the data from each respective subject. But then between-session variance would probably smother any task-related effects.

You are right on the first count, one can enter data as if they were from a single session and "trick" SPM into looking at age effects. I very strongly suspect you are right on the second count as well, inter-session variance originating from other sources of variance would dominate over any "true" age effects. The question is related to one by Kris Boksman a few days ago.

Would all this mean that I can only look for age or IQ effects post hoc, e.g. looking at Z scores or mean signal change within ROIs etc.?

I don't really think it is meaningful in any way to look at main effect of age using T2* weighted data. You could look at main effect of age using a morphological technique (e.g. using T1 weighted and Voxel based morphometry) or you could use a quantitative technique for measuring perfusion (i.e. PET or perusion MRI).

With T2* data you may look at task-by-age interactions, i..e. how age affects the response to a given stimuli. To do this you would generate the appropriate contrast for each subject (say you are interested in how the difference between conditions 1 and 3 changes with age, then you would enter the contrast [-1 0 1 ...] for each subject. In "Basic models" pick "simple regression (correlation)" and enter the resulting con*.img images as input images and enter the age of each subject as your covariate. This will constitute a Random effects model and will allow you to make proper population inferences.

From your suggestion above (i.e. to look at z-scores post-hoc) I suspect it may really have been a condition-by-age interaction you were after in the first place. Note though that by using z-scores (rather than the linear combination of parameter estimates offered by the con*.img images) you would be assessing reliability of activation across subjects rather than magnitude. That is slightly different and more akin towards a meta analysis. [Jesper Anderson, 12 July 2000]]

I'm analysing some PET data, and am interested in looking at regions where activity correlates with RT, and was wondering what was the best way to proceed. We have 6 subjects, 12 scans per subject When I choose the multi-subject covariates only design, there appear to be two ways to proceed. One is to not have covariate * subject interactions, to mean centre the RTs, and then to make the following contrasts: 1 to look at regions where activity correlates with increasing RT and -1 the converse. The second is to select covariate by subject interactions, and then to have the following contrasts: 1 1 1 1 1 1 and -1 -1 -1 -1 -1 -1. Could someone help explain what the differences are between these two analyses?

If you don't model the subject by covariate interaction, you assume that there is a common slope of the covariate over all subjects. This saves you some degrees of freedom, but essentially you're assuming that the slope of these covariates is the same for each subject, if there is some component in the observations, which can be explained by your covariate.

If you do model subject by covariate interaction, you don't make this assumption of the same slope for each subject, but allow for fitting a different slope for each subject. Note that this model can also be used to generate subject specific contrast-images, which you can use as input to a second level analysis. [Stefan Kiebel 3 Aug 2000]

Back to Outline.

Contrasts

Contrasts: F-contrasts explained

Thank for you help. I have done what you suggested, and now, I'm stuck at the contrast stage.

and use F-contrasts to test for mean effect and differences in the epoch-related responses.

I have four conditions and the design matrix now has 8 columns, 2 for each condition. I'm not used to looking at F maps. Usually, if I was comparing t-maps, I would enter a contrast 1 -1 0 0 if I wanted to compare condition A with B. Do I now enter: 1 1 -1 -1 0 0 0 0 or do I look at the individual activations 1 1 0 0 0 0 0 0 and 0 0 1 1 0 0 0 0 and then use making to look at commonalities/differences?

First use the F-contrasts computed by default in the results section - contrast manger, under F-contrasts. These will give you the condition specific F-contrasts. For example the responses due to condition 1 should look like:

1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0

To look for the differences between condition 1 and 2 use something like:

1 0 -1 0 0 0 0 0 0 1 0 -1 0 0 0 0

If you have used 'mean and decay' than you can also try the conventional T-contrasts looking for differential respones in terms of mean or decay. For the 'mean' it is simply:

1 0 -1 0 0 0 0 0

An F-contrast can be though of as a collection of T-contrasts, that you wish to test jointly. [Karl Friston 28 Jun 2000]

Contrasts: Time x Condition

I'm still struggling a little with understanding how contrasts need to be specified in spm99... The timecourse does not always look like I expect it to! I have the following blocked fmri design: control task control task control task If a set it up as one condition and use the contrast [1] then I see areas that are activated with respect to the control. The contrast [-1] reveals areas that are deactivated with respect to control. However, I'm interested in the task vs. control (activation) x time interaction. So I specified the design as 3 conditions. I thought that I could mask [1 1 1] (task activated wrt control) with [1 2 3] (increasing trend over time) or [3 2 1] (decreasing trend over time). However, the [1 1 1] contrast appears to show regions that are both activated and deactivated for task wrt control. Can anyone explain this?

If I understand correctly, you are primarily interested in time-by-condition interactions, where you have a control and a task condition that alternate over the time-series.

I would set up the design matrix in the following way.

Your design matrix should have four othogonal covariates:

1. Control condition (control condition not modulated over time)
2. Task condition (task condition not modulated over time)
3. Control condition x time (control modulated by time)
4. Task condition x time (task modulated by time)

(iii) and (iv) are simply (i) and (ii) multiplied by a trend (linear or non-linear - see below). SPM99 implements this in the following way. If I remember correctly, after you have entered the details of your variables, it will ask about parametric specifications. The choices offered are "none", "time" or "other". Your choice should be "time". The next choice will relate to the nature of the expansion ("linear", "polynomial" or "exponential"). I would choose "linear"as a first pass. You will then be asked which trial types to apply the expansion to. Select both your trial types.

If you choose to modulate both your conditions, there will be four covariates of interest in your design matrix (first four columns). Your contrasts will be:

1. Task > control (not over time) -1 1 0 0
2. Control > Task (not over time) 1 -1 0 0
3. Task > control (interaction with time) 0 0 -1 1
4. Control > task (interaction with time) 0 0 1 -1

This provides a flexible design matrix in which you may wish to compare (iv) either with (iii) (control and task both change over time). Alternatively, if there is good reason to believe that the control condition itself remains unchanging, the time-by-condition interaction may be found by -1 0 1 0 and 1 0 -1 0 (compare (i) with (iii)). [Narender Ramnani 14 Aug 2000]

However, I'm interested in the task vs. control (activation) x time interaction. So I specified the design as 3 conditions.

actually what I specified was block one (first presentation of control, task) as condition 1, block two (second presentation of the same control, task) as condition 2, and block three (third presentation) as condition 3.

Thanks. Sorry not to have worked this out first time around. Now I think I see the idea behind of your contrasts. The contrast 1 2 3 was supposed to pick out voxels which show an increase in task-specific activity over time. In fact, this contrast asks a completely different question (not one that you are interested in), i.e. is the sum of the parameter estimate for the first column, plus twice the parameter estimate for the second column, plus three times the parameter estimate for the third column, significantly different from zero. A voxel which shows, for example, exactly the same task-related activity in all three blocks (i.e. no condition by time interaction) will show up in this contrast.

The contrast 1 0 -1, however, is beginning to get towards what you are after. This will pick out voxels which show significantly more task-related activity in block 1 than in block 3, suggesting a time-dependent decrease. Similarly, contrast -1 0 1 will show voxels which show significantly more task-related activity in condition 3 than condition 1, suggesting a time-dependent increase.

However, if I specify it as one condition, use the parametric modulation with time, and then use the contrast 0 1 0, will this tell me the task x time interaction across all of the scans? (ie rather than blockwise as I tried to specify it above)? does this tell me the interaction between areas activated (task wrt control) that increase with time?

Yes, it would tell you the task x time interaction across all of the scans. This new model is more tightly constrained, in that you have to specify the shape of the parametric modulation with time. I would guess that you are happy with a linear model, but SPM also offers you an exponential model and one other model too (I can't quite remember what).

If a voxel shows a net increase in task-related activity over time, then in the new model it should show up with the contrast 0 1 0. The voxels which will be best fitted by the model are those in which the increase over time is linear (e.g. one in which the task-related activity is 1 in the first block, 2 in the second block and 3 in the third block). However, even a voxel in which the relationship with time is more complicated will be fitted to some extent, provided there is an increase overall during the experiment (i.e. there is a linear component to the relationship).

and 0 -1 0 would tell me areas of activation that decrease with time? these are the two contrasts I'm interested in.

You have to be careful with use of language here. 0 -1 0 does not tell you 'areas of activation that decrease with time', in the sense that it would also include 'areas of deactivation which increase with time'. The contrast shows the voxels whose task-related BOLD becomes more negative during the experiment. The question of whether, over the whole experiment, there is a net positive task-related signal is an orthogonal question.

If by 'areas of activation that decrease with time' you mean voxels which satisfy the following 2 criteria ...

1. the signal during all of the task epochs is greater than the signal during all of the control epochs (taking the whole experiment together (to give you 'areas of activation...'); and
2. the task-related signal becomes more negative during the course of the experiment (to give you '...that decrease with time') ...then you should probably mask the contrast 0 -1 0 with the contrast 1 0 0

Similarly, if you wanted to find 'areas of activation that increase with time' then you might mask the contrast 0 1 0 with the contrast 1 0 0. [Richard Perry 14 Aug 2000]

Dear Narender, Richard- aha yes!! I was using F contrasts rather than t. Suddenly these contrast maps look much more like what I expected them to... How do the F and t contrasts differ? (I can also look this one up on the archives..)

There's a whole lot of stuff to F contrasts and how to use them (only some of which I understand). But briefly, in a situation like yours the F contrast is the t contrast squared. Hence it doesn't matter whether you use +1 or -1, the result is the same.

In general, F contrasts test whether the covariates that you specify contribute significantly to modelling the variance, regardless of the sign of the parameter estimate. Your contrast 0 1 (or 0 -1) was pulling out all of the voxels within which the second covariate explained a significant amount of the variance in the data. [Richard Perry 15 Aug 2000] Contrasts: A > (B & C)

What would be the apropriate Contrast for three conditions A B C that tests for A > B & C ? The contrast [2 -1 -1] tests if the activity in A is larger than the mean activity in B and C. The conjunction between contasts [1 -1 0] and [1 0 -1] tests if the activity in A is larger than the activity in B, AND is larger than the activity in C. I suspect the latter makes a little more sense. [Jesper Andersson 10 Jul 2000] I have a simple contrast question. I have a paradigm wiht 3 conditions (a, b, ab). If I set a contrast of -1 -1 1 does this show me the areas where ab>a+b? If not what does this contrast represent? The contrast [-1 -1 2] would show you, where ab>a+b, i.e. where ab has a larger effect than the averaged effect due to a and b. [Stefen Kiebel, 14 July 2000] I'm not sure if you are working with PET or fMRI. Certainly with PET paradigms and most fMRI paradigms, contrasts across conditions within a group, need to total zero; i.e. they need to be balanced. It appears from what you have written that you have three conditions a, b, and ab. If so you need to specify contrasts across these conditions that add up to zero. So to test for areas that are more active in b than a, you would need to enter: -1 1 0. For areas more active in ab than a AND b: -1 -1 2; which I think answers the latter part of your query. [Alexander Leff, 16 July 2000] Conjunctions

Conjunctions: General

I have a simple question (I hope). I have completed a multi-subject (6) multi-condition (3) fmri study. I have completed individual analysis and am working with a fixed effect group analysis because my degrees of freedom are too small. I am interested in the best method to determine the area(s) that are activated in the study population across all three conditions. In other words, what are the common areas activated across the three conditions in this study. If anyone has suggestions I would be grateful.

A conjunction analysis would be appropriate but would necessitate each active condotion being referred to its own control, to ensure three orthogonal contrasts. Conjunctions are specified by holding down the 'control' key during contrast selection. [Karl]

I have a few questions concerning conjunction analyses. I have a study with 9 subjects. I would like to do a conjunction analysis but have several questions.

1. Should corrected or uncorreted p values be used for the analysis. The mailbase continually refers to uncorrected p values.

In SPM99 a conjunction SPM comprises the minimium t values of the component SPMs. These minimum t values have their own distributional approximation which allows one to compute both corrected and uncorrected p values, just like ordinary SPMs. The criteria for using corrected or uncorrected inference is exactly the same as for any other SPM.

2. In relation to question 1, should a height threshold be used. I have worked with SPM and realize that extent thresholds cannot be used with conjunction analyses.

The distributional approximations for the spatial extent of a conjunction SPM are not known (at present) and therefor inference based on spatial extent is precluded. Consequently height is currently used to specify thresholding. This can be corrected or uncorrected and both pertin to the final significance of the conjunction SPM (Pconj) (not the components - in SPM99b the uncorrected height threshold refered to the components).

3. What p value is most appropriate? A recent mail from Karl stated that between .5 and .05 is most approptiate. At what level (i.e. simple threshold or height threshold) is this entered? Also, could you explain why such a p value is appropriate.

