verb-generation experiments in the domain of information used (language vs. spatial/motor), output modality (verbal vs. manual), and sensory information used (visual vs. kinesthetic/somatosensory).

One final introductory point worth making: The examples that will be presented here are of the effects of practice on a set of items, or on a particular learning instance. As such, they do not explicitly represent the development of a general skill, like typing or playing tennis. We will argue, however, that the nature of results themselves suggests relevance to skill acquisition in a more general way than might be thought given the item-based nature of the practice effects.

Background/Experimental Design. In an early set of studies (11, 12), subjects performed a hierarchy of four single word processing tasks. These tasks were done with both visual and auditory word stimuli. During separate scans, subjects did the following: 1, fixated on a centrally displayed fixation point; 2, fixated and passively viewed or heard nouns; 3, fixated and repeated the seen or heard nouns; and 4, fixated and said aloud a verb appropriate for heard or seen nouns.

In that original study, a hierarchical subtraction design was used to identify regions at different “levels” of the processing of single words. In other words, in each modality, task 1 was subtracted from task 2, task 2 from task 3, etc. and only positive changes were reported (see Table 1). This produced some interesting results, but had some design problems, which will be discussed below.

Results. In the first subtraction (passive presentation minus fixation), activation was seen in modality-specific primary and nonprimary sensory processing regions. For visual input, bilateral primary and extrastriate regions were clearly activated. For auditory input, bilateral primary auditory and auditory association areas were active, as well as a region at the left temporoparietal junction. By virtue of the localizations and the tasks involved these activations were attributed to modality-specific processing of the word stimuli. These experiments could go no further in assessing the type or specificity of that processing [although further work on this issue has been done (16, 17)].

In the second subtraction (word reading or repetition minus passive presentation), areas commonly related to motor processing were seen. Bilateral primary motor and insular cortex, premotor, SMA, and medial cerebellum were active irrespective of the modality of stimulus input, and these activations were attributed to the common output demands of the reading and repetition tasks.

In the final subtraction (verb generation minus word reading or repetition), activation was seen in the left prefrontal cortex, the anterior cingulate, and the right cerebellar hemisphere. Again, these activations were seen irrespective of modality of input, and they were attributed to the additional processing demands of the more complex verb-generation tasks.

“Problems” with “Cognitive Subtraction.” In several instances in the literature, hierarchical subtraction designs such as described above have been criticized as having inherent assumptions that may well not be met. The most problematic of these assumptions is that of “pure insertion.” This difficulty is not new with the use of subtraction in functional imaging, and indeed it has been discussed for several decades in psychology, particularly in dealing with certain types of reaction time studies (e.g., ref. 18).

The problem with pure insertion can be conceptualized like this. Take two tasks—for example, seeing a word and saying it out loud (word reading) and seeing a word and saying an appropriate verb related to that word (verb generation). Both tasks entail visual processing and motor output, and processes to translate the visual input into a motor output. When the reaction times for these two tasks are studied, it turns out that word reading on the average takes about 550–600 ms, and verb generation about 1000 ms (15). One way of interpreting these results is that all of the processing that occurs during word reading also occurs during verb generation, plus some more processing that takes about 400 ms. This interpretation assumes “pure insertion” in that the extra processing that is related to the verb generation task is just added to that of the reading task.

The problem with such a subtractive design and interpretation in a reaction time experiment is that it is quite possible that certain processes that occur during the word reading task may not occur during the verb generation task. In other words, some of the processes used in word reading are replaced when verb generation is performed. It is the sum of the time saved by dropping some processes used in simple reading and the time for additional processes for verb generation that in total contribute to the longer reaction time. By just looking at reaction times for these two tasks, the two alternative explanations cannot be deconfounded. Different approaches to this issue in cognitive psychology have led to more complex experimental methods, such as additive factors and factorial designs, that allow the investigation of process interactions in two tasks. Some of these approaches have also been used in imaging research (e.g., ref. 19).

The hierarchical subtraction analysis used in the imaging study presented above “buys into” this problem, because the design of the analysis and the interpretation drawn from it uses pure insertion logic (and assumptions).

However, this is where the problem ends and confusion in the imaging literature begins. The confusion results from the lack of appreciation for the distinction between “cognitive subtraction” as an experimental design and interpretive strategy, and image subtraction as an analysis methodology. As seen above, cognitive subtraction can be used to design and interpret an imaging study, and imaging subtraction can be used to mirror the cognitive strategy. However, image subtraction does not make the assumption of pure insertion: experimental designs, analysis choices, and interpretive strategies do. Image subtraction is performed in part to mirror experimental design strategy, but more importantly it is done to reveal the differences in the hemodynamic signal between two conditions by subtracting the large amplitude complex anatomical background present in hemodynamic images.

In recognition of the interpretive problems with the early analysis presented above. Fiez et al. (13, 14) reanalyzed the