What are the units of a plot in SPM?
In SPM, when you click on “Plot” and ask for a “Contrast estimates and 90% CI”, what exactly are the units of the thing plotted? I get this question so many times I wanted to record my answer once and for all.
The short answer is: 2-3 times an approximation to percent BOLD change, if you have a long-block design; otherwise, it depends. Here’s the long answer.
The final units of a plotted contrast of parameter estimates depends on three things:
- The scaling of the data.
- The scaling of the predictor(s) that are involved in the selected contrast.
- The scaling of the selected contrast.
The scaling of the data
For a first level fMRI analysis in SPM the scaling is done “behind the scenes”… there are no options to change the default behavior: The global mean intensity for each image is computed, and the average of those values (the “grand grand mean”) is scaled to be equal to be 100. However, SPM’s spm_global produces a rather dysmal estimate of typical intracerebral intensity (note that the author is listed as “Anonymous”; to be fair, it worked well on the tightly-cropped, standard-space PET images that it was originally designed for). The estimate of spm_global is generally too low, and so when the data are divided by it and multiplied by 100, the typical brain intensity is well over 100. In my experience, if you examine a beta_XXXX.img corresponding to a constant in a first-level regression model, typically gray matter values range from 200 to 300 when it should be 100; for first-level analyses done in subject-space (where there are even more air voxels) the bias can be even greater. Let’s call the typical gray matter intensity of your BOLD T2* images to be B for later correcting this effect (e.g. B = 250 or 300 or whatever you find).
At the second level, you depend on the contrast estimates passed up from the first level to have a valid interpretation. So all three of these rules should apply to the first and second level models, but the data scaling is done at the first level (no further scaling should be done at the second level; e.g., in the Batch Editor, second-level models have “Global normalisation -> Overall grand mean scaling” set to “No”).
The scaling of the predictor(s) that are involved in the selected contrast
At the first level, the predictors must be scaled appropriately to give the regression coefficients the same units as the data. For BOLD fMRI, this means the baseline to peak magnitude should be 1.0… i.e. a one unit change in a beta coefficient will result in a one unit change in predicted BOLD effect. In SPM, for loooong blocks/epochs, 20 seconds or longer, the scaling is as expected, with the baseline-to-plateau difference being 1.0. For shorter blocks however, the peak can be higher – as this corresponds to the ‘initial peak’ part of the HRF, as the duration is too short for the flat plateau to be established (this ‘early peak’ is due to symmetry of convolution, as it must exactly mirror the post-stimulus overshoot; there is some physiological basis for this as well, though) – or they can be lower. For events, the HRF is normalized to sum to unity, which doesn’t help you predict the peak.
If you have anything but long blocks, then, you have to examine the baseline-to-peak (or if you prefer, baseline-to-trough) height, and make a note of it for later scaling (call it H). (The design matrix is in
SPM.mat; plot individual columns to judge the scaling). For event-related designs, however, it can be tricky to uniquely define the baseline-to-peak height; if you have a slow design, with isolated events, you’ll find a different peak height than if you have a design with some closely-spaced events as the responses pile-up additively. A reasonable unique definition of a event-related response magnitude is the height of an isolated event; again, call this H.
At the second level, you must also make sure that your regressors preserve the units of the data. For example, if you are manually coding dummy regressors, like gender, as -1/1, realize that the unit effect of beta coefficient for gender will result in a two unit change in the data. Safer is using a 0/1 coding, though that then re-defines the intercept; best yet is to use -1/2 and 1/2 as dummy values for a dichotomous variable.
For parametric modulation and covariates, note that a regression coefficient’s units are determined by the units of the covariate. For example, if you have a parametric modulation in a reward experiment, where the units are £’s that are at risk in a bet, then your regressor coefficient has units (if everything else is working as it should) “percent BOLD change per £ at risk”. Likewise, if you use an age predictor in the second level, if you enter age in years (centered or uncentered), the units of the age coefficient are “percent BOLD change per year”
Finally, beware of trying to compare between percent BOLD change in Block Designs vs. Event Related designs. Due to the nonlinearities of the BOLD effect, a percent BOLD change for an event related response is really a different beast from a percent BOLD change for an block design. When reporting one type, be sure it is clear what sort of effect you’re reporting, and don’t try averaging or comparing their magnitudes.
The scaling of the selected contrast
The General Linear Model is a wondrous thing, as you’ll get the same T-values and P-values for the contrast [-1 -1 1 1] as you do for [-1/2 -1/2 1/2 1/2]. However, the interpretation of the contrast of parameter estimates (contrast times beta-hats) is different for each. The following rules to scale contrasts will usually work to ensure that a contrast preserves the units of the regression coefficients:
- Sum of any positive contrast elements is 1.
- Sum of any negative contrast elements is -1.
If the contrast elements are all positive or all negative, the interpretation is clear, as you are ensuring that the beta-coefficents (each which should have unit-change interpretations) will be averaged, giving a (average) unit change interpretation. If there are a mix of positive and negative elements, I see this rule as carefully ensuring there is a positive unit change quantity subtracted from a negative unit change quantity, giving the expected difference of unit change. (This rule should work for covariates and simple 0/1 dummy variables; this FSL example shows a case with -1/0/1 dummy variables where the contrasts violate these rules yet actually give the correct units.)
When it goes right, when it goes wrong
If everything had gone well, then the plotted units are “approximate percent BOLD change”. The “approximate” qualification results from only using a global-brain mean of 100; an “exact” percent BOLD change requires that the voxel in question has had baseline (or grand mean) intensity of exactly 100. We don’t grand-mean-normalise voxel-by-voxel because of edge effects: At the edge of the brain, the image intensity falls off to zero, and dividing by these values will give artifactually large values.
Of the three scalings in play, the first two of these three items are basically out of the control of the user. As hinted, first-level fMRI data in SPM are generally mis-scaled, resulting in global brain intensity between 200 and 300, i.e., 2 to 3 times greater than anticipated. Also, if you don’t have long blocks, you’ll need to apply a correction for the non-unity H you measured. The corrections are as follows: Dividing the reported results B/100 (i.e. multiplying by 100/B) will correct the problem of the data being incorrectly scaled. Multiplying by H will correct the units of the regression coefficients. After both of these corrections, you should get something closer to approximate percent BOLD change.
This all sounds rather depressing, doesn’t it? Well, there is a better way. If you use the Marsbar toolbox it will carefully estimate baseline intensity for the ROI (or single-voxel ROI—make sure it isn’t on the edge of the brain though!) and gives exact percent BOLD change units.
Postscript: What about FSL?
FSL has an good estimate of the global mean, because it uses morphological operations to identify intracerebral voxels. It scales grand mean brain intensity to 10,000, however (because it used to use a integer data representation and 100 would have resulted in too few significant digits.) Block-design predictors are scaled to have unit height; the default HRF doesn’t use a post-stimulus undershoot, and so baseline-to-peak height is the same as trough-to-peak height. Event-related predictors are not scaled to have peak height of unity; however, the recommended tool for extracting percent BOLD change values,
featquery, estimates a trough-to-peak height and then appropriate scales the data. However, as Jeanette Mumford has pointed out, this isn’t really the right thing to do when you have closely spaced trials; see her excellent A Guide to Calculating Percent Change with Featquery to see how to deal with this.
Thanks to Jeanette Mumford & Tor Wager for helpful input and additions to this entry, and Guillaume Flandin for the additional references.