Thresholds are entered in the results section after specifying which contrasts are to enter into the conjunction (by holding down the control key). The recommendations above probably refered to uncorrected p values for the component SPMs (Pcomp) (in SPM99b). In SPM99 a corrected p value of 0.05 or an uncorrected p value of 0.001 would be sensible. These might correspond to 0.5 or even more from the perspective of the component SPMs. Note that for uncorrected p values Pconj = Pcomp^n, where n = number of [orthogonalized] contrasts. [Karl Friston]

We're piloting a motor learning paradigm, in which subjects are scanned while performing a motor task A and a motor task B, then are trained for a week on A alone, then are scanned while doing both tasks again. What we are looking for ideally is 1) strength of signal--tose voxels which show a different activation for A than B in the second scan, but not in the first scan; and 2) extent of activation--a larger or smaller area for A than B only in the second scan, but not in the first scan. We have three subjects, and will only run more if we think the results look promising, since this is a pilot study.

Given the preliminary nature of the study, is it a fair assessment to compare cluster sizes for each subject individually for A before and after training, and for B the same way? Or is it better to put all three subjects in a single design matrix and look for conjunctions? Or is some other approach even better?

The conventional approach to this problem would be to create a SPM of the effect of interest, namely the condition x session interaction. This obtains from putting all your data into a single, session- separable model and testing for (A1 - B1) > (A2 - B2) or (A1 - B1) < (A2 - B2) interactions with the appropriate contrast. This approach controls for non-specific time effects and should be practice-specific.

The question about differential areal activation is implicitly answered in the interaction SPM (i.e. if the area contracts the penumbra will show a negative interaction and if it expands it will show a positive one).

A more sensitive analysis obtains if you use a conjunction of the interaction and main effects: i.e. test for a conjunction of the two hypotheses: This motor responsive area (Hypothesis 1 = main effect of A vs B) shows learning-dependent adaptation (Hypothesis 2 = interaction). [Karl Friston 17 Apr 2000]

I have a question regarding how to determine p-value related to spatial extension (voxel number) in a conjunction analysis using spm99. I noticed that spm99 does not provide any p-value for spatial extension for conjunction analysis, in contrast to simple contrasts. My understanding is that statistical assumptions underlying computations of those p-values are no more valid in conjunctions of contrasts. Is it right? If so, how can we decide significant spatial overlap between several contrasts. Any insights, help or references adressing this question would be greatly appreciated. Thank you.

The preliminary answer is that spatial extent values are not provided because they have not been worked out for conjunction analyses. It also sounds like what your trying to work out is how to do a conjunction on the differences between groups. i.e. how does one decide whether activations seen on a conjunction analysis are significantly different between groups.

As far as I know this can't be done in spm. Conjunctions basically tell one about main effects. For the interactions one has to do either fixed effects or random effects-type analyses. This of course leads to difficulties sometimes interpreting results because the interactions are not being looked at in the same way as the conjunction. Perhaps something is possible with masking, but I can't see that this would let you reject the null hypothesis. Hopefully there will be a more expert addendum. [Darren Gittelman]

I have two groups of subjects (actually they are the same subjects, but this is two distinct fmri sessions) and I wanted to know which regions are involved in the same contrast in both groups. A conjunction analysis seems appropriate (the two contrasts are orthogonal). My problem is that as a result I got clusters including e.g. 1, 2, 10 or 50 voxels. Which extend threshold to use to select regions with significant joint activations? (in the example, 1 and 2 voxels appear too small, 50 voxels significant, and 10 voxels in between). My first answer would be to use the same threshold as in single contrast (e.g. 15 voxels, i.e. p<.05), because the conjunction has already taken into account the joint test (at least on height threshold). However, it seems to be inapproriate in some instance: Consider for example that each single contrast activates a cluster including 20 voxels and the resulting conjunction analysis provides a cluster of 12 voxels within each single-contrast clusters. Using the same theshold (15 voxels) would reject this conjoint activation, but 12 voxels over 20 voxels which co-jointly activated seems to be quite significant. I was thinking that in this instance, we indeed use the implicit assumption that we are testing for joint activations within given clusters and a standard masking analysis might be more appropriate. But, the same spatial extend threshold problem appears to occur again, if I am right. What do you think?

The corrected p values are sensitive to the spatial extent threshold for single contrasts. In other words the p<0.05 corrected height threshold is lower if you specifiy an extent threhold of 8 voxels than if you use 0 voxels. For conjunctions you are forced to use 0 voxels because, as Darren points out, the theory does not exist for > 0. Therefore even a cluster with 1 voxel in a conjunction SPM is significant. Remember the conjunction SPM finds the overlap among a series of component regions. This overlap can be small but very significant. [Karl Friston]

How can I integrate conjunction analysis in a multi subject random effect model? Single subject conjunction of two contrasts is easy to perform, but what do I enter at the second level or how do I perform the conjunction analysis over a group of subjects? You do not (generally). A conjunction analysis at the first level, over subjects, addresses the same thing as a single contrast at the second level (using the subject-specific contrast images). In some instances you may want to do a conjunction of two contrasts pertaining to different effects at the second level. This requires the two or more [orthogonal] contrasts to be entered into the second-level model with a simple conjunction (e.g. [1 0] and [0 1]). However, you are making strong assumptions about the sphericity of the error terms, at the second level, by doing this. These assumptions might easily be violated (for example subjects who activate in contrast 1 may be more likely to activate in contrast 2). If you are obliged to enter more than one contrast per subject into a second-level analysis then you should qualify your inferences along these lines. [Karl Friston 4 Aug 2000] Conjunctions: With SVC

Following up on the issue of conjunctions, in looking at a paper by Keith Worsley and yourself (A test for a conjunction), it suggested [to me] the possibility that a conjunction could be taken over a region of interest as well as at the voxel level. Would this be possible, that is- designating a region of interest, and performing a conjunction analysis which would test the null hypothesis that all subjects did not have an activation in a certain area? Then the localization would then pertain to the area and not to a particular voxel.

Unfortunately the maths of the paper escape me, so if this is possible I'm not sure how to implement this over an area as opposed to a voxel. Any help appreciated.

By using the corrected p value one can test the null hypothesis that one or more subjects did not activate within the search volume (to which the correction applies). By using a small volume correction within a conjunction SPM one can restrict the inference to a VOI.

I am not sure that this is what you had in mind but it represents a useful combination of SVC and conjunctions. [Karl Friston 5 Apr 2000]

Conjunctions: Differences between SPM96, SPM99

After huge digs in the SPM archives I still didn't find the answers to some of my questions...Could you please reexplain to me:

1. What are the theoretical and practical differences of conjunction analyse between spm96 and 99?

Conjunctions in SPM96 were based on getting a significant main effect (averaged over all the contrasts that entered into the conjunction) in the absense of any interactions among the contrasts. This was suboptimal because it relied on accepting the null hypothesis of 'no interactions'.

Conjunctions in SPM99 are tested by ensuring all the contrasts are jointly significant. The conjunction SPM would look similar to the SPM that obtains from exclusively masking all the contrasts with theselves to reveal the common areas or intersections. The t values are the smallest among all the component SPM(T) values and the conjunction SPM is a 'minimum T-field'. Gaussian field theory is now applied to this SPM to give corrected p values. We could not do this in SPM96 because the minimum T-field theory had not been devised. [Karl Friston 22 Mar 2000]

I am uncertain about the difference and interpretation of a conjunction analysis compared to masking of contrasts. For example, is inclusive masking of one contrast by another the same as a conjunction between the two contrasts ? When is it best to use conjunctions and when is it best to use masking ?

Jose' Ma. Maisog wrote:

I'd be interested in seeing some sort of answer to this question, too. My understanding from reading Cathy Price's paper (Price CJ & Friston KJ, Neuroimage 5:261-270 (1996)) is that conjunction analysis is not the same as masking one contrast by another, and that a conjunction analysis between two contrasts is done as follows:

(1) Include as a regressor in the GLM the interaction effect between the two contrasts.

(2) Threshold the Z map for the interaction effect, generating a mask.

(3) Use this mask to mask the Z maps for the two contrasts.

Is this right? Or, have there been new developments (e.g., the paper by Worsley & Friston) which render the above obsolete? Has conjunction analysis changed between SPM96 and SPM99?

Thanks,

Joe.

Yes Conjunctions have changed in SPM96 and SPM99. In SPM96 and SPM97, conjunctions involved calculating the main effect of two contrasts (A1-B1) + (A2-B2) and then subtracting areas that showed an interaction. In this context, conjunctions provided a Z score and the associated probability for the main effect where there were no interactions. However, there were a number of interpretation problems. First, the probability generated relates to the overall main effect rather than the liklihood that two events occur independently. Second, the conjunction relied on the sensitivity of finding an interaction. In other words the voxels identified were those where there was no significant interaction. The objection here is that you can not prove a null effect. Indeed, when we plot out some of the areas identified by conjunction analysis in SPM96 and SPM99 we find that the effect sometimes only comes from one contrast because the difference between contrasts doesn't reach significance.

The conjunction analysis in SPM99 overcomes these problems by using multiple masking and reporting the probability that relates to the co-occurrence of two or more effects in the same voxel. Probability decreases as the number of contrasts in the conjunction increases. The masking option is still included because it allows you to (i) specify a mask that is not included in the conjunction; and (ii) specify an independent threshold for the mask; (iii) Mask with contrasts which are not orthoganol to those in the conjunction. For example, the conjunction (A1-B1) + (A2-B2) could be masked with A1- Rest and A2 - Rest. Including these masks might be useful if it is necessary to differentiate increases in A from decreases in B. [Cathy Price 21 July 2000]

Conjunctions: Orthogonal Contrasts

I want to run an analysis that I have done by you just to make sure that my logic is sound.

I have an fmri study with multiple stimulus conditions in each series. I am interested in two questions. The first is to locate the areas of activation that overlap across stimulus conditions. I have performed conjunction analyses at the individual and group level (fixed effects). My conjunction included the positive contrasts for both stimulus conditions.

My second question is: Do areas of activation from one stimulus overlap with inactivations from the other stimulus. To do this I performed conjunctions analyses (again at the individual and group level) and selected the positive contrast from one condition and the negative contrast from the other condition. In other words the did a conjunction of the contrasts 1 0 and 0 -1.

Are these analyses valid? They seem to follow what has been addressed on the mailbase but I am quite surprised by my results and want to verify that I have not made some stupid misinterpretation.

Your approach was quite reasonable, except that you have made the assumption that the contrasts 1 0 and 0 -1 are orthogonal, and I guess that this is probably not the case. (After all, when you have stimulus A, presumably you can't also have stimulus B at the same time?) If they are not orthogonal, then the conjunction which you have tried is not really interpretable.

To take an over simplistic example, imagine a situation in which there were only two conditions, A and B, and these were both modelled with box cars, and in fact the covariate for B was equal to that for A multiplied by -1 and with +1 added (i.e. when A had ones, B had zeros and when A had zeros, B had ones). In this overspecified model, the same variance in voxels which were actually 'activated' by condition A but not B could either be modelled using covariate A (with a positive parameter estimate) or with covariate B (with a negative parameter estimate). Many of these voxels would end up being modelled by some combination of the two, and these would show up in the both contrasts, 1 0 and 0 -1, and would therefore also show up in the conjunction of these two, in spite of the fact that in this example there is no response at all to B.

Really to answer your question you need to have more conditions. Ideally you should have condition A and its own baseline condition, and condition B with its own baseline condition (you can't use the same baseline for two conditions which you want to use for a conjunction analysis). With these four covariates you could do the conjunction of 1 -1 0 0 and 0 0 -1 1, and get a meaningful answer. I guess that what you have ended up with, in your conjunction of 1 0 and 0 -1, are many more voxels than you expected. These are not necessarily voxels in which 'areas of activation from one stimulus overlap with inactivations from the other stimulus'; it may just be telling you that your covariates are significantly co-linear. [Richard Perry 8 Aug 2000]

Richard thanks for your reply. I suspected that there may be a problem with non-orthogonal contrasts but I did not fully understand. What is SPM doing when it "orthogonalizes" the contrasts in a conjunction? Does this have any meaning and why does it not account for the non-orthogonal nature of the contrasts?

Sorry, I was forgetting that SPM99 tries to take account of this problem. I must admit that I haven't used SPM99 for this purpose, so I don't quite know how it works. I still don't think that it helps you, for the following reason.

As I understand it, when you have two covariates, then the variance modelled by these can be partitioned into three components:

1. variance which can only be explained by covariate A
2. variance which can be explained by either covariate,
3. variance which can only be explained by covariate B.

The way in which you specify your orthogonalization order (SPM99 prompts you for this after you have chosen your contrasts) will influence the parameter estimate for one or other covariate, and I have to confess that I cannot remember which way round it works, as I find it a bit confusing. I think that the first contrast which you specify is left unchanged (i.e. the same contrast is applied to the same parameter estimates), but the second contrast is modified to compensate for the fact that your parameter estimate is for components 2 and 3 rather than just component 3. Thus your parameter estimates stay the same, but you will see that the second (I think!) contrast now looks slightly different, and includes non-zero values even for some of the covariates which only appeared in the first contrast when they were originally specified.

