In general, MEG experimental design is no different from other types of design; successful designs will derive from specific hypotheses concerning brain function. In particular, users familiar with FMRI design will find that typical "box-car" designs may be used almost unaltered in MEG. However, there are several significant differences between FMRI and MEG, and successful designs must also adhere to the special constraints of MEG.
In addition, certain types of analysis do better with certain types of design. Thus, the anticipated type of analysis can and should play a role in the design of the experiment.
It depends on the modality. The stim computer (Ika), can for example trigger the Grass stimulator, which can deliver pulses with durations measured in tenths of milliseconds. As for auditory, in the click task we're doing, we're using a 5 ms click, which is considered long, as clicks go. At a 44 kHz sampling rate (typical for auditory stimuli), one sample is about 23 microseconds, so we can go pretty fast. A 1 ms click would be very doable. Note that the plastic tubing used to deliver the stimuli only has a bandwidth of about 5 kHz, though. For visual, the graphics card runs at 60 Hz usually, so you can theoretically leave a stimulus up for 16 ms, but the LCD projector's slow response time will limit that a little-—it won't turn completely on in that amount of time so the image will be dim. It takes more like 2 frames (32 ms) to go from all black to all white. You can increase the refresh rate of the card to perhaps 80 Hz, but that won't speed up the LCD projector.
The MEG system can sample at rates up to 6 kHz easily, with a 1500 Hz bandwidth. It is rumored that bandwidths of 3000 Hz are also possible.
The brain, scalp, and other tissues of the head are largely transparent to magnetic fields. Thus, the magnetic fields generated by electrical currents within the brain emerge from the head with almost no distortion. This is in contrast to EEG, where the electrical conductivity of the tissues creates a blurring of the signal. For more information on the source of the signal, please see the Lead Field page.
Because of the low distortion, high spatial resolutions are possible, although this will in general depend on the type of experiment being carried out. Stimulation of a small patch of skin, for example, can be expected to generate activity in a relatively small area of sensorimotor cortex. Tasks with high cognitive loads, such as working memory or semantic judgement tasks, can be expected to activate relatively large areas of cortex.
The type of analysis you do will depend on the type of signal you expect. If the signal is likely to be highly focal, and time-locked to a stimulus event, a simple dipole fit may work well. For cognitive tasks, a SAM analysis is more likely to succeed.
Dipole fitting and SAM are both examples of "source analysis" as opposed to "sensor analysis". That is, they both give you information about the sources inside the head, instead of just channels (which record magnetic field outside the head).
Besides that, they are totally different methods, used in different ways.
Dipole fitting is used with averaged data, at a specific time point or small window. You might be interested in the location of the generator of a stimulus-evoked field. Then you would do a dipole fit.
SAM uses the unaveraged trial data. Actually, it might be better to say that SAMsrc uses the covariance of the data (not the average). From the user's point of view, though, you do dipole fits on an averaged dataset and SAM on the raw (but probably baseline corrected, high-pass filtered) dataset. SAMcov computes the covariance matrix.
So SAM uses second-order statistics (variance), while the dipole fitter operates on first-order statistics (mean). Both are useful methods for doing source analysis.