High level thinking and reasoning: How the brain makes some kinds of inferences

Introduction

For the purposes of this chapter, let us mean by "reasoning" the act of reaching new conclusions (however tentative) on the basis of facts and suppositions already present in the mind. For example, concluding that it will rain today on the basis of weather forecasts and the appearance of the sky is an act of reasoning. The "new conclusion" in this case is the conviction of rain whereas the starting "facts and suppositions" are composed of the weather forecasts and the perceptions obtained by viewing the sky. Another act of reasoning is concluding that your keys are in the bedroom from the assumptions that they are not in the kitchen and that they are either in the kitchen or the bedroom. You’ve probably noticed that human reasoning is extraordinarily proficient, having arrived at conclusions that are the basis of the stunning technology that penetrates society. Human reason can also be breathtakingly obtuse, as you have also no doubt observed when your friends don’t agree with you (or by reading the newspaper and perceiving what a mess is being made of human affairs).

Both the virtues and defects of human reasoning have no parallels in the rest of the animal kingdom. So this commonplace activity --- drawing inferences --- must count as one of the great miracles of life. Just as much as perception, language, and motor behavior, it is reason that allows us to interact successfully with the physical and social environment. Crucial to our experience, and virtually a defining feature of the human species, you would think that it would be among the first cognitive functions to have been explored and illuminated by Cognitive Neuroscience.

Alas, such is not the case. Little is known about how the brain produces an inference or evaluates the credibility of one that is presented to it. Part of the difficulty of conducting research in this area is the complexity of the task accomplished during inference. The reasoner must comprehend the facts and suppositions that get the inference started then fix upon a credible and useful conclusion while excluding a host of alternatives that are either useless or not credible. Making matters even harder are the subtle distinctions among kinds of reasoning that must be respected if research is to be meaningful. Is the reasoning logical or probabilistic (or something else)? If probabilistic, does it have to do with objective or subjective probabilities (or another kind)? Each branch of this tree of distinctions can be elaborated to several levels of depth.

In the present chapter, we aim to give you just a peek at the kind of research currently carried out on the complex topic of reasoning. We’ll describe a single study that attempts to illuminate the difference between two of the principal categories of reasoning, which we can call deductive versus probabilistic. A partial guide to the literature that we don’t describe is provided in the reference section.

Two kinds of reasoning

The goal of the study is to identify and contrast the brain areas underlying deductive versus probabilistic reasoning. Deduction underlies the intuition of necessity that accompanies inferences like: "No senator is both lazy and corrupt, therefore every senator is either not lazy or not corrupt." Probabilistic reasoning yields intuitions of likelihood (in contrast to certainty). For example, many people judge it likely that the ice caps will melt before 2050 on the assumption of increased oil consumption in the next 100 years.

The psychology literature provides competing perspectives about the functional neuroanatomy of reasoning. According to one point of view, deductive reasoning is grounded in the structure of language (Braine, 1978, Rips 1994). Such a view makes it plausible that logical thinking depends on left-hemisphere language areas. For example, one introductory psychology textbook affirms that "at least for right-handed people, the left hemisphere tends to be dominant for language [and] logic …" (Weston, 1999, p. 116.) A rival perspective interprets logical implication in terms of the nonlinguistic interpretations that render sentences true or false (Johnson-Laird, 1994). Since the interpretations are conceived as schematic in character, right-hemisphere participation in logical thinking is foreseen (as in Johnson-Laird, 1994). Theories that conceive probabilistic and logical reasoning as drawing on similar mental operations (e.g., Johnson-Laird et al., in press) raise the possibility of similar neural substrates for the two kinds of thinking. Just this hypothesis is proposed in Cosmides and Tooby (1996, p. 61).