However, regardless of the implementation, I think that the idea is that you are ascribing the variance which can be modelled by either covariate (component 2 above) to one or other. You don't actually know which one it comes from, and there is no way to find out. You could still be mislead in your situation. Thus, you might set things up so that the common variance (component 2) is explained by covariate B, when in reality it is entirely attributable to covariate

1. The remaining variance which can only be explained by covariate A (component 1) is appropriately modelled by this covariate. Once again you have a situation where variance which actually comes from one condition appears to be attributable to a combination of both, and so you have voxels showing up spuriously in your conjunction.

But I may be wrong about this. It may be that SPM99 discounts the common variance (component 2), so that the conjunction would now ask whether the data from a voxel includes both 'component 1' and 'component 3' variance. If there is considerable co-linearity between the contrasts, so that much of the variance is 'component 2', then this test would obviously be rather insensitive, but I think that the results might be meaningful even in your case. However, if this is what SPM99 does, then I wouldn't have thought that it would need to ask you for an orthogonalization order.

I hope that someone else will be able to give a more expert reply, and tell you which of these SPM99 actually does. Some of the real experts are away at the moment, though. If this question is important, though, I would seriously consider doing another experiment in which each condition has its own baseline, as described before! [Richard Perry 8 Aug 2000]

Richard is essentially correct. Conjunctions in SPM99 use minimum t-field theory, which requires that the t-maps are independent, which they approximate when the contrast spaces are orthogonal (and there are many dfs). Note that it is possible that even though contrast weights are orthogonal, eg [1 0] and [0 1], the subspaces defined by the contrasts are not (if there is correlation between the covariates for example).

Thus SPM99 may ask you for an orthogonalisation order, such as [1 2]. This will orthogonalise the contrast space for contrast 2 with respect to that for contrast 1. This will modify the second contrast to include the shared variance (that Richard talks about). (Rather confusingly, for the special case of two covariates, this is equivalent to a [0 1] contrast after orthogonalising the first covariate with respect to the second.) [Rik Hensen 16 Aug 2000]

Back to Outline.

Random Effects

Random Effects: General:

1. Is there a preferred method for doing the first level analysis: should I calculate the con*.img files for the second level analysis on an individual basis or should I use a multi-subject fixed-effects model, and then calculate the single subject con*.img files by presenting the other sessions (subjects) as null events (e.g. 2 conditions, subject 1: 1 -1 0 0 0 0.....; subject 2: 0 0 1 -1 0 0........ ; subject 3: 0 0 0 0 1 -1 0 0......etc).

In the latter case, the individual con*.img files are not independent, I would think.

In fact they are. Even when using AnCova to remove the confounding effects of global activity SPM is set up (if you use the defaults and model subject x condition interactions) to be subject seperable. This means that the contrast images are the same as if you had analyzed each subject seperately. As such doing a multi-subject first-level analysis is much more convenient (for selecting subject-specific contrasts, as you specifcy, for the second level).

2. Is it possible to do a RFX analysis with an unbalanced design, i.e. a random-presentation event-related design with a different model for each individual subject?

Yes, if the designs are sufficiently similar (i.e. roughly the same number of events etc.) the second-level analysis of an unbalanced design can be considered a mixed-effect analysis (with assumtions about sphericity). There has been some discussion about this in the archives. [Karl Friston 24 Mar 2000]

I have a few more questions concerning the multisubject correlation analysis. Does "Covariates only: interaction with subject" model take into account the dependence within subject? If it does, how does it do that? This is important because we have 8 scans per subject.

We specified the following contrasts (11 subjects) to test

1. positive correlation 1 1 1 1 1 1 1 1 1 1 1
2. negative correlation -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1. Are these t-contrasts correct for testing the average positive and negative correlation?

I wonder whether these results are fixed effects analyses, that cannot be generalized into population level.

You are correct that this is a fixed effects analysis, and cannot strictly speaking be generalised into a population level inference.

The way to perform a random effect analysis on this study is to generate a contrast parameter estimate map for each subject (by simply looking at contrasts [1 0 ... 0], [0 1 0 ... 0] ... [0 .. 0 1] in the results section). You then perform a one-sample t-test (i.e. compare the average of your parameter estimates with zero) on these (con_00*.img) maps. You do the same thing with the [-1 0 ... 0], [0 -1 0 ... 0] ... [0 .. 0 -1] contrasts to check the negative correlations.

The RFX model will tell you if all subject "activate" in the same location and with roughly the same magnitude. If you want to answer the slightly less stringent question "do they all activate in the same location?" you can use a conjunction across subjects instead. Its real easy. Once you have entered all the individual contrasts above you simply select them all (positive and negative separately) in the results section using the control button. [Jesper Andersson 12 July 2000]

when we do a 'multi-group conditions and covariates' analysis, we can not produce the subject specific contrast images, because the columns of the design matrix contain the different conditions for all subjects. How can we obtain the contrast images to enter into the second level analysis ? Do we have to use the results from the single subject analysis ? You could proceed by treating each subject as a separate group. If you are using covariates ensure you select 'covariate x subject interactions'. [Karl Friston 12 July 2000]

...or just use "Multi-subject: cond x subj interaction 7 covariates", and put all the subjects in together. This will ask you less questions, but effect the same result as Karl's suggestion.

Although you have two groups, you only want the individual subject level contrasts, which you then will assess at the second level. So, the group membership isn't important in the first level of the analysis since the model fits each subject separately.

As Karl notes, all effects must be fitted as interactions with the subject effect to ensure subject separability. [Andrew Holmes 1 Aug 2000]

How can I integrate conjunction analysis in a multi subject random effect model? Single subject conjunction of two contrasts is easy to perform, but what do I enter at the second level or how do I perform the conjunction analysis over a group of subjects? You do not (generally). A conjunction analysis at the first level, over subjects, addresses the same thing as a single contrast at the second level (using the subject-specific contrast images). In some instances you may want to do a conjunction of two contrasts pertaining to different effects at the second level. This requires the two or more [orthogonal] contrasts to be entered into the second-level model with a simple conjunction (e.g. [1 0] and [0 1]). However, you are making strong assumptions about the sphericity of the error terms, at the second level, by doing this. These assumptions might easily be violated (for example subjects who activate in contrast 1 may be more likely to activate in contrast 2). If you are obliged to enter more than one contrast per subject into a second-level analysis then you should qualify your inferences along these lines. [Karl Friston 4 Aug 2000] Random Effects: 2 groups, 2 conditions

suppose the following study: Group A: 12 patients Group B: 12 controls. Each groups performs Task I and II.

We are interested in task specific effects withingroups and group differences between tasks.

To analyse within group effects I perform two fixed effects analysis, for each group seperately with 12 subjects, compute subjects specific con-images and feed them into a second level, Rx-analysis. So far so clear.

But what if I want to know something about the group differences for Task A? I see two possibilities: Either I take the con-images of each group specific Fx-effects analysis. OR I do a third Fx-Effects analysis with all 24 subjects in one group. The main difference is that global normalisation in the first case is group specific in the second case not. But isnt the second case, i.e. to make a new Fx-effects analysis not the correct way because I want to compare both groups?

to do the between-groups analysis, use the two-sample t-test option under basic models and then provide the subject-specific con*.img files for each of your two groups. this will test for differences in that effect between the groups (in either direction depending upon the contrast you specify).

Is it better (more correct) to take the con*images from one Fx-Effects analysis (24 subjects, both groups) or from two seperate Fx-Effects analysis (12 from one group, 12 from the other).

you should run a separate fixed effects analysis on each subject and then enter the con* images from those analyses into the two-sample model (one for each subject). [Russ Poldrack] As far as I understand the Con images should be the same no matter how many subjects is in the design matrix. The difference between the two setups lies within the residual mean square (ResMS) and thus the Ressels per voxel (RPV) image and the T maps. Because one should only put contrast images into the second level analysis it makes no difference which setup you use. If you also want to make fixed effect analysis, you will need the design matrix including all subjects, and in that case I would just use the big one. If you do not care about first level analysis and have many subects I would go for one designmatrix per subject (parameter estimation works faster with smaller designmatrices). [Torben Ellegaard Lund]

I have two groups of fMRI subjects (10 young subjects and 10 old subjects). Both groups performed an AB block design where B was a control task. All data was spatially normalized. I have computed the effects of task A in each individual subject, and for the youngsters as a group, and for the oldsters as a group and for the total group.

How do I determine regions that were significantly more active in the young group than in the old group (and vice versa).

I would recommend the random effects approach to test the interaction between groups:

1. compute individual (first level) contrast images of A-B for each subject:

2. Second level analysis: hit the basic models button. select a two sample t-test You dont need global normalization here since you did it (hopefully) at the first level. Enter the contrast images of A vs B in the young group for one sample and then do same for older group.

3. The set up contrasts with Results button. Your contrast should be: 1 -1 Young > Old -1 1 Old > Young

[Sterling C. Johnson 30 Jun 2000]

I just saw that Sterling already answered, but I didn't want to throw away my answer...

I would put all subjects of both groups into one model. Each subject gets a separate session, in which you model condition A. You can then choose 2 contrasts to test young vs old. Given that you entered first all young subjects, then all old subjects the contrast for testing, where your young activate more than your old subjects would be: 1 1 1 1 1 1 1 1 1 1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1

and old more than young: -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 1 1 1 1 1 1 1 1 1 1

You could also proceed to a 2nd level analysis. To do this you create one contrast-image for each subject, where the appropriate contrast for young subject #7 would be: 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0

You then enter the 20 images into a 2-sample t-test in 'Basic Models'. [Stefan Kiebel 30 Jun 2000]

We are analyzing PET data of subjects from 2 groups who underwent the same stimulation. Group1 contains 7 subjects, Group2 contains 12 subjects. We can do a single subject analysis and a RFX group evaluation of group 2. For the group evaluation of group 1 the number of 7 subjects might be not enough. (?) We want to do a comparison of the 2 groups. Which model do we have to use for the first and second level analysis?

For the 1st-level analysis use a 'multi-group conditions and covariates' for the 2nd-level analysis compare the two sets of subject-specific contrasts with a two sample t test under 'basic models'. The latter will have reasonable power because you are using all 19 subjects. [Karl Friston 10 July 2000]

Random Effects: 3 cond A, B, Rest

Let me ask one simple question about fMRI Random Effects Analysis(SPM99). I have 3 conditions, A, B, and Rest. I compared A and B for each subject, specifying the t-contrast (1,-1), and obtained con_.imgs. Then I tried to perform a group analysis using random effects model. I selected "one sample t test" from Basic Stats menu and selected the con_.imgs. So far so good.

Yup - that's all fine.

Then the program requested to specify contrasts again. My question is how to specify this second level contrast and what it means. I assume that one sample t-test refers to a significant test against the null hypothesis: the populaton mean=0, so I guess that the second level contrast should be "1". Is this correct? and If so, why should we specify it? (because I cannot think of any other numbers. As long as one sample t test is concerned, it must be always 1, isn't it? What does it mean when it is , say, -1?)

You're right: for the second level analysis you specify, your H0 is that the population mean is not significantly different to 0 (i.e. no consistant mean effects). Doing an F at the second level would test for ANY mean effects significantly different to 0 - WITHOUT any constraints on the direction of these effects.

However, using your one-sample t, you can test for the direction of the difference if you have prior predictions that some areas will show increases while other may show decreases (I am sidestepping the issue of how one wishes to interpret 'deactivations' or mean decreases in fMRI). So a contrast of (1) means 'test for a significant mean (+ve) effect' and a contrast of (-1) means 'test for a significant mean (-ve) effect'. [Dave McGonigle] Random Effects: 1 group, 2 cond., 1 covariate, 2 nuisances

I will first start with what we have: Within an fmri study, One group Five subjects Two conditions Auditory Monitoring versus its own baseline Working Memory versus its own baseline Two nuisance variables anxiety score (one score per subject) Depressive mood score (one score per subject)

One covariate of interest error score on the working memory task This is what we did Design Description Desgin: Full Monty Global calculation: mean voxel value (within per image fullmean/8 mask) Grand Mean scalingL (implicit in PropSca global normalization) Global normailzation: proportional scaling to 50 Parameters: 2 condition, +1 covariate, +5 block, +2 nuisance 10 total, having 7 degrees of freedom leaving 3 degrees of freedom from 10 images

Is this a valid way of looking at this? We are concerned with the large degrees of freedom that we are using up. Also how would we accurately interpret such a model? Does the statistical map only represent activations that are associated with the covariate of interest after controlling for anxiety and depression scores?

Firstly I assume this is a second level analysis where you have taken 'monitoring' and 'memory' contrasts from the first level. If this is the case you should analyse each contrast separately. Secondly do not model the subject effect: At the seond level this is a subject by contrast interaction and is the error variance used for inference. Thirdly a significant effect due to any of the covariates represents a condition x covariate interaction (i.e. how that covariate affects the activation).

I would use a covariates only single subject design in PET models (for each of the two contrasts from the first level). A second-level contrast testing for the effect of the constant term will tell you about average activation effects. The remaining covariate-specific contrasts will indicate whether or not there is an interaction. [Karl Friston 17 July 2000]

Random Effects: 4 cond, 4 matched rest

Further to my question to you earlier this week which was:

17. My paradigm is a block design with 4 different active blocks each followed by it's respective null block, i.e I have 4 different null blocks. How do I go about specifying the design matrix for a second level analysis taking these different null blocks into account? e.g. if my 4 active blocks are: A1 A2 A3 A4 and my 4 null blocks are: N1 N2 N3 N4

If I specify trials 1-8 in the following order: A1 N1 A2 N2 A3 N3 A4 N4

how do I contrast [A1-N1] - [A2-N2]? or vice versa?

