GLM FAQ
From VoxBoWiki
Contents |
Which covariates should be convolved with an HRF?
The goal in designing your covariates is to construct vectors that you expect to correlate with the BOLD signal in some way. The HRF is an estimate of how we'd expect neural activity to map onto BOLD signal. So for covariates that you expect to be related to neural activity (e.g., task conditions or performance covariates), convolve with the HRF to get the shape of the expected variation in BOLD signal due to that activity. For covariates that aren't related to neural activity (e.g., global signal or run effects), don't.
Which covariates should be mean centered?
If you have an intercept, mean centering a covariate doesn't affect that covariate — its beta, t value, and p value will still be the same. However, if it's not mean centered, it will affect the weight on the intercept. Basically, if you want the intercept to reflect the mean signal (e.g., for calculating percent signal change), then you generally want the rest of your covariates mean centered.
If you don't have an intercept, then we assume you know what you're doing and why.
For more details, try this Mean centering demo.
What does it mean if my GLM comes out all red (or all something)?
One common frustrating outcome from a GLM is the infamous all-red statistical map (new versions may change the color, but you get the idea). A typical cause of this is Mean scaling your data when you set up the GLM, even though you expect the mean to be a small number, near zero. When you ask VoxBo to mean scale your data, it divides the data from each run by a mean signal value, to help get rid of scaling artifacts. This can help your sensitivity for analyses with mutiple runs. But you can't divide stuff by zero, and in practice you can't even divide stuff by really small numbers. When you try, you'll get unstable results, which show up as all red (or yellow) in VoxBo. This typically comes up with second-tier analyses, and sometimes for perfusion data. There's no reason to mean scale those kinds of data anyway, so turn off mean scaling and you'll be happier.
When should I mean scale (aka mean norm) and/or linear detrend my data?
Both of these are options applied independently to each time series in each voxel from each tes file. The mean scale option divides each time series by its mean. The detrend option removes any linear trend from the time series. Both options are applied directly to the time series data at the time of the regression, so they are reflected in the regression results but they don't modify your data files. (You could apply either operation as a preprocessing step, but we don't currently make that option available in VoxBo.)
Mean scaling can remove scaling artifacts, the sort you'd expect if signal values from the scanner were multiplied by a different constant in each run (magnifying signal and noise). It's not the same as the effect removed by scan effect covariates, which can account for a different constant added to each run. Both are conceivable sources of unexplained variance, and it costs little to remove them when appropriate. Note that mean norming is pointless when you just have a single run. And you can't do it when you expect the mean of your time series to be close to zero (not typically the case for BOLD, but certainly the case for second-tier group analyses, and perhaps for perfusion data).
Linear detrend removes "drift" artifacts, which are partly but not perfectly captured by high-pass filtering. These artifacts can be caused by physical processes in the scanner (heating), and if your study isn't designed to detect a linearly increasing effect (it probably shouldn't be for BOLD), it's helpful to remove this useless and predictable source of variance.
Which frequencies should I filter out?
When constructing your GLM, you can filter out high frequencies, low frequencies, and arbitrary frequencies in the middle. Filtering properly can improve your sensitivity by removing a lot more noise than signal. But you can't go overboard, because the more frequencies you remove, the fewer effective degrees of freedom you have.
Things aren't that sensitive at the high end, since there is generally little if any power there in either your design or the data. Historically, we have recommended filtering out a single high frequency, which supposedly can catch some acquisition artifacts but shouldn't affect things substantively.
At the low end, it depends a bit on your design. For blocked or coarse-structured designs, it can be fairly simple. Your covariates of interest will generally have a well-defined first peak, and you should filter out all frequencies below that peak. If you have multiple covariates of interest, make sure to leave intact all the frequencies that are important for any of your covariates of interst. VoxBo's G matrix designer makes an effort to tell you where the power cut-off is in the frequency structure of your covariates. If it tells you frequencies 0-8 contain less than 1% of the power for a given covariate, then you can remove 8 frequencies (the zeroth frequency doesn't count) without harming the sensitivity of that covariate.
For event-related designs in which your comparisons are between different periods during a trial, you could in principle remove many low frequencies without removing anything useful, since your comparisons of interest lie at high temporal frequencies. However, the more frequencies you remove, the lower your effective degrees of freedom. After a certain point you're doing more harm than good. As a rule of thumb, we suggest removing 2-5 frequencies, which should get most of the low frequency noise.
For event-related designs in which your comparisons are between trial types (e.g., rapid event related designs), filtering is a slippery slope. Your design may have roughly equal power (or expected power) at all frequencies. In this case, it's worth paying close attention to the software-recommended cut-offs, but it's probably best not to remove more than one or two frequencies.
