Limitations of t-maps

(1) Spatial imprecision after smoothing: Spatial smoothing has a nonlinear effect on the voxel variances. For example, when smoothing over two regions with a positive contrast but independent error terms, the minimum variance is found between these regions.
index_html_m5090030b
This may lead to sometimes-counterintuitive artifacts in t-maps, especially after smoothing over different types of tissue with different variability (e.g. gray matter, white matter). Most white matter findings can be explained by this effect. It is important to understand that only t-maps (t-map = 'signal-to-noise' image) are affected, but not contrast images (contrast = 'signal' = the estimated parameter that was divided by its standard error to obtain the t-value).
Figure (worst case): 2D-Simulation of three black squares (5*5pixel) representing gray matter, surrounded by white matter. A positive contrast Δµ=1 is present in the upper and the lower region, in the middle, there is only noise, no signal. Note that after smoothing with a FWHM = 14pixel, the nearest gray matter of Tmax is the region with Δµ=0!

(2) Another question is that of which quantitative measure the investigator is interested in most. A statistical measure, for example, a t-value, allows to test if an estimated parameter (=contrast) differs significantly from zero, that is, if a 'true effect' has been observed. In other words, the t-map reflects the 'signal-to-noise' ratio, where 'noise' comprises measurement errors as well as authentic between-subject variance that is not explained by the statistical model. However, if there is no doubt about the significance of a finding, the investigator may be interested in the signal itself (e.g. a between-group difference of receptor availability) rather than a measure that depends on the local noise.

(3) The strength of the standard SPM procedure is to detect significant findings. A more difficult if not impossible challenge is to determine the exact outline of an effect. Many people choose the original voxel-level threshold to define the displayed region, however this is not without problems: first, any t-threshold is subject to the same artifacts as described above. Second, this threshold may be too conservative with respect to the outline of a finding and may thus reveal only the tip of the iceberg. What is the a priori likelihood for a voxel next to a (statistically proven) true effect? I don't know, but I believe that - once a cluster is significant - we can afford more liberal thresholds for illustrative purpose.