Your answer was: A.You would simply specify 8 conditions (A1 - N4) and use the appropriate contrasts.

Unfortunately, 'use the appropriate contrasts' is the bit we don't know how to do now that we have so many different nulls.

I have specified the conditions 1-8:A1 N1 A2 N2 A3 N3 A4 N4 for simple contrast A1-N1 I've used:1 -1 0 0 0 0 0 0

& for contrast A2-N2 I've used: 0 0 1 -1 0 0 0 0 How do I specify a 2nd level contrast looking at the activity in A1 minus it's null N1 versus the activity in A2 minus it's null N2, i.e. [A1-N1] - [A2-N2]?

If I use:A1 N1 A2 N2 A3 N3 A4 N4 1 -1 1 -1 0 0 0 0 then surely this is just adding the activity in A1 and A2 and taking away the activity in N1 and N2 which is not what we want to do. In fact this is [A1-N1] - [A2-N2] and is exactly what you want. I think the confusion may be about the role of the 2nd-level analysis. To perform a second level analysis simply take the above contrast [1 -1 1 -1 0 0 0 0] and create a con???.img for each subject. You then enter these images into a one sample T test under 'Basic Designs' to get the second-level SPM. To do this you have to model all your sujects at the first level and specify your contrasts so that the effect is tested in a subject-specific fashion:

i.e. subject 1 [1 -1 1 -1 0 0 0 0 0 0 0 0 0 0 0 0 ... subject 2 [0 0 0 0 0 0 0 0 1 -1 1 -1 0 0 0 0 ... ... [K Friston 19 July 2000]

Random Effects: variable event-related

I am currently doing an event-related fMRI study using a memory task. I would like to do an RFX analysis on hits - baseline. However, the problem is that the individual subjects have considerably different hit scores, so that different numbers of events are taking into the first level analysis, giving different degrees of freedom (ranging from 90 - 130). I understand that this can be problematic in that RFX analysis assumes equal degrees of freedom. Is there any way to correct for this?

In fact the second-level analysis, used in SPM, assumes that the design matrices are identical for each subject. To do your analysis properly you would have to fit a hierachical model with different variance estimators for each subject. However, remember it is the variance of the contrast that is the critical variance here and the d.f. for each contrast will be the same (assuming the same scan number for each subject). The only thing you have to assume is that the variable number of 'hits' in each subject does not induce substantial differences in the variance of the contrast estimators. I think that as long as you qualify the results, and make it clear what you have assumed, you should be OK.

We are currently working on this issue but it is too early to quantify how robust the two-stage approach to RFX analyses is at this stage. [Karl Friston 11 Jul 2000] fMRI

HRF (Hemodynamic Response Function)

We are trying to specify a subject-specific HRF as a basis function at the individual level of analysis. The 'subject-specific HRF' is an independently collected and averaged timecourse of signal intensity in a cluster in the motor cortex during a separate (motor) task.

It appears that this can easily be added as an option in spm_get_bf.m, by adding a 'user specified hrf' option for the Cov variable around Line 91, and inserting a condition to import a user-specified vector for bf, instead the default SPM's hrf around Line 150, which is now: [bf p] = spm_hrf(dt);

However, I was wondering, are there any properties that a Gamma basis function has (and that a user-specified vector may not have) that are used at some later stage of the statistical modelling and the benefit of which would be lost in the case of a user-specified vector?

I can't think of any properties that a gamma bf has that would influence the statistics of your analysis above and beyond being a better or worse fit to your actual data. A single gamma bf just happens to be a shape that looks 'hrf-ey' and has the added advantage of being described completely by a single parameter. The more mathematically gifted may wish to correct me if I've over-simplified things here, or just got them plain wrong!

As far as I am aware, the main problem with using any single parameter basis function to describe a complex waveform such as the hrf is one common to all attempts to fit a model to data in linear regression - the model may not be well specified. You can't fit a square peg into a round hole, and so it is often the case that the choice of bf is not appropriate. There are options in SPM that take this into account, and allow the modelling of neurovascular responses by a basis set of more that one function (either the Fourier or 3 gamma bfs options). These will fit any example of the 'family' of responses that can be described by a linear combination of your bfs. The disadvantage is that it becomes harder to relate these more complex fits back to the underlying neural activity that we assume generates the hrf, and so unambiguously talk about differences between evoked responses that we wish to describe by a difference in parameter estimates.

A recent study by Geoff Aguirre and colleagues showed a great deal of shape difference between the hrfs of different subjects in the region of the central sulcus to a transient motor response (The variability of human, BOLD hemodynamic responses; Neuroimage 1998 Nov;8(4):360-9), but less variability in hrf shape within subjects when studied over a number of different runs/sessions. This group now regularly defines subject-specific hrfs which are then used in subsequent analyses - the approach is described in 'Using event-related fMRI to assess delay-period activity during performance of spatial and nonspatial working memory tasks. Brain Res Brain Res Protoc 2000 Feb;5(1):57-66'. This seems to suggest that your strategy of independently defining subject-specific hrfs is a sound one, assuming you wish to use these to fit responses in the same region, as hrfs may show spatial variability even in the same brain. An SPM answer just wouldn't be an SPM answer without a few caveats, would it? [Dave McGonigle]

In the diagram of my efMRI model basis set (hrf with time and dispersion derivatives), it looks as though the dispersion derivative is the 2nd derivative of the hrf reflected about zero. Is that the case? If not what is it? It is the derivative of the HRF with respect to dispersion (see spm_hrf) for how dispersion is parameterized. This is actually very similar to the second derivative. [K Friston 12 July 2000]

I am learning to use SPM 99 to generate experiments. It seems that the onset times for the various stimuli are contained in the matrix Sess{1}.ons for the nth event type.

I've noticed with several different designs that these onset times are 0.125 seconds later than I would expect: for example, the beginning of the first epoch for a block design -- Sess{1}.ons{1}(1) -- is 24.125 sec when I specify 12 scans (at 2 sec TR) as the time to first trial. This also occurs for event-related designs. This is a reflection of the fact that onset times are specifed in time-bins of TR/16 = 0.125s for you. One starts acquiring the first scan at t = 0 seconds (scans) and finishes at t = TR seconds (1 scan). To ensure indices do not start at 0, 1 time-bin is added to every onset. 125ms does not matter one way or the other given the time constants of the hemodyanic response. [Karl Friston 17 July 2000]

we just stumbeled over a question that might have been discussed earlier. Our fMRI-paradigm is slightly asynchronous to the image acquisition. From the performance of execution of the paradigm we derive a covariate. As during the images at the beginning and end of the block the execution starts or ends, we do not know what to choose as a covariate for these images (for all the other images, we use either the average of the performance data acquired during the image or zero (during rest). In this context we also would like to know, whether the covariate data is convolved with the hemodynamic delay function, and if so, how ?

If the covariate is presumed to have a neuronal correlate then it should be convolved with the HRF (because this is how that effect will be expressed in the data). One way to model this is to treat your block design as an event-related design, where each block is a train of trials or events. Using the parametric option simply modulate each trial with the perfomance measure for that trial. This approach eschews any problem with asynchrony between acquisition and trials and automatically ensures approoriate convolution with the HRF in 'microtime' (i.e. time-bins of TR/16). [Karl Friston 21 July 2000]

I'm trying to model event-related activity that is due to two stimuli presented sequentially with an SOA of 1.5 seconds. In looking at the raw data, the hemodynamic response to such an event often has a wider peak than does the typical HRF for a single event. I'm wondering whether modelling the data using a dispersion derivative (i.e., hrf + time + dispersion derivatives) will enable SPM to fit a canonical hrf to the data that has the appropriate width. In other words, does the dispersion derivative allow the width of the of the canonical HRF to be adjusted analagous to the way that the temporal derivative allows the onset of the canonical response to be adjusted? Absolutely. [Karl Friston 21 July 2000]

You might consider using the 'parametric modulation' option to explore time-dependent effects by creating boxcar regressors that are modulated by an exponential (or other) function of time. This is slightly more complicated to set up, but has the potential merit of greater flexibility in characterising the nature of any time-dependent effect. [Geraint Rees 26 July 2000]

Can anyone tell me the best design matrix that I can specify for the SPM analysis? Also, what are the effects of selecting "SOA-variable", "Convolve with HRF" and "Add temporal derivatives" for fmri model setup? TR: 3000ms TE: 60ms ON-OFF experiment: 12-12-12-12¡Kstart with off state Multi-phase: 62

One possibility is to make a design matrix with one condition. Your SOA is fixed (the time between the start of each ON block) and these onsets represent EPOCHS that are 12 scans long. Do convolve with the HRF, don't add temporal derivatives (at least at first), select the default high pass filter and use hrf for low pass filtering with no modelling of the autocorrelations. Use a contrast of [1] to visualise ON>OFF. That should do the trick I hope!

To answer your questions:

SOA-variable is used when either the start of the ON blocks, or the timing of invdividual events in an event-related design, is irregular.

'Convolve with hrf' convolves the regressors with a synthetic haemodynamic response function. This is usually a good idea unless you have made up your own user-specified regressors that are already pre-convolved.

'Add temporal derivatives' makes a second column in the design matrix for each regressor that approximates the temporal derivative of that regressor. The idea is that some (unknown) combination of these two columns, appropriately weighted, can model a (small) temporal shift in the fit of the original regressor. This can be used to improve the overall fit of the model if precise timings are uncertain, or to formally test for differences in the onset of the BOLD response for different conditions. [Geraint Reese 12 Aug 2000]

Back to top.

Extract fMRI time-course

I would like to extract time courses for the whole brain, however I quickly run into memory problems when trying to extract regions with a radius larger than 40mm. Is there a way of getting the time courses more directly? i.e. is there an object somewhere that contains all the time courses (preferably high pass filtered)?

One possibility is to use the Y.mad file to extract the raw time course of voxels (it will only have information about those voxels that survived the upper F-threshold specified in you defaults file, so you might want to raise this threshold to 1 in order to write all voxels' timecourse). Then in matlab you could do something like

load SPM and then to extract a voxel with voxel coordinates [39, 19, 1] and put the timecourse in variable Y, type: idx = find(XYZ(1,:)==39 & XYZ(2,:)==19 & XYZ(3,:)==1); Y = spm_extract('Y.mad',idx);

To extract a voxel defined in terms of mm coordinates, you'd have to first convert the XYZ variable above, into mm coordinates index:

XYZmm = M(1:3,:)*[XYZ; ones(1,size(XYZ,2))];

and then do the same as above but this time using the XYZmm variable and the estimates of the voxel location in mm. I think you can also apply the filtering you've specified at the design specification stage (both high and low if they've been specified) by:

Y = spm_filter('apply',xX.K,Y); [Kalina Christoff] Slice Timing

I am learning to use SPM, and I have a question about slice timing correction that I haven't been able to find an answer for. To wit, does SPM 99 properly correct the slice timing for coronal slices? The program seems to expect axial slices, and if it works for slices on other axes, I don't know how it determines the axis to use.

The slice timing program doesn't care about orientation per se, just that you have a time series of images. In terms of specifying the reference slice you have to very careful about this if the acquisition is not ascending or descending as the interleaved order is not intuitive (you might just specify the order yourself using the option provided).

SPM in general expects axial slices for lining up the activations on the glass brain templates but for running the statistics the orientation doesn't matter as long as the origin is specified correctly. If you've figured out which slice is 1 for your origin in the coronal direction then use that numbering scheme to pick the proper reference slice. [Darren R. Gitelman 15 July 2000]

Thank you for your reply, Darren. I'm afraid I still don't understand exactly what is going on, though. Here's my situation: I have 140 3D image files from a scan series, one every 2 seconds, that I want to work on; each file contains 20 64x64 coronal slices. (I made these 3D files because I thought that's what SPM needed to work with.) Is it that I need to provide SPM with a set of 20 coronal slice files instead, and repeat the timing correction for each of the 140 scans? I had this issue for a study myself. If you have reconstructed your coronal slice set as a volume in the correct spm orientation then YES you do need to be careful as spm will be expecting the data to be acquired in the z-direction (ie SI).

My solution: use AIR to rotate the volumes with yz then do slice time correction, then rotate back with zy option. Of course you will need to double check what becomes the effective top of the volume for specifying to spm the slice acquisition order. The slice timing is done on your session (140 scans). [Robert C. Welsh, 15 July 2000]

Christian Buechel's post

http://www.mailbase.ac.uk/lists/spm/1999-07/0154.html

gave a formula for correcting an onset time, to account for SPM allowing a user-specified choice of which time point to sample regressors (fMRI_T0). Is this formula still applicable, or has spm99 now do this automatically with no user input? That formula is still valid, and nothing has changed in SPM99. We thought about linking the reference slice chosen in slice-timing correction to the value of fMRI_T0, but decided against it because some people do not use the slice-timing stage during preprocessing (eg with long TRs, when the interpolation error may be large, or for blocked designs, when model/slice timing differences are often negligible).