Some of the foregoing claims have been evaluated via neuroimaging studies of syllogistic reasoning. Syllogisms are inferences that involve categorical premises and conclusions, such as "Some of the mayors are women." Using positron emission tomorgraphy (PET) to record brain blood flow in men reasoning about syllogistic stimuli, two studies observed distinct neural activations for probabilistic versus deductive problems (Goel et al., 1997; Osherson et al., 1998). But in one study deduction was primarily in right posterior and right frontal brain areas (Osherson et al., 1998) whereas in the other it was mostly in left frontal and left temporal brain areas (Goel et al., 1997). The finding of Goel et al. (1997) was replicated in a follow-up experiment in which men reasoned about syllogistic, spatial relational, and non-spatial relational deductive problems (Goel et al., 1998). Even for spatial relational syllogisms (involving inferences about relative height) Goel et al. (1998) observed activations only in the left hemisphere. This is unexpected in light of earlier reports of impaired reasoning on spatial relational problems by patients with insult to the right hemisphere (Caramazza et al., 1976; Hier and Kaplan, 1980, Read, 1981; Grossman 1982; Grossman and Haberman, 1987).

So you see that the findings of previous studies have been rather inconsistent. Part of the inconsistency may result from different methods of contrasting the brain activations associated with alternative forms of reasoning. In Osherson et al. (1998), identical stimuli were used to elicit both probabilistic and deductive reasoning. To get the subject to engage in one versus the other form of reasoning, instructions and preceding problems were used to "set" the thought processes applied to a given stimulus. In Goel et al. (1997), distinct stimuli were employed to elicit the two forms of reasoning, thus leaving open the possibility that observed activations depended as much on the stimuli as the reasoning task. The discrepant findings might also be partly explained by the special character of syllogisms, which are known to encourage a variety of reasoning strategies, including mentally represented diagrams (Ford, 1995). It may be that subjects in the different studies relied on mental diagrams to different extents, leading to different activation levels in brain areas hypothesized to underlie visual imagery. (These regions are bilateral or predominantly right posterior, frontal, and prefrontal regions areas. See Jonides et al., 1993; Kosslyn, 1994; Parsons and Fox, 1998; Courtney et al., 1998; Dehaene et al., 1999; Farah, 1999).

To reduce the role of visual-spatial imagery, the present study described below employed tasks based on propositional logic, which lends itself less readily to diagrammatic reasoning. For example, from the premise If it is cold in Sacramento then it is cold and rainy in San Francisco, propositional logic allows deduction of the conclusion If it is not rainy in San Francisco then it is not cold in Sacramento. A sophisticated reasoner might discover a diagrammatic scheme for verifying the validity of such inferences, but this seems less likely than for syllogisms. Propositional logic stimuli were therefore constructed that could support both probabilistic and deductive reasoning, as well as a control tasks involving no more than comprehending the sentences involved in the reasoning tasks. PET measures of regional cerebral blood flow were then obtained while subjects performed the three tasks with identical stimuli. The different kinds of reasoning (probabilistic, deductive, comprehension) were psychologically "set" by changing the instructions to subjects, and by embedding the scanned tasks within longer series of problems involving reasoning of the desired kind.

In sum, the study was designed to determine the extent to which deductive and probabilistic reasoning rely on the same brain regions, as well as to shed light on the role of language areas and visual-spatial processing areas in each type of reasoning.

Materials and Methods

Now we describe the stimuli used in the study, and the procedure. We’ll rely on the following terminology. By an argument will be meant a list of two premises followed by a conclusion. (A solid line separates the conclusion from the premises.) The task posed by an argument is to determine whether the conclusion "follows" from the premises. There are two distinct senses in which the conclusion might follow from the premises. It might follow "logically" in the sense of being guaranteed to be true provided that the premises are true. Or else the conclusion might follow "probabilistically" in the sense of having a greater than 50-50 chance of being true given that the premises are true.

In overview, subjects in the experiment faced three tasks. The deduction task required distinguishing valid from invalid arguments (a mix of valid and invalid arguments was given). The probability task required judging whether the conclusion of an argument was more likely to be true than false given the premises (based on pilot studies, we were able to construct arguments that elicit a range of intuitions). The comprehension control task required detecting anomalous content in premises or conclusion. By content being anomalous we mean that it is strange enough to conflict with the common meaning of the words involved (as in "hot ice cube"). We now provide details, starting with a description of the kind of arguments used across the three tasks.

