FAQ.md

@@ -120,8 +120,7 @@ No, the basic functionality of PhysIO, i.e. creating nuisance regressors for you
 Currently, PhysIO natively supports the following physiological logfile types:
 
 - Brain Imaging Data Structure (BIDS)
-    - [Standard for peripheral recordings]
-    (https://bids-specification.readthedocs.io/en/stable/04-modality-specific-files/06-physiological-and-other-continous-recordings.html)
+    - [Standard for peripheral recordings](https://bids-specification.readthedocs.io/en/stable/04-modality-specific-files/06-physiological-and-other-continuous-recordings.html)
     - both raw physiological traces and pre-computed pulse events are
     supported
 - BioPac formats
@@ -304,7 +303,36 @@ The most likely explanation, however, is that the setup of PhysIO on your machin
     - If yes, try `rmpath(genpath('path/to/spm12'));addpath('path/to/spm12')` and run the example again.
     - More details on this issue can be found in our [GitHub Forum](https://github.com/translationalneuromodeling/tapas/issues/166).
 
-## 17. I cannot find the answer to my question in the FAQ. Whom do I ask for help?
+
+## 17. Which models do I have to include in my physiological regressor matrix? And which number of regressors (model order / delays) per model?
+
+The question about how many and which regressors you need for successful noise removal depends a lot on your experimental design and your research questions. Here are a couple of rules of thumb:
+
+1. To have sufficient degrees of freedom to estimate the parameters of the general linear model (GLM), it is often recommended to have a least 10 times as many datapoints as regressors in your model. That means, for 26 regressors in the GLM, it is advisable to have 260 volumes or more.
+2. 26 regressors is the standard output if you include 
+    1. RETROICOR (3rd order cardiac = 6 regressors (1 cosine and one sine per order), 4th order respiratory = 8 regressors, and 1st order multiplicative terms (4: cos*cos, sin*sin, cos*sin, sin*cos), 
+    2. HRV (1 regressor), 
+    3. RVT (1) and 
+    4. motion realignment parameters (6)
+    5. This list also gives an indication of the order of the regressors in the matrix, but see [question 10](https://tnurepository.ethz.ch/physio/physio-public/-/wikis/FAQ/edit#10-what-is-the-order-of-the-regressor-columns-in-the-multiple-regressors-file) for details.
+3. You can now reduce the number of regressors by the following observations:
+    1. The RETROICOR 3/4/1 order is taken from a paper that optimized physiological noise removal in the brainstem [1]. You might not need the full model, if you are interested in other brain areas
+    2. For example, in our PhysIO paper, figure 9 [2], where we evaluated this model for 35 subjects, you   can see that the multiplicative terms explain noise mostly in the midbrain, acquaeduct and more inferior parts of the brainstem.
+    3. Furthermore, also in that figure, you can see that the cardiac RETROICOR terms explain most of the variance (temporal SNR (tSNR) gains of up to 70%, compared to only motion correction), whereas the effect of both respiratory and multiplicative terms is one order of magnitude smaller (5% and 3% tSNR gains, respectively).
+    4. So, you could probably leave out the multiplicative terms and reduce the number of regressors by 4, or reduced the respiratory terms to 2nd or 1st order, reducing them by 4 or 6, respectively. I have also seen 2nd cardiac order models, reducing by a further 2 regressors.
+    5. In total, you would end up with 14 regressors for a RETROICOR 3/2/1, 10 regressors for RETROICOR 3/2/0 and 8 regressors for 2/2/0. In each case, you would still add the 8 regressors from HRV, RVT and motion.
+4. To get a first idea which sets of regressors contribute to noise removal for your data, you can run F-Tests over the columns for the respective regressors only (see also [question 11](https://tnurepository.ethz.ch/physio/physio-public/-/wikis/FAQ/edit#11-how-do-i-know-whether-the-physiological-noise-correction-worked) and see where in the brain they explain significant variance (you can also use tapas_physio_compute_tsnr_gains to compute tSNR as in our figure).
+    5. To formally compare whether it’s warranted to include a set of regressors into your GLM, you would have to do model comparison (as in [1]). There is a toolbox for SPM that can do this automatically for the kind of regressor matrices that are output by PhysIO [3].
+
+
+[1] Harvey, A.K., Pattinson, K.T.S., Brooks, J.C.W., Mayhew, S.D., Jenkinson, M., Wise, R.G., 2008. Brainstem functional magnetic resonance imaging: Disentangling signal from physiological noise. Journal of Magnetic Resonance Imaging 28, 1337–1344. https://doi.org/10.1002/jmri.21623
+
+[2] https://www.sciencedirect.com/science/article/pii/S016502701630259X#fig0045 from: Kasper, L., Bollmann, S., Diaconescu, A.O., Hutton, C., Heinzle, J., Iglesias, S., Hauser, T.U., Sebold, M., Manjaly, Z.-M., Pruessmann, K.P., Stephan, K.E., 2017. The PhysIO Toolbox for Modeling Physiological Noise in fMRI Data. Journal of Neuroscience Methods 276, 56–72. https://doi.org/10.1016/j.jneumeth.2016.10.019
+
+[3] Soch, J., Allefeld, C., 2018. MACS – a new SPM toolbox for model assessment, comparison and selection. Journal of Neuroscience Methods 306, 19–31. https://doi.org/10.1016/j.jneumeth.2018.05.017
+
+
+## 18. I cannot find the answer to my question in the FAQ. Whom do I ask for help?
 
 We are very happy to provide support on how to use the PhysIO Toolbox. However, 
 as every researcher, we only have a limited amount of time. So please excuse, if