Minor related question: is there any reason that fMRI_T (the number of sample points per TR for e.g. the hrf) should be the same as the number of slices? E.g., if I have 24 slices and slice-timing correct to the middle slice, I think it should be OK to use other values of fMRI_T, as long as fMRI_T0/fMRI_T = 0.5. Is that right? You are right. The value of fMRI_T could match the number of slices, but doesn't have to - it is the ratio you describe that is important. Basically, increasing the value of fMRI_T may gave you more temporal precision, but the advantage is likely to be negligible (unless you have a long TR, eg >4s) and the computation of covariates will take longer. [Rik Henson 11 Aug 2000]

I'd be much obliged if someone could confirm that my understanding from the spm_archives is correct. To slice time correct a volume of 21 slices, acquired in an interleaved manner from bottom to top as:

"1 3 5 7 9 11 13 15 17 19 21 2 4 6 8 10 12 14 16 18 20"

(Where the numbers are the Analyse format spatial positions (i.e.1=bottom)) Is the above number phrase exactly what I would need to enter into the GUI under the user specified option?

Assuming that is the particular interleaving of your scanner, yes. [Rik Hensen 17 Aug 2000]

| I have applied slice timing to my event-related series. For motion correction | (spatial realignment) I would like to use a non-SPM software. This software can | read the aV*.img images but not aV*.mat files. | | My question is: have aV* images been actually written with the slice timing | correction imbedded? That is, if I use them outside SPM without their .mat | files, will slice timing be still taken into account? If not, what should I do?

I think you can quite happily delete these .mat files as they probably don't contain any additional information. Whenever any new images are derived from existing ones, then the positional information is preserved in the new set of images. This positional information is derived from the .mat files if they exist, but otherwise from the voxel-size and origin fields of the .hdr files. If the origin contains [0 0 0], then it defaults to mean the centre of the image, which, depending if you have odd or even image dimensions, is either an integer value or halfway between two voxels. Because the origin field can only store integers, half voxel translations can not be represented in the .hdr files, so this information is written to .mat files. [John Ashburner 25 July 2000] Back to outline.

Interslice gap

| I'd like to know how to handle EPI scans w/ interslice gaps when analyzing | using SPM99 (typical realign>coreg>norm>smooth'ing). For example, let's say | I have 14 slices of epi scans w/ slice thickness of 7mm and gap of 2mm. Do | I spcifically setup the header file or just handle it as if it's 9mm thick? | Any comments are appreciated.

Just enter the third voxel as 9mm. [John Ashburner 20 July 2000] Small Volume Correction

1. Uncorrected p value.

I am having some difficulty determiniing what is acceptable in terms of an uncorrected p value. For example lets say I set the voxel threshold to an uncorrected level of p<0.01 and get a list of regions, most of which I would predict. Typically The areas I am interested in are between the range of p<0.009 and p <0.0001 (uncorrected). When I set this to p<0.001 I obviously loose some of the areas. Is using a threshold of p <0.01 (uncorrected) level acceptable or should I use p<0.001

I would report the results at two levels. First go through the anticipated regions using a SVC (at p <0.05). Second report anything you had not predicted (but is interesting) 'descriptively' at p <0.05 uncorrected. The second reports can be used by you or others to guide anatomically specific hypotheses in future experiments. Say at the beginning of the results that you will be reporting the results at these two levels.

My next step is to focus on the areas of interest, basal ganglia (e.g globus pallidus) or frontal cortex and use the SVC option. When I run the SVC do I use the voxel or cluster level (corrected) p value ?

Use a voxel-level SVC p value.

2. SVC

In one of your previous emails to me you suggested using a SVC with a radius of 16 mm for anticipated areas and a correction for the whole volume for non anticipated areas.

In my case the anticipated areas are cortical (eg parietal or frontal gyri) and sub-cortical (basal ganglia) and as such will have different shapes and volumes. Is it appropriate therefore to use a fixed value (16 mm) for all these regions or should I use different radii for each of the different regions ? I have been using the sphere option and specifying a 16 mm radius for the sub-cortical regions but am unclear what to use for the cortical regions. In one of your previous papers,

10. Neuroscience Dec 1999 p 1087, you based your corrections on the volume of interest and cited Filipek et al 1994. and Worsley et al

1996. I have the Worsley paper and know there is a table that lists the cc values of different areas.

I think you have the lattidude to change the radius depending on the structure involved. You could use any published data as a guide. The Caudate, for example, would have a small radiou (sqy 6mm) whereas parietal cortex should be greater unless you specify which part (e.g. IPS).

[Who/when ???]

Hemisphere (L/R) effects

5. In the last mail, you said that for flipped vs non-flipped analyses I had to consider flipped images as coming from different subjects. As I work in PET I should thus do a multi-study analyse? Can't I just say the flipped images come from the same subject but are from a different ? condition? No because you want to remove the main effect of hemisphere. Treat each the flipped image as another subject. This gives a better model that accomodates both the subject effect and hemisphere effect as confounds. Hemisphere x condition effects can be tested using the appropriate T-contrast.

http://www.mailbase.ac.uk/lists/spm/2000-05/0142.html [K Friston ?]

Back to top.

Image Scaling

| Does anyone have a C program which can be implemented within | MATLAB or otherwise, which can read in an image in Analyze format, | then multiplies each and every voxel by a user defined scale factor | to produce a new 'scaled' image?

The easiest way is to change the scalefactors in the .hdr files, which can be done something like: V = spm_vol('imagename.img'); V.pinfo(1:2,:) = V.pinfo(1:2,:)*scalefactor; spm_create_image(V);

Alternatively, the ImCalc button will do this. Select the image you want to scale, then enter an output filename. The operation you want to perform is then something like i1*scalefactor . [John Ashburner]

The easiest way of converting to a different datatype [e.g. from 64-bit to 16-bit] is probably something like:

VI = spm_vol(spm_get(1,'*.img')); VO = VI; VO.fname = '16_bit_version.img'; VO.dim(4) = spm_type('int16'); dat = zeros(VO.dim(1:3)); for i=1:VI.dim(3), dat(:,:,i) = spm_slice_vol(VI,spm_matrix([0 0 i]),VI.dim(1:2),0); end; spm_write_vol(VO,dat); clear dat

[John Ashburner 25 July 2000]

Back to top.

Output Files

I need to write a script to pull out number of voxels activated in a given ROI for a group of subjects. I have done the estimation and results etc to build contrast images etc. I presume the con images just hold the z score (???) for a given location. It looks like the spmT images are just coordinates?

The con-images are contrast-weighted estimated parameter images. The t-images are the computed t-values for a given contrast.

So one, can someone tell me the correct files to use to do the analysis I would like to? And second, a more general question, is there a description somewhere of what information each file holds etc that is produced at Results time. I can piece together information from here and there, but it would be nice to have a concise description of all the files from one locale.

Yes, the documentation seems to be a bit sparse here. Type 'help spm_getSPM' for some information about the output-images.

in the img files, spm_read_vols returns a 3d volume and another matrix which appears to be a set of coordinates. What is the purpose of the xyz matrix (just curious and trying to understand the innards of spm) XYZ denotes the coordinates of (original) image intensities as stored in Y.mad. There is more information on this in spm_spm.m (l. 148 - 169), which should clarify things. [Stefan Kiebel 24 July 2000] Back to Outline.

4. Concepts

1. Spatial Normalization

Realignment is not a panacea for all movement artifacts within the scanner. Some of the artifacts it does not correct are:

Interpolation error from the resampling algorithm used to transform the images can be one of the main sources of motion related artifacts. When the image series is resampled, it is important to use a very accurate interpolation method such as sinc or Fourier interpolation.

When MR images are reconstructed, the final images are usually the modulus of the initially complex data, resulting in any voxels that should be negative being rendered positive. This has implications when the images are resampled, because it leads to errors at the edge of the brain that can not be corrected however good the interpolation method is. Possible ways to circumvent this problem are to work with complex data, or possibly to apply a low pass filter to the complex data before taking the modulus.

The sensitivity (slice selection) profile of each slice also plays a role in introducing artifacts.

fMRI images are spatially distorted, and the amount of distortion depends partly upon the position of the subject's head within the magnetic field. Relatively large subject movements result in the brain images changing shape, and these shape changes can not be corrected by a rigid body transformation.

Each fMRI volume of a series is currently acquired a plane at a time over a period of a few seconds. Subject movement between acquiring the first and last plane of any volume leads to another reason why the images may not strictly obey the rules of rigid body motion.

After a slice is magnetised, the excited tissue takes time to recover to its original state, and the amount of recovery that has taken place will influence the intensity of the tissue in the image. Out of plane movement will result in a slightly different part of the brain being excited during each repeat. This means that the spin excitation will vary in a way that is related to head motion, and so leads to more movement related artifacts.

Ghost artifacts in the images do not obey the same rigid body rules as the head, so a rigid rotation to align the head will not mean that the ghosts are aligned.

The accuracy of the estimated registration parameters is normally in the region of tens of micro m. This is dependent upon many factors, including the effects just mentioned. Even the signal changes elicited by the experiment can have a slight effect on the estimated parameters. [John Ashburner 4 Aug 2000]

| 1) Is this correct: the more basis functions are used the better the result | ?

In theory, the best results can be obtained with the most basis functions, although the amount of regularisation may need to be tweeked to achieve this.

| | 2) Why is the default ( 7 8 7 ) , not for example ( 7 7 7 ) ? In other | words, why not a same number for each axis?

The dimensions of the SPM template images are 91x109x91. With the 7th lowest frequency basis function, there are 3 whole cycles((n-1)/2) over the 91 voxels, meaning that a period covers 30.333 voxels. The 8th basis function has 3.5 cycles, so a whole period lasts for 31.142 voxels.

| | 3) How does the choice of basis functions ( x y z ) affect the spatial | normalization of each axis ?

This is a very difficult one to explain without pictures. Basically, the choice of basis functions determines the types of deformations that can be modelled. Displacements in all three directions are modelled by the same number of parameters. The dimensions (e.g., [7 8 7]) reflect how many low frequency coefficients of a 3D DCT are used to model displacements in each of the directions. It is difficult to explain, but I have included a few lines of Matlab that may illustrate the point. The effect of the regularisation is not modelled though. To change the number of basis functions in the different directions, you would modify the first line, before copying and pasting into Matlab.

d = [8 3]; [X1,X2]=ndgrid(1:64,1:64); B1=spm_dctmtx(64,d(2)); B2=spm_dctmtx(64,8); Y1=X1+B2randn(d(1),d(2))20B1'; Y2=X2+B2randn(d(1),d(2))20B1'; plot(Y1,Y2,'k',Y1',Y2','k'); axis image xy off [John Ashburner 27 Mar 2000]

Back to outline.

2. Smoothing

3. Smoothness estimates.

I have a couple of basic questions regarding the code contained in the spm_spm.m file for the estimation of map smoothness from the residuals.

The smoothness estimation is described in

SJ Kiebel, JB Poline, KJ. Friston, AP. Holmes, KJ. Worsley Robust Smoothness Estimation in Statistical Parametric Maps Using Standardized Residuals from the General Linear Model. Neuroimage 1999; 10, 756-766

Note that the general principle used for the smoothness estimation, as described in this paper, is employed in the current version of SPM (SPM99), however there are differences in the algorithm/implementation due to new developments made by Keith. These developments allow for non-stationary smoothness and involves taking the expectation of the determinant of the [co]variances of the first spatial partial derivatives of the residuals as opposed to taking the determinant of the expected [co]variance. For stationary fields these are the same. For non-stationary fields the new estimator is valid.

1. Is the FWHM calculated for the t-field components of ei, the t-field t-null or the Gaussianized T-field Z-null? I suspect what is desired is the second (t-null), but could not identify code that converted from the components ei to t-null.

In SPM99, the FWHM is calculated on the residual fields. This allows one to estimate the smoothness in terms of FWHM of the resulting t-field under the null hypothesis. The smoothness of the residual fields are an estimate of the smoothness of the underlying component fields. These component fields are not just those of a t field, they could also be for an F field or any other statistic. They are a generic representation of the data (not any statistical process derived from the data).

2. This question may be the answer to my first. What is the purpose of this line of code at position 1142?:

%-adjust FWHM such that prod(1/FWHM) = (unbiased) RESEL estimator %----------------------------------------------------------------------- FWHM = FWHM*((RESEL/prod(FWHM(1:N))).^(1/N));

The important thing here is that instead of taking the determinant of the sum of the squared partial derivatives, one takes the sum of the determinants of all squared partial derivates (see above). Effectively, this enables one to estimate the RESEL count in non-isotropic and non-stationary data. We no longer deal with FWHM or smoothness per se but consider the number of RESELS the search volume comprises. The computation of corrected p values, for any statistic, only requires this [smoothness-independent] volume metric. The FWHM characterization above, saved by spm_spm, is the equivalent stationary [non-isotropic] FHWN that corresponds to the number of RESELS in the search volume (it is not actulally used by later routines). This stationary FWHM equivalent is given by the line above.