Materials

On each trial, subjects viewed an argument composed of two premises followed by a conclusion. The first premise was shown for 3 seconds, then joined by the next premise for 3 more seconds, and then joined by the conclusion for 14 more seconds. All arguments in the experiment were composed of sentences bearing on a man drawn randomly from a small Texas town known to the subjects. The chosen man was denoted "he." Sample arguments are shown in Figure 1. In Figure 1a are examples of valid and invalid stimuli from the deduction task. Figure 1b contains examples of stimuli from the probabilistic reasoning task. In Figure 1c are examples of stimuli with and without anomalous content from the comprehension control task.

Six invalid arguments were evaluated on three separate occasions, once for validity (deduction task), once for high probability of the conclusion (probabilistic reasoning task), and once for anomaly (comprehension control). This is the key feature of the experiment. Since the very same arguments were used to elicit all three types of reasoning, differences detected in the brain activity associated with different kinds of reasoning cannot be attributed to the stimuli. Figure 1d shows all six of the invalid arguments used in the three different reasoning tasks.

Fig. 1a

Fig. 1b

Fig. 1c

Fig. 1d

Figure 1. Sample argument stimuli. (a) Two examples of valid arguments followed by one example of an invalid argument, all used in the deduction task. (b) Three examples of arguments used in the probabilistic reasoning task. (The first was constructed to elicit a judgment of high likelihood for the conclusion; the second was constructed to elicit a judgment of low likelihood; the third was intermediate.) (c) Two examples of arguments with anomalous content followed by one without. (d) The six arguments that were evaluated on three separate occasions, once for validity, once for high probability of the conclusion, and once for anomaly.

On each PET scan, or task, subjects evaluated a sequence of five distinct arguments. In the context of suitable instructions, the first two arguments in a task were designed to exercise a specific kind of reasoning (involving probability, deduction, or comprehension). The subject was only scanned while reasoning about the third and fourth arguments. To summarize, a task was a sequence of 5 arguments of which just the 3rd and 4th were involved in scanning; the others were used to establish "set" and to exercise the desired form of reasoning.

Tasks were organized into "matched sets." A matched set consisted of a probability task, a deduction task, and a comprehension task with the property that the scanned arguments (3rd and 4th) were identical across the three tasks. Three matched sets were constructed, each with its unique set of arguments. The nine resulting tasks were presented to the subject in random order. Let us now describe in more detail the three kinds of task making up each matched set.

In the probability task (Fig. 1b), subjects judged whether a conclusion was more likely to be true than false, assuming the truth of the premises. All of the arguments in the probability task were (logically) invalid. The information in the premises was therefore insufficient to force the conclusion to have either high or low probability. The judgment was thus subjective in character, and required the reasoner to integrate background knowledge (e.g., about jobs and recreation) not explicitly presented in the argument. Notice that the nature of the probabilistic task made it necessary to use invalid tasks in this task. Valid arguments leave no room for probability since the conclusion is guaranteed to be true if the premises are.

In the deduction task (Fig. 1a), subjects judged whether the conclusion followed logically from the premises. This judgment involves the detection of logical necessity rather than probability. No background knowledge is required for the reasoning since the information provided by the premises suffices to determine whether the inference to the conclusion is logically valid. Approximately half the arguments in a given deduction task were valid, the rest invalid.

In the language comprehension task (which served as a control), subjects judged whether there was anomalous content in any premise or conclusion, with no need to judge the relationship between statements (Fig. 1c). None of the arguments in the comprehension task were valid.

The two arguments common to all three tasks in a matched set were always invalid with no anomalous content (Fig. 1d). Only during performance with these common stimuli were subjects scanned. Identical stimuli were therefore processed with different psychological "set," eliciting either deductive, probabilistic or semantic reasoning. Subjects were not informed about the presence of identical arguments across different tasks, and never remarked upon this fact spontaneously. Subjects were not informed about the onset or offset of PET scanning, for which there was no perceptible change in the immediate environment.

