(2006) to improve confidence in reverse inference. The first is to increase
the selectivity of the activation pattern for the cognitive process
of interest. Focusing on sets of smaller regions instead of a single large
region is likely to improve selectivity, because specific processes most
probably engage specific functional networks and subloops within
them (Stiers et al., 2010). It is clear that this suggestion is maximally
met in multivariate pattern analysis. The training identifies an optimal
number of the smallest spatial units with a selective response towards
the processes of interest, without the need for a priori region selection
based on theoretical assumptions. Poldrack's second suggestion was to
increase the prior probability that the cognitive process of interest is actually
engaged by the cognitive task studied — i.e., if we are certain that
the task engages the cognitive process, the prior probability is 1.0. This is
under control of the experimenter. We designed the overt recall task
used here as reference task in such a way that the immediate and long
term trials only differed in the (hidden) neural source from where the
targetwordwas retrieved. As a consequence,we zoomed in on the processes
of retrieval regardless of their content or outcome (being successful
or not) and confirmed their operation in a thorough validation
study.
An alternative interpretation of our results might be that the classification
doesn't reflect retrieval specific differences, but merely a generalized
difference between immediate and long term trials, such as for
instance higher effort in the latter. Neither the univariate results, nor
the multivariate results agree with this interpretation, however. In univariate
analyses task-general effects are associated with a typical distribution
of activations and de-activations in the attention/salience and
defaultmode network, respectively (see for instance Fig. 2A). The distribution
of voxels that differentiate between immediate and long term
recall trials shows only a limited amount of overlapwith this typical activity
pattern (Fig. 2C). This suggests that retrieval specific effects are
being observed. Moreover, some differentiating regions showed effects
in the direction opposite to the direction of the presumed generalized
effect (for instance, cluster 3 in Fig. 5B–C). In the multivariate results,
on the other hand, voxels contributing to the classification fell in large
part outside of the regions showing a univariate difference between immediate
and long term recall trials (Fig. 4, left panel). Moreover, classification
accuracies were high at all iterations of recursive feature
elimination (Fig. 3B) and even without univariate feature reduction
(Figure S1A), suggesting that the result is not critically dependent on
any specific subset of regions. This was confirmed by a follow-up
MVPA analysis inwhich all voxels showing a significant univariate effect
(p b 0.001, uncorrected for multiple comparisons) had been eliminated
from the feature matrix. This analysis yielded classification accuracies
nearly identical to those in the original analysis (see Figure S3).
While multivariate pattern analysis was effective in decoding cognitive
processes, the interpretation of the spatial map of contributing
voxels is less straightforward than in a traditional univariate analysis
(Jimura and Poldrack, 2012; Poldrack, 2011). First of all, farmore voxels
contribute to the classification than would be expected from a univariate
approach. This is due to the higher sensitivity of the multivariate approach
(e.g., DeMartino et al., 2008; Haxby et al., 2001), but in our study
it most likely also reflects individual differences in strategy. Secondly,
the voxels identified as contributing to a classification are not all the
voxels that carry relevant information. This it clear from the fact that
in the recursive elimination process a wide range of iterations, each
with a different subset of voxels, yielded high classification accuracies.
Thirdly, that a voxel contributes to classification still doesn't tell us