Perhaps this is the conversion that I'm looking for?

No, not really. As stated in the paper (p 759, 2nd column) the conversion you are looking for no longer exists in SPM99. The smoothness of the Gaussianized T-field was last estimated in SPM96, but was abandoned on SPM97 for the more robust estimators based on the residuals.

3. I couldn't find any evidence of the effective degrees of freedom correction that is necessary to compensate for working with the estimated component fields [i.e., the correction factor (v-2)/(v-1)]. Is this actually in the code someplace and I'm missing it?

You are right, this factor is now redundant because we are taking the expectation of the determinant (as opposed to the determinant of the expectation). [Stefan Kiebel 21 Mar 2000]

Back to outline.

4. Gaussian Field Theory

I need a reference for the adjustment that is performed to correct for nonindependence of voxels when determining the signficance level of activations in the SPM(t) maps for fMRI.

I'm not certain which paper would be the correct one to reference. Guidance would be most sincerely appreciated!

This is simply Gaussian Field Theory:

Friston KJ Holmes AP Worsley KJ Poline JB Frith CD and Frackowiak RSJ Statistical Parametric Maps in functional imaging: A general linear approach Human Brain Mapping 1995;2;189-210

Worsley KJ Marrett S Neelin P Vandal AC Friston KJ Evans AC A unified statistical approach for determining significant signals in images of cerebral activation Human Brain Mapping 1996;4:58-73 [Karl Friston 11 Jul 2000] Back to outline

5. Fixed vs. Random Effects.

Why and when should I use second level analyses with PET? If, for instance, I want to compare flipped with non-flipped images in one group of subjects, I think I can do that without second level analyses. If I want to compare the difference between flipped and non-flipped in one group compared to another group of subjects it seems to me I can still do this without second level analyses but I think in one of the mails of the archive I read that you should do that with second level analyses, why?

The same applies to PET and fMRI. First-level analyses use within-subject variance and provide for inferences that generalize to the subjects studied. Second-level analyses use between-subject/session differences whose estimator is the correct mix of within and between-subject error to give a Mixed-effect (i.e. Random) analysis. This allows you to generalise to the population from which the subjects came.

In PET (but not fMRI) the similarity of between- and within-subject variances and between the number of scans per subject and the number of subjects means that the difference between first and second-level analyses are much much less severe. Traditionally PET studies are analysed at the first level. Second-level analyses are usually employed when you want to make an inference about group differences given some within-subject replications. In your example you would collaspe repeated measures of flipped vs. unflipped into one contrast per subject and then compare the contrasts at the second level. This would be less sensitive but would allow you to generalise to the populations from which the subjects came. [Karl Friston 22 Mar 2000] |We have problems with a simple activation task for two groups: patients |and healthy controls. We are using a block-design with alternating rest |and activation condition (Tr=3s, 100 measurements, 10 measurements per |epoch). Everything was preprocessed and put in one big statistical |matrix. The individual contrasts show that the healthy controls have |activation in brain regions where the patients don't activate. I get |corresponding results when I set up contrasts like 0 0 0 0 1 1 1 1 where |all patients are set to one and the controls are set to zero and vice |versa. Also contrasts like -1 -1 -1 -1 1 1 1 1 seem to give me good |results for interaction. Now, is this method valid?

This depends on the inference you want to make from the comparisons. You are describing a fixed effects model, so the statistical inference is restricted to the specific group of patients and specific group of controls you are studying. Usually in comparing patients and controls you would like to generalise your inferences to the population of patients and controls. This can be implemented in SPM by the 'second level' analysis you describe, effecting a random effects model where the error variance is solely the inter-subject (i.e. intra-population) variance.

|When I do a second level analysis with the individual contrast images |(what information contain these images anyway ?????)

The contrast images represent spatially distributed images of the weighted sum of the parameter estimates for that particular contrast. In essence and for your particular case, it's like a difference image for (activation-rest). You need one contrast image for each patient and each control. By doing that you are collapsing over intra-subject variability (to only one image per contrast per subject) and the image-to-image residual variability is now between subject variance alone.

| With a two-sample |t-test or with one-sample t-tests for each group I get really strange |results. Suddenly the patients are activating more then the controls in |brain regions where not even one patient activated individually. Things |change again dramatically (and get even stranger) when I use |proportional scaling in the first level analysis. Now I would like to |know, what exactly the second - level analysis tells me and whether it |is valid just to work with the first level analysis. There are many possible reasons for the differences, which are basically telling you that the error map for the fixed and random effects models are different (as might be expected). Usually proportional scaling would be used in the first level of analysis, because you want the contrast images entering into the second level of analysis to be on the same scale. You don't tell us how many subjects are in each group, but I infer from the fixed effect contrasts that you have four subjects in each group. In this case, you will not have very many degrees of freedom for the second level of analysis and therefore lack power. In general, the recommendation would be to use 10-12 subjects per group for a second level analysis so adding subjects will help considerably.

As to which method is 'valid', that depends on the nature of the statistical inference you wish to make. In general for comparisons between patient populations and control populations a random effects model will be more appropriate, as one would like to generalise the result beyond the specific individuals studied to the population. You might be interested in the excellent summary of the rfx discussions prepared by Darren Gitelman, which is at http://www.brain.nwu.edu/fmri/spm/ranfx.html [Geraint Rees 9 Feb 2000]

We have a 2-groups PET study (9 Healthy(G1) vs 8 Patients(G2)) with 6 conditions (N,H,A,B,C,S) and 2 replications by condition. N and H are control conditions for the S activation condition A is control condition for B and C activation condition. we wanted to compare the two groups for the S-N, S-H and C-A contrasts. We first defined individual contrasts and we then performed second level analyses in order to compare the two groups, i.e., S-N(G1) and S-N(G2) interaction; S-H(G1) and S-H(G2) interaction; C In a RFX analysis, one wants to look at the variance over subjects. You do this in SPM99 by fitting a model to the weighted parameter estimate images of the 1st level analysis (contrast images). The error variance of this 2nd level model is then over subjects, not over scans, because you have got one image per subject. The degrees of freedom for the estimate of this error variance is here 'number of subjects' - 'rank of 2nd level design matrix', i.e. the degrees of freedom is lower than in the 1st level analysis, in your case 15 = 17-2. [Stefan Kiebel 30 Jun 2000] -A(G1) and C-A(G2) interaction. These analyses have very low degrees of freedom (15). But, since each individual contrast involved in these analyses was obtained from 4 scans (2conditions, 2 replications), the total number of scans involved in the analyse is 4 x (9 + 8) = 68. Is there mean to correct df with the real number of scans?

No, not really. There is a critical distinction between a fixed effects (FFX) and random effects (RFX) analysis of your data with respect to the degrees of freedom and the inference. You described the RFX analysis, leaving you with degrees of freedom, which is a function of the number of subjects and the design matrix employed at the 2nd level analysis.

In terms of inference, the difference between a FFX and RFX analysis is that with a RFX analysis you generalise your inferences to the population of the subjects/patients. With a FFX analysis, you make inferences only about your measured data. However, the more general inference facilitated by a RFX analysis has its price in the lower degrees of freedom available (given that you have more than 1 scan/subject).

There is a helpful website about RFX analyses put together by Darren Gitelman, where he compiled some references and many of Andrew's answers to the SPM-mailbase into a knowledge base about RFX analyses.... http://www.brain.nwu.edu/fmri/spm/ranfx.html

Concerning the degrees of freedom: In a FFX analysis, you analyze the data only at the 1st level. The estimated error variance at a voxel is a function of the model and the actual fit to the data, i.e. you look at the variance over scans. Here one usually has got high degrees of freedom for a group study, because the degrees of freedom is 'number of scans' - 'rank of design matrix'. (PET)

I have a few more questions concerning the multisubject correlation analysis.

Does "Covariates only: interaction with subject" model take into account the dependence within subject? If it does, how does it do that? This is important because we have 8 scans per subject.

We specified the following contrasts (11 subjects) to test

1. positive correlation 1 1 1 1 1 1 1 1 1 1 1
2. negative correlation -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1. Are these t-contrasts correct for testing the average positive and negative correlation?

I wonder whether these results are fixed effects analyses, that cannot be generalized into population level.

You are correct that this is a fixed effects analysis, and cannot strictly speaking be generalised into a population level inference.

The way to perform a random effect analysis on this study is to generate a contrast parameter estimate map for each subject (by simply looking at contrasts [1 0 ... 0], [0 1 0 ... 0] ... [0 .. 0 1] in the results section). You then perform a one-sample t-test (i.e. compare the average of your parameter estimates with zero) on these (con_00*.img) maps. You do the same thing with the [-1 0 ... 0], [0 -1 0 ... 0] ... [0 .. 0 -1] contrasts to check the negative correlations.

The RFX model will tell you if all subject "activate" in the same location and with roughly the same magnitude. If you want to answer the slightly less stringent question "do they all activate in the same location?" you can use a conjunction across subjects instead. Its real easy. Once you have entered all the individual contrasts above you simply select them all (positive and negative separately) in the results section using the control button. [Jesper Andersson 12 July 2000]

In fact the second-level analysis, used in SPM, assumes that the design matrices are identical for each subject.

I've seen this point mentioned before but I think I may have missed its importance. Does this mean that if you randomize the order of presentation across subjects then you cannot use an RFX because each subject has a different design matrix? I think this is rarely done in fMRI studies but it is normal practice for PET. I'm wondering whether this means that the assumptions underlying an RFX analysis are violated when analysing (typical) PET data.

The critical thing, about using the simple 2-stage analysis to implement an exact RFX analysis in SPM, is that the contribution from the error variance (Ce) at the 1st level is the same for each subject. This is

pinv(X)Cepinv(X)'

where X is the 1st-level design matrix. Because

pinv(X)Cepinv(X)' = pinv(X(i,:))Cepinv(X(i,:))'

where i is any permuation of indices, randomizing the order of conditions over subjects will have no effect. Indeed randomizing the onset times of different trial types in fMRI will have no effect (ignoring minor interactions with serial correlations). The only situation where one should be careful is when the number of trials, in X, varies substantially from subject to subject.

Would it make a difference if one analysed a group of subjects (with different stimuli presentation ordering) using the 'conditions x subj' option before generating the subject-specific contrasts? In that case, each subject would be part of the same design matrix although the contrasts would come from independent (and different) subsets.

Yes. For PET one must always model the effects in a subject-separable fashion (i.e. 'conditions x subj'). This is enforced in the fMRI setup because each session is specified separately. [Karl Friston 12 July 2000]

Reading SPM 99 documentation, I have understood that the statistics corresponding to "random effect" are done using a two-stage approach, i.e. calculating one contrast image for each subject (as if only one determination has been performed on each subject, so that the residual df is number_of_subject-1) and then running a second level analysis (I did not find out how this analysis is performed. Is it by comparing the mean t value to 0?). Is this correct?

Yes, the t (and F) tests are against the null hypothesis of zero mean, using the one (or two)-sample t-test option in SPM99.

In books concerning variance analysis, the random effect (mixed models) is generally performed by calculating a F value as the ratio : (main effect linked variance) / (interaction variance). Then, the interaction df is: (number_of_subject-1)*(number_of_replication_per_subject - 1). The contrasts of interest are then calculated in the same way than in SPM but the interaction variance is taken as residual variance.

1. Did I correctly understood the random analysis in SPM?

Yes. In the special case of two conditions (two levels of your main effect) and one replication per subject (or data averaged over balanced replications), the "conventional" F-test you describe and the F-contrast [1] on an SPM one- sample t-test are equivalent. Because the con*imgs already contain the effect parameters for each subject, the residual error in the SPM model is identical to the subject x effect interaction, the denominator of the conventional repeated measures ANOVA.

When there are more than two levels of your factor and you want an omnibus F-test (rather than a specific planned comparison, ie t-contrast), you must use a PET design. However, the resulting analysis uses a pooled error term (even for factorial designs) and there is currently no correction for sphericity violations (so your p-values may be invalid). This is why we advise keeping second-level models to one/two-sample t-tests on specific t-contrast images.

2. As the number of values for the contrast is always low (the number of subjects), is it better to use a non parametric test to compare the mean t value to 0?

It may be, particularly with ~10 or less subjects - see: http://www.mailbase.ac.uk/lists/spm/2000-07/0053.html (though if you use permutation tests, as in SnPM, you are not really treating subjects as a random effect - but then again how often are subject samples for imaging experiments true random samples from the population?)

3. Is the first order risk (false positive) in the two-stage approach in SPM the same (or lower or greater) as in the classical approach (one_stage analysis, F determination and contrasts deducting using of interaction variance in place of between replicates variance)?

