 |
 |
Dear Priyantha, I noticed no-one else had answered this so I thought I'd give it a quick go. 1. I know that there are 4 different options to model the response function in the SPM99b.. namely... Fixed box car, Discrete Cosine set etc. I noticed that the first two, DCS and the mean and expo. decay models additional regressors, where as box car and half sine functions do not. How does one select what response function is the best for a particular paradigm?I realize that the fancy stuff like DCS etc are more suitable for efMRI, but for the usual blocked designs, what is best? None of the suggested options are always the 'best' model to use: they each depend on the form of the neuronal response to your paradigm. The 'box-car' function assumes that an immediate, tonic response results from your behavioural intervention (i.e. your experimental task/stimulation), and ceases more-or-less immediately when it does (hence its shape and name). However, although this may hold for the primary sensory cortices, other models may be more appropriate for different spatial locations and paradigms. For example, use of the half-sine regressor as your stiumulus waveform produces a smoother boxcar with a gradual rise and decay, and the mean and exponetial decay option allows you to model hemodynamic responses which decay over the period of the task. This can be understood in the context of a neuronal response which undergoes habituation to repeated stimulation, for example. If you play with the 'fMRI design' button, and review the design matrices afterwards, you can see these changes in detail, and also what they look like after convolution with the hrf. To sum up, each regressor is essentially asking a different question - you have to choose which one is appropriate for you. 2. I also noticed that depending on modelling intrincic correlations of the time series and many other manipulations that one does,the effective DF for a given contrast varies, although the number of parameters estimated in the model is the same in both situations ( ie AR+ or not etc..). How is this possible and what does it mean in terms of the outcome of the analysis? This is an example of where fMRI differs from PET. In PET, each scan is treated as an independent observation. In fMRI the presence of temporal autocorrelations violates this assumption (basically, each scan is part of a timeseries, and so the correlation between successive terms is not zero, and so they are not independent). The SPM approach is to use 'temporal smoothing (old school term)' or 'low-pass filtering (new school term)', and, by extension, the use of first-order auto-regressive models (the AR(1) option) to overcome this problem. There are a number of postings in the mailbase discussing the theory motivating this approach in more detail. Basically, as the degrees of freedom are affected by the number of independent observations in the data, and changing the autocorrelation structure (by filtering/smoothing/whatever) will effect this, I'm sure you can see how your dfs change. 3. In general, how does one asses the goodness of the fit in a particular model for any given paradigm? Having modelled the data set in many different possible ways, what factors should favour the selection of a particular model and the emanating SPM t maps over the others? This is known as a 'model selection' problem, and basically boils down to a kind of Hobson's choice: should I attempt to include every component of interest in my model, and so model the experimental variance well but lose dfs - or vice versa? This problem is usually dealt with is by adopting one of two strategies: forward or backward selection. In forward selection, one takes a very sparse model and adds regressors to it until you can no longer explain a significant amount of variance by the inclusion of your terms into the model. A simple example is the use of a fourier basis set to model hemodynamic responses: although using a basis set removes one from a priori assumptions about the shape of the hrf, it is always problematic to assess how many basis functions one should use - the inclusion of successive terms could be tested by model selection. Backward selection is the exact opposite: you start with a' megamodel' and assessing the effects of removing regressors in a similar fashion. An extra sum-of-squares F statistic is used to assess significance in both cases. The topic is well described in 'the book' (Human Brain Function, on pages 78-80) by Andrew H, and a good recent example is contained in the paper by Aguirre and colleagues in Neuroimage, looking at the inclusion of subject-specific regressors in a study examining the variability of hemodynamic responses. The approach differed slightly, however, in that the authors employed a partial F test, not an ESS F (apologies if I've remembered that wrong!). The above is a 'bare bones' approach to these questions - further comments are encouraged. Best Dave McG.