As noted, subjects were scanned during the deduction task only when they were reasoning about invalid arguments. Is this a limitation to the study inasmuch as reasoning on valid arguments was never scanned? We don’t think that much is lost by limiting reasoning to invalid arguments. For, the reasoning mechanisms engaged during scanning in the deduction task were likely to be the same or similar to those involved in evaluating valid arguments. This is because (a) subjects did not know when they would be scanned, (b) scanned arguments were surrounded by both valid and invalid arguments, and (c) subjects did not know prior to reasoning whether a given argument was valid or invalid.

It is also worth emphasizing that a yes/no response was required in the probability task, instead of a numerical judgment. A Yes response meant that the conclusion was more likely than unlikely given the premise; a No response meant that the conclusion was more unlikely than likely. Identical response options were thus available for the probability and deduction tasks; all that differed was the kind of reasoning needed to choose between them. (We elected not to use numerical responses in the probability task because they would have introduced a confounding asymmetry vis-à-vis deduction.) Observe as well that there were no objectively correct answers to the probability questions. We decided not to frame probability questions with objectively correct answers (e.g., involving urns) because that would have converted them into disguised deduction problems (e.g., requiring the calculation of posterior odds from prior odds and likelihood). In contrast, the intuitions about chance at issue in this chapter are based on subjective assessments that cannot be determined by logic alone. Subjective assessments are often the focus of psychological research (e.g., Kahneman et al., 1982; Yates, 1990; Johnson-Laird et al., in press).

Our stimuli underwent extensive pilot testing to ensure that subjects (a) could successfully perform the deduction tasks, (b) felt intuitively that the deduction and probability tasks required distinct kinds of reasoning, (c) could sustain reasoning to complete a judgment during the allotted response time interval, and (d) agreed with the experimenters and each other about which arguments contained semantic anomaly. As a final check, 23 Rice University undergraduates were asked to carry out all nine tasks in front of a computer terminal, and then to answer follow-up questions about their thought processes. The average accuracy rate for the 15 deduction problems (which included 8 valid arguments and 7 invalid arguments) was 86.4% (S.D. = 9.1). This is reliably better than the 50% expected from mere guessing (P < 0.0001). There was also statistically reliable agreement among the subjects’ responses on the probability trials (Kendall Coefficient of Concordance W = .11, chi-square = 36.2, P < 0.001). Virtually all of the students spotted all the anomalies introduced into the arguments in the comprehension control task.

The follow-up questionnaire used in the preliminary study included the following query: "Did the logic and probability tasks differ only in how much time it took you to decide? Or do you think that you used different reasoning in the two tasks?" All 23 subjects said that "the two tasks required different reasoning." On the other hand, there was some role for deduction in probabilistic reasoning. Thus, 12 students reported that "a considerable amount of my time on the probability questions was devoted to deduction," and 11 reported that only "a small amount of my time on the probability questions was devoted to deduction." Thirteen of the students reported that deductive reasoning was either "rather different" or "very different" from that for probability; 10 reported that the two tasks were "a little different;" no one denied any difference at all.

The use of deduction in the probabilistic reasoning task was expected: it is widely appreciated in the literature devoted to subjective probability that deductive relations among propositions play a role in their probabilities. (See presentations of the axioms for subjective probability in, e.g., Skyrms, 1986, p. 168, or Earman, 1992, p. 36.) Moreover, various psychological theories envision a role for deduction in probabilistic contexts. This is true, for example, of Rips' (1994, p. 278 ff.) account of the psychology of deduction. Indeed, Rips takes deduction to be the core of human cognitive architecture, hence involved in most aspects of problem solving, planning, and categorization. It is therefore likely that the probability task recruited elements of deductive reasoning.

Regarding the comprehension task, 19 out of the 23 subjects in the preliminary study reported thinking "about each statement by itself," as requested in the instructions. Four students admitted that they "tended to think about connections between different statements."