4. The same question for second order risk (power).

They are the same in both cases (if I have understood you correctly), for the reasons given above. [Rik Hensen 17 Aug 2000]

6. Contrasts

The contrast 1, 0 tests whether the parameter estimate for the first covariate is significantly different from zero. The contrast 1 -1 tests whether the parameter estimate for the first covariate is significantly different from the parameter estimate for the second covariate. So if a voxel is picked up by the 1, -1 contrast but not the 1, 0 contrast, then this presumably means that the parameter estimate for the second covariate is negative (unless its something funny to do with the error used for each of these comparisons, but I don't think so). Obviously a glance at the parameter estimates themselves will establish whether this is the case, and indeed at least some of these same voxels may show up in the contrast 0, -1 (although others may be sub-threshold in this comparison).

I am not sure that contrasts such as 1, 0 are really 'typical' in event-related experiments. Many groups design their experiments to look for differential effects between different types of event, and I must admit that personally I would always feel much more comfortable with that approach. Otherwise you are imaging everything that is time-locked to your events, much of which may not be relevant to the cognitive component that you are interested in.

Also, if you only have two types of events, then there might be a significant degree of co-linearity between your covariates. In the worst case, for example, where events A and B alternate with a fixed SOA of, say, about 12 seconds, the second co-variate might look a bit like the first one multiplied by -1 ( just because event B occurs during the bits of the time series when event A isn't occurring). If so, you could easily be misled looking at main effects contrasts like 1, 0 or 0, -1. Whether a voxel shows up in the first or second of these contrasts may be determined largely by noise. [Richard Perry 2 Aug 2000]

In the SPM course notes chapter, "Statistical Models and Experimental Design," it is written that "a contrast is an estimable function with the additional property c^T beta-hat = c'^T Y-hat = c'T Y ... Thus a contrast is an estimable function whose c' vector is a linear combination of the columns of X."

Note that being a contrast is a property of c, not c'. I claim that any estimable function c (shorthand for the estimable function c^T beta) must be a contrast.

I think what was meant here is that an estimable function is defined by the property above. (which is not really, this is just a corollary of the usual definition that c^T beta is estimable if there exists c' such that E(c'^{T Y) = c}T beta for any beta. I guess it could be used as a definition ...) I do agree that the sentence in that text is misleading ....

Why: c estimable means that there is a c' such that c^T = c'^T X.

yes, and your "means" is an "if and only if" ...

Pick such a c'. Now define c'' to be the orthogonal projection of c' onto the range of X. Since c' - c'' is orthogonal to the range of X, (c' - c'')^{T X = 0. Hence c}T = c''^T X.

yes, more simply put, c' needs not be unique, but its ortho projec. onto the range of X is unique. Another way of seeing c' is to think of it as a constraint on the model X ... (so it must lies in its space !)

Moreover, c''^{T (Y-hat - Y) = 0, since c'' lies in the range of X, and Y-hat - Y is orthogonal to it. Thus, c}T beta-hat = c''^T X beta-hat = c''^T Y-hat = c''^T Y.

Is this correct?

absolutly, and it is used in the spm routines.

Furthermore, a natural extension of this is that c is estimable (and hence a contrast) if and only if it is orthogonal to the null space of X.

yes, or more simply that the contrast lies in the range of X^T, and that's how we check estimability ... (have a look at spm_SpUtil, spm_FcUtil, and spm_sp if you want more details on how this is handled. In fact, because as you point out c'' is in the range of X, we use the coordinates of it in an orthonormal basis of X to save memory and computation time and work in the parameter dimension rather than the temporal dimension.

A more specific question about contrasts: The same section also states that "For most designs, contrasts have weights that sum to zero over the levels of each factor." If the most liberal definition of contrast is used, which I claim above is equivalent to c being orthogonal to the null space of X, then this would imply that the row sums of the design matrix vanish.

not quite sure what you mean here, what was meant is that in simple factorial designs, valid contrasts usually have weights that sum up to zero.

(I assume that the constraint that c_0 vanish is to insist that the contrast not "see" the constant term; it's not made necessary by the definition itself.)

yes. But in any case, the spm interface would not allow a non valid contrast. [Jean-Baptiste Poline 13 Feb 2000]

I still have some basic questions about the Contrasts ins SPM99. I'm looking at a quite simple paradigm with the conditions [A B R] and would like to test for Voxels where the activation for A is significantly higher than for B. If I look at the results for the Contras 1 -1 and plot the event/epoch-related response for these voxels, all plots show me activation for A and deactivation (negative response function) for B. I just wonder what happened to the voxels where there is activation for both contrasts but more for A.... Is this a problem with my data or do I understand something wrong about SPM?

I don't think there is necessarily a problem with your data. It might simply be a question of power. Obviously the largest differences in your contrast [1 -1] will be found in the cases weher there is an activation in A and a deactivation in B. If your power isn't very high (i.e. if you have relatively small no. of events) it may be that you are unable to detect the more subtle differences resulting from an activation in A and a smaller activation in B. Try to lower the threshold (you may have to lower it in the SPM-defaults to ensure data are saved for plotting) and have a look at some voxels with smaller z-scores and see what you find. You should also be aware that an "activation" or a "deactivation" is always relative to some baseline which may be more or less well defined. If you are using rapid stimulus presentation (short SOA) without null events it will be less well defined, and it will be very difficult to determine between activations and deactivations. In that case the interpretation of a positive finding in the contrast [1 -1] can be larger activation in A than in B, or less deactivation in A than in B, or anything in betweeen. If R in your design denotes null events you are in a better position and the question of activations or deactivations should not be determined from plots of event-related responses of event types A and B, but rather be based on the [1 0 -1] and [0 1 -1] contrasts. [Jesper Andersson 10 Jul 2000]

What is the statistical value of the contrast of parameter estimates?

A contrast is just a specific weighting of the parameter vector. This is used to specify a null hypothesis at each voxel (e.g. there is no activation in condition 1 as compared vs. the rest condition). At each voxel, a t-value is then computed by dividing the scalar product of the contrast and the parameter vector by the estimate of the standard error. After this a p-value (corrected for multiple comparisons) is computed for each t-score. You can then assess the significance of this p-value (e.g. p <0.05). [Stefan Kiebel 11 July 2000]

If the selected contrasts for a conjunction analysis are not orthogonal, SPM asks the orthogonalization order of contrasts. The results seems to depend on which contrast is first in orthogonalization order. The first contrast remains the same, but those who are non-orthogonal regarding the first contrast change. One assumption made for a conjunction analysis in SPM99 is that the used contrasts are orthogonal in such a way that the spaces spanned by the contrasts are orthogonal to each other, i.e. if X is your design matrix and c1 and c2 two contrasts, then orthogonalizing c2 with respect to c1 means that c2 is changed such that

c1'*pinv(X) is orthogonal to c2'*pinv(X) where pinv(X) = (X'*X)^(-1)*X'

How to interpret the results of an orthogonalized conjunction analysis? I don't know of an generally applicable interpretation about orthogonalized conjunction analysis, but why not just say that you modified the contrasts such that the spaces spanned by the contrasts are orthogonal to each other. This removes common subspaces spanned by more than one contrast from all but one contrast, which would otherwise make your conjunction analysis invalid. Maybe others might want to comment here. [Stefan Kiebel 20 July 2000]

Back to top.

7. Confounds

I would like an opinion on the following issue: we have submitted a paper on a PET experiment, where we collected data from 11 volunteers and performed different comparisons between 3 experimental conditions and a baseline with SPM96. During the experiment, we collected on-line reaction times and accuracy scores. In the paper we report that the behavioral performances (both reaction timies and accuracy) of the 11 subjects were significantly different from each other (p=0.0001). However, we didn't include the behavioral scores in any way in the SPM analysis (one reason was that we had no hypothesis about any possible physiological impact of the behavioral scores). We now have been asked from a referee to include the behavioral data for each single subject in the SPM analysis as a confound. Are you aware of any publications, where as a similar approach was adopted? And, more in general, do you think such an approach would be correct?

This is my opinon:

This is a difficult question to answer in a general way. There are an enourmous number of instances where behavioural or psychophyisical data are entered into the design matrix as explanatory variables. The objective here is generally to find the neurophysiological corrrelates of the data in question. On the other hand it would be ridiculous to include a behavioural response variable as a confound if its variance was caused by an experimental factor already in the design matrix (for example including visual analogue scores of pain as a confound in a pain study would be silly).

The question you have to address is whether the RT and accuracy data contain information that is independent of the effect you are interested in and, if so, would the analysis be better if you used the RT and accuracy data as surrogate markers for this confounding effect. You then have to think carefully about the orthogonalization scheme you would adopt in the case of collinearity.

If the RT and accuracy data represent measures of the process you are interested in, then you could include the RT and/or accuracy data as regressors of interest and report these effects.

In general reviewers who specify that a particular statistical model should be adopted before the report will be considered for publication are, in my opinion, in danger of over-stepping their brief. On the other hand, reviewers are, generally, only trying to help you present your ideas in a valid and clear fashion. It may be that the reviewer thinks you are mis-attributing activations to one experimental cause when there is another more parsimoniius explanation that is reflected in the RT and accuracy data. [Karl Friston 21 July 2000]

Back to top.

8. Eigenimages

"eigenvariates" and "eigenimages" refer to a factorisation of functional imaging data of the following form. Say you have your data stored in an nxm matrix Y where n is the number of voxels in each image volume, and m the number of scans in the time series. We may now factor Y in the following form

Y = USV'

where U is an orthogonal nxm matrix, S is a diagonal mxm matrix and V an orthogonal mxm matrix. This is a bit like trying to factorise a number into the product of three other numbers, i.e. there are a lot of ways of doing it. To make the factorisation unique the first column of U multiplied with the first number in S multiplied with the first column of V has to be the combination which explains the maximum possible amount of the variance in

25. The second combination has to be that which explains the most variance in Y after the first one has been removed etc. This is called singular value decomposition.

In neuroimaging the columns of U are typically/often denoted eigenimages, and the columns of V may be called "eigenvariates", "eigentimecourses" or something like that.

Its usefulness in neuroimaging comes from the similarity between the factorisation above and the multivariate version of the general linear model

Y = P*X' + E

where Y is still the data, X is the good old design matrix, the columns of P contains the parametric images (one for each column of X) and E is the error matrix.

For well behaved data (e.g. PET) the SVD factorisation can sometimes be useful in that the "automatic generation" of the "design matrix" V can give new insights to the experiment. It can also be used as a data reduction method or as a preconditioning of data prior to e.g. ICA.

As a general reference I quite like the explanation given in "Numerical Recipies in C" by Press et al. For Neuroimaging purposes you might refer to Friston et al 1993 in JCBFM, or you could look for papers by S. Strother. [Jesper, 20 Jun 2000] Back to top

1. Power Analysis

I have a question regarding the power of SPM results. Specifically I have FDG PET studies obtained in 17 adult controls and 7 children older than 6 year of age. I performed a routine SPM analysis at the 0.05 level and no significant differences can be detected. The reviewer however argues that this is a null finding of group differences between small groups and thus I cannot conclude that subtle differences do not exist. How can I determine what power my results have?

I assume that you have the t-values from the SPM analysis comparing responses at particular areas of interest between the adults and children. Although the t-values themselves are not a good measures of effect size (because they are a function of sample size), for between group designs there is a simple transformation that converts t to the effect size statistic d, assuming that there is no treatment x subject interaction.

The general formula is d = [1/sqrt (p*q)] t / sqrt N , where p and q are the proportion of cases in the two groups and N is the total sample size.

It seems to me that you would be able to use this effect size estimate in the standard formulas for power (e.g., Cohen, 1977) to determine the power for the specific effects in your study, to estimate sample size required for future studies, and to reply to the reviewer.

Note that the situation is not as straightforward if one is using a within subject design or a mixed design. One would generally want to test specifically for a treatment x subject interaction as well as consider the test-retest reliability of the dependent variable when computing the effect size estimate. In SPM conjunction analysis seems to have been used to address similar issues. {Frank Funderburk 26 Jun 2000]

I have a quick question: performing spm analysis of PET data comparing two groups (17 and 7 subjects) and not detecting any significant differences, what is the power for my null findings? In other words, how confident can I be that I did not missed effects larger than say 10% difference?

In theory you should be able to do this using the standard error of the contrast of parameter estimates at a representative voxel. After plotting the contrast of interest the variable SE in working memory represents the standard error of that contrast. The probability of detecting an activation of A% at a specificity of defined by a T threshold u (given the standard deviation of the parameter estimate is SE) is (I think):

1 - spm_Ncdf(u*SE,A,SE^2)

This is the power or sensitivity.

