>Henson (2006) explains Forward Infrence or Reverse Association
>also called
>Function-to-Structure Deduction (Henson, 2005).
>[r1, r2 = regions; c1 and c2 = conditions of interest which
>are different
>for one function; c0 = condition of control]
>Very briefy there are three criteria for Forward Inference: 1)
>interaction
>R(r1, r2) × C(c1, c2) suggesting a dissociation between r1 ed r2; 2)
>positive correlation between r1 and r2 for the contrast c2 and/or c1
>compared with c0 suggesting an assocition betwee r1 and r2; 3) negative
>correlation between r1 and r2 for the contrast c2 compared with c1.
Let me preface my views by noting that the default SPM position is that one
should not test region-by-condition interactions, primarily because it is
likely that different regions have different mappings from neural activity
(what one is really interested in) to BOLD (what one measures). (It is also
the antithesis of the massive univariate approach that is SPM!)
Nonetheless, the criteria I outlined in those papers were primarily to
overcome such potential differences in the neural-BOLD mappings in different
brain regions: the only assumption made being that those mappings are
monotonic (either increasing or decreasing). (Indeed, these criteria conform
to what psychologists call a "reversed association").
>Are the calculations related to the three criteria done on the % signal
>change for c0, c1 and c2?
Short answer is it doesn't matter, so it's probably simpler to do on the
contrast/beta values. In SPM, % signal change is derived from the betas (and
basis functions), with a grand mean scaling that is constant across voxels
(ie regions). Scaling can affect additive interactions, but cannot undo
crossover interactions. (So even voxel-specific scaling shouldn't matter).
>Is it correct to extract % signal change for c0, c1 e c2 from
>the same t-map comparing c2-c1?
Good question. Ideally the regions should be defined independently, eg on
contrasts orthogonal to those that define a reversed association. Otherwise,
defining r1 by the contrast c2-c1 and r2 by the contrast c1-c2 will bias you
to finding a reliable crossover interaction.
>Can I use a condition of rest (look a cross) explicitly modeled as c0 or
>Forward Inference is invalidated?
Rest (eg fixation) is fine as c0. Anything will do!
>Can I derive a Forward Inference when the three criteria are
>satisfied but some or all % signal changes for c0, c1, c2 are negative?
Yes. The criteria (and apologies that they have not always been spelt out
clearly) are 1] crossover interaction between c1,c2 and r1,r2 (ie c1>c2 in
r1 and c2>c1 in r2, or vice versa), 2] common activation/deactivation of c1
or c2 vs c0 in both r1 and r2 (eg c1>c0 in both r1 and r2). The signs of
contrasts c1 and c2 vs c0 do not matter (as long as 2] is met). I think.
Best wishes
Rik
--------------------------------------------------------
DR RICHARD HENSON
MRC Cognition & Brain Sciences Unit
15 Chaucer Road, Cambridge,
CB2 7EF England
EMAIL: [log in to unmask]
URL: http://www.mrc-cbu.cam.ac.uk/~rik.henson
TEL +44 (0)1223 355 294 x522
FAX +44 (0)1223 359 062
MOB +44 (0)794 1377 345
--------------------------------------------------------
|