In summary, the design of our stimuli allows probabilistic and deductive reasoning to be carried out over identical stimuli by changing instructions. The resulting thought processes are intuitively distinct although partially overlapping. The comprehension task requires merely understanding the individual sentences of an argument, and elicits only a slight tendency to link premises to conclusion.

Subjects

For the PET procedure itself, we recruited ten healthy right-handed adults. There were five males and five females. They ranged from 23 to 43 years of age (means, 32 for the male group and 32 for the female group). Subjects were screened to ensure no pre-experimental training in formal logic.

Procedure

During the PET session, each subject performed three matched sets of tasks (hence, nine tasks in all). In addition, they were scanned twice with their eyes closed, at rest. The three tasks in a given matched set involved judgments of probability, deductive validity, and semantic well-formedness, as described above and illustrated in Figure 1. Before the PET session, subjects received supervised practice in each task (using stimuli not appearing in the experiment). During the PET session itself, subjects lay stretched out in the PET instrument, with the head immobilized by a closely fitted plastic facial mask. Stimuli were displayed on a monitor suspended in front of the subject’s eyes. The first premise of each problem was presented for 3 s, then joined by the second premise for 3 more s, and then joined by the conclusion for 14 s (total display/trial time, 20 s). At the conclusion of the PET session subjects responded to a questionnaire eliciting introspections about the phenomenology of the three tasks.

For each task, the subject evaluated five arguments. During the third and fourth arguments, brain blood flow was imaged. For each argument, subjects were instructed to perform the task throughout its 20 s period; they were to continue to double-check their judgment to ensure accuracy if they finished before the stimuli were removed from view. Note that subjects reported their judgments only at the end of each task (hence, after all five arguments were presented). If the responses had been made during the presentation of stimuli, the resulting brain activations would have reflected the motor planning needed to generate a vocal response. By collecting judgments only at the end of the stimuli in a task, this irrelevant source of activations was eliminated. As an aid to memory, the five arguments in a given task were presented a second time at the end of scanning, and subjects recalled their opinion about each.

Measures of Task Performance

Overall accuracy on the deduction problems, assessed against the criterion of standard logic, was 72% (69% for the scanned trials), which is comparable to logical performance without time stress (Rips, 1994), and is reliably better than chance (P < 0.001, binomial test). Regarding the probability task, across all trials the subjects judged 41% of the conclusions to be more likely than not (45% for the scanned trials). They also agreed with the experimenters' judgment 72% of the time. There is no objective standard of accuracy for the probability task since the laws of probability do not impose any particular value for the arguments we used. Thus, the best measure of the quality of the stimuli is the extent to which the subjects agreed with one another. The subjects’ judgments were indeed concordant (Kendall Coefficient of Concordance = .349, chi-square = 39.1, df = 14, P < 0.001). Responses to the semantic comprehension test were virtually perfect. These results indicate that the subjects successfully performed the three different tasks. It is thus pertinent to examine the brain activations provoked by the three tasks.

Brain activation results

The PET scanning produced images of the brain blood flow that occurred during the three tasks. Our analysis of these data was designed to discover four things. First we hoped to locate the active regions that are common to the two kinds of reasoning. Second, we aimed to uncover the regions (if any) that were distinctive to reasoning compared to linguistic processing. The third goal was to determine whether the distinction between probabilistic and deductive reasoning interacted with left versus right hemispheric processing. Finally, we sought specific contrasts (if any existed) between the brain sites active for probabilistic versus deductive reasoning.

Activations Common to the Two Kinds of Reasoning

We found that many brain areas were active when the comprehension control task was subtracted from the probabilistic and deductive reasoning tasks. Such a "subtraction" involves removing the level of activation found at a given brain area for a specific task from the level of brain activation found at the same brain area for another task. A subtraction of this nature is often called a "contrast." The two contrasts at issue in our first analysis are those that (a) subtracted the comprehension task from the deduction task, and (b