For example if your corrected threshold is 4.2, SE = 2.6 and the grand mean of your data was 100, then the power to detect a 10% activation would be:

1 - spm_Ncdf(4.2*SE,10,SE^2) = 0.3617

i.e. 36% power. I am not sure how rigorous this is (I am not very expert in this) but you could certainly check this with your local statistician. [Karl Friston 26 Jun 2000]

I have a question regarding the power of SPM results. Specifically I have FDG PET studies obtained in 17 adult controls and 7 children older than 6 year of age. I performed a routine SPM analysis at the 0.05 level and no significant differences can be detected. The reviewer however argues that this is a null finding of group differences between small groups and thus I cannot conclude that subtle differences do not exist. How can I determine what power my results have? For voxel-wise power estimates, one can apply the non-central t or F distributions - see for example Van Horn et al, NeuroImage 7, 97-107. I once wrote some software to do voxelwise power calculations on SPM96 analyses that may or may not be useful to you:

ftp://ftp.mrc-cbu.cam.ac.uk/pub/imaging/Power

However, I'm not sure whether the voxel-wise approach is answering the correct question. For example, an obvious question might be how much power you have to detect a change in a given brain region, and this question is more complex to answer within SPM, as even for a small region you are likely to have more than a single measurement's worth of data. You could of course reduce the problem by using regions of interest. [Matthew Brett 26 Jun 2000]

I know of at least three articles that deal with power in spatially extended statistical processes, in addition to articles from the FIL that address power in previous versions of SPM. Note that as some important aspects of SPM have changed in the latest version-- such as relaxation of the constraint that smoothness is equal at all voxels-- the calculations in these latter papers may not hold precisely. note that all these articles are available online at http://www.idealibrary.com free of charge, if your institution has a license.

Mapping Voxel-Based Statistical Power on Parametric Images John Darrell Van Horn, Timothy M. Ellmore, Giuseppe Esposito, and Karen Faith Berman NEUROIMAGE 7, 97–107 (1998) ARTICLE NO. NI970317

Factors That Influence Effect Size in 15 O PET Studies: A Meta-analytic Review Sherri Gold,* Stephan Arndt,* , ? Debra Johnson,* Daniel S. O?Leary,* and Nancy

3. Andreasen* NEUROIMAGE 5, 280–291 (1997) ARTICLE NO. NI970268

Estimation of the Probabilities of 3D Clusters in Functional Brain Images Anders Ledberg,1 Sebastian Åkerman, and Per E. Roland NEUROIMAGE 8, 113–128 (1998) ARTICLE NO. NI980336

not to mention SPM roots: Detecting Activations in PET and fMRI: Levels of Inference and Power

11. 10. FRISTON,A.HOLMES, J-B. POLINE,C.J.PRICE, AND C. D. FRITH NEUROIMAGE 40, 223–235 (1996) ARTICLE NO. 0074

and COMMENTS AND CONTROVERSIES How Many Subjects Constitute a Study? Karl J. Friston, Andrew P. Holmes, and Keith J. Worsley NeuroImage 10, 1–5 (1999) Article ID nimg.1999.0439

[Christopher Gottschalk 26 Jun 2000]

Dr Friston: you specify below data with a grand mean of 100-- does this apply to an analysis in which grand mean scaling was set to 100-- even though the raw data do not [usually] have this mean?

Yes. The grand mean is simply set to 100 arbitrary units.

Equally true if ANCOVA or PS were used?

It does not really matter. You simply have to specfiy your activation size in the appropriate units. These are usually adimensional and scaled to the grand mean specified.

and a further clarification: it appears what you propose applies to a given voxel or region, not over the SPM as a whole--true?

Absolutely.

The formula for the power you specified is, I assume, for the type II error (the beta)? In other words this is the false rejection rate (the probability to call a voxel as not significantly different between groups when they really are different)?

Strictly speaking the power is the probability of correctly rejecting the null hypothesis. This is 1 - p(type II error) conditional on the alternate hypothesis being true. The standard error should be the contrast of parameter estimate divided by the t value if this helps. [Karl Friston 27 Jun 2000]

Excerpts from SPM-help#: 8-Aug-100 SPM and Power by Kris [email protected].

Can anyone recommend a source for information on how to derive power estimates for fMRI analyses as implemented in SPM?

See these recent SPM list answers:

http://www.mailbase.ac.uk/lists/spm/2000-06/0193.html http://www.mailbase.ac.uk/lists/spm/2000-06/0191.html

(Search http://www.mailbase.ac.uk/lists/spm/search.html for "power" for more.) [Thomas Nichols 8 Aug 2000]

Back to top.

10. fMRI Time Modeling

I am working on a paper and I need reference literature which explain two functions I used. I was using the 'mean & decay function' for my experiment to describe time within-epoch adaptation.

Here you use 2 basis functions, which should capture an underlying within-epoch adaption. I would look at it as a specific case of the general linear model as implemented in SPM99. This could be covered by the following two references: The first is about the general linear model approach, providing the framework for the core of SPM.

---

Friston KJ Holmes AP Worsley KJ Poline JB Frith CD and Frackowiak RSJ Statistical Parametric Maps in functional imaging: A general linear approach. Human Brain Mapping 1995;2;189-210

---

The second is about basis functions for epochs, where your mean & exponential decay function would just be another set of basis functions.

---

Friston KJ Frith CD Turner R and Frackowiak RSJ Characterizing evoked hemodynamics with fMRI. NeuroImage 1995; 2;157-165

---

Also I used 'parametric linear modulation over time' to describe a general adaptation for ongoing stimulus tasks within a session.

The parametric linear modulation over time is implemented by a convolution of the chosen basis function set with a series of 'stick functions', where the height of each stick is modulated over time. Essentially, this approach is largely equivalent to the one described in the following paper, where the authors specify basis functions, which are modulated versions (up to 2nd order) of a stimulus function.

---

Buchel C Wise RJS Mummery CJ Poline J-B Friston KJ Nonlinear regression in parametric activation studies. NeuroImage 1996;4:60-66 [Stefan Kiebel 27 Jun 2000]

I need to explain where in the analyses of the fMRI images the intrinsic correction for temporal autocorrelation occurs. I've not been able to figure this out on my own. (sorry if this question is a little daft)

In SPM, the temporal autocorrelation is taken into account at the parameter estimation stage and in the statistical inference.

The design matrix and the data are both convolved with a filter kernel, which is usually a bandpass filter, i.e. it is effectively a combination of a user-specified lowpass and highpass filter. This changes the autocorrelation structure of the data such that the actual autocorrelation structure is given by convolution of the (unknown) intrinsic autocorrelation with the bandpass filter kernel. One goal of this filtering is to impose an autocorrelation structure on the data, which is not too different from the assumed autocorrelation (s. below).

At the level of statistical inference: to compute a t-value at each voxel, one has to estimate the intrinsic autocorrelation of the data. Currently, in SPM99, you can do that by assuming that the intrinsic autocorrelation before the convolution with the bandpass filter kernel is a unity matrix or by estimating the autocorrelation with an AR(1)-model. Anything what follows at this stage, e.g. the computation of the effective degrees of freedom is based on these estimates of the intrinsic and actual autocorrelation structures.

Some part of all this is described in KJ Worsley and KJ Friston, 1995. Analysis of fMRI Time-Series Revisited

• Again. Neuroimage, 2:173-181

As far as I know, a paper by Karl Friston et al. about temporal filtering is in press (NeuroImage). [Stefan Kiebel 11 Jul 2000]

Back to top

5. Data Acquisition Pointers

1. fMRI Susceptibility Artifacts

Dear SPM users, I received a number of highly informative responses to my question about temporal lobe susceptibility effects. Since I assume that they may be of general interest, I list the collected responses below. Thanks again. [Peter Indefrey 10 Jul 2000]

Peter Indefrey wrote Q1)I'm currently considering the pros and cons of PET vs. fMRI for a language paradigm with the inferior temporal lobes as the principal region of interest. While my preference is PET , considering the known susceptibility problems of fMRI in this region, I'd nonetheless like to learn what can be done in fMRI to minimize (the effect of) magnetic field inhomogeneities. Any suggestions?

1) From: Jose' Ma. Maisog [email protected] Hi Peter, check out this abstract from the recent Human Brain Mapping conference in San Antonio. They used a post-processing statistical correction to minimize the effect of susceptibility artifact. Perhaps one of the authors can offer more advice.

Devlin J, Russell R, Davis M, Price C, Wilson J, Matthews PM, Tyler L, "Susceptibility and Semantics: Comparing PET and fMRI on a Language Task," NeuroImage Volume 11, Number 5, May 2000, Part 2 of 2 Parts, S257.

Q2) Peter Indefrey wrote:

maybe I should be more specific on this: it was apart from my own experience with PET and fMRI on similar paradigms just this talk that confirmed my opinion that PET would be the adequate thing to do, since the statistical correction did not seem to fully compensate for the susceptibility artifacts. So my question was rather: does everybody agree on the conclusions of this paper or are there procedures, post-processing or other, such as shimming, choice of slices and angles, that have proven effective in minimizing the problem to such an extent that fMRI is at least as good as PET in scanning ventral temporal regions.

2)From: Alejandro Terrazas [email protected] Peter- I have some experience with field corrections. I am writing up a paper comparing fieldmap corrected to non-corrected data. Clearly the raw data is improved but it is still not clear whether there is an improvement of the activation maps. For one thing, people use smoothing to "improve" their images and fieldmap corrections in spiral are like unsmoothing. Things are different for EPI sequences where you get geometric distortions. There are correction methods for this as well.

PET vs. fMRI is a tough question. It depends on the temporal dynamics of what you wish to see. PET is probably more reliable for deeper structures.

3) From: Matt Davis [email protected] Hi Peter, If you read the full paper that has just been published in NeuroImage:

Joseph T. Devlin, Richard P. Russell, Matt H. Davis, Cathy J. Price, James Wilson, Helen E. Moss, Paul M. Matthews, and Lorraine K. Tyler (2000) Susceptibility-Induced Loss of Signal: Comparing PET and fMRI on a Semantic Task. NeuroImage 11(6): 589^Ö600

you'll see references to three acquisition methods that have been proposed to alleviate susceptibility-induced problems in the anterior and inferior portions of the temporal lobe. These are:

1. tailored RF pulses:

Chen, N., and Wyrwicz, A. M. 1999. Removal of intravoxel dephasing in gradient-echo images using a field-map based RF refocusing technique. Magn. Reson. Med. 42:807^Ö812.

2. Z shimming:

Constable, R. T. 1995. Functional MRI using gradient echo EPI in the presence of large static field inhomogeneities. J. Magn. Reson. Imag. 5: 746^Ö752.

Yang, Q. X., Dardzinski, B. J., Li, S. Z., Eslinger, P. J., and Smith,

13. 2. 1997. Multi-gradient echo with susceptibility inhomogeneity compensation (MGESIC): Demonstration of fMRI in the olfactory cortex at 3.0 T. Magn. Reson. Med. 37:331^Ö335.

Yang, Q. X., Williams, G. D., Demeure, R. J., Mosher, T. J., and Smith, M. B. 1998. Removal of local field gradient artifacts in T-2*-weighted images at high fields by gradient-echo slice excitation profile imaging. Magn. Reson. Med. 39:402^Ö409.

Constable, R. T., and Spencer, D. D. 1999. Composite image formation in z-shimmed functional MRI. Magn. Reson. Med. 42: 110^Ö117.

3. Spiral scanning:

Crelier, G. R., Hoge, R. D., Munger, P., and Pike, G. B. 1999. Perfusion- based functional magnetic resonance imaging with single-shot RARE and GRASE acquisitions. Magn. Reson. Med. 41:132^Ö136.

As yet we have no practical experience of any of these methods are. There is a Constable paper in NeuroImage 12(1) showing that z-shimming allows detection of activations in the hippocampus, but to my knowledge there has been no equivalent demonstration for activations in the temporal pole. If the inferior (especially anterior) temporal lobe is a region of interest, I would agree with you that PET may be a superior imaging modality for your purposes.

If you hear of any other suggestions other than the ones referred to here I would be interested to hear of them.

4) From: Russ Poldrack [email protected] I didn't follow this whole thread, so I'm not sure if anyone has suggested these steps:

1 - reduce TE to increase the level of bold contrast in the regions where the dropout occurs - the tradeoff is that you reduce the bold contrast in other regions. at 1.5T you might try something like 25 ms as a tradeoff

2 - use smaller voxels (something like cubic 3 mm is probably best) - this definitely reduces the dropout

3 - different orientations may result in different levels of artifact - you should do test runs with different orientations and see which work best for your particular area of interest

I've not read the paper mentioned below, but I am skeptical that these artifacts can be overcome solely using post-processing stats.

5) From: L.K. Tyler [email protected] In our studies on semantic processing comparing activations in PET and fMRI (3T) using a similar paradigm and materials, we definitely find more robust inferior temporal lobe activation in PET than in fMRI. However, there are procedures you can use to maximise the signal in the temporal lobes in fMRI, although it's not yet clear how much this will improve things. For the time being, PET seems to have the advantage if you want to activate the ventral temporal regions.

6) From: Joe Devlin [email protected] From discussions with our physicists, it sounds like there are things one could do to reduce the macroscopic susceptibility artifacts -- particularly if you are specifically interested in one region. If the lateral surface of the ventral temporal region if the area you are most interested in, you can often use manual shimming to reduce the field gradients in that region. The problem is that this typically distorts the signal from other areas but if you are doing an ROI analysis, this may not be important. Obviously, using a small volume statistical correction for this region will further increase your sensitivity.

It may be worthwhile piloting your experiment with one or two subjects in fMRI and looking at their data before deciding further.

转载于:https://www.cnblogs.com/minks/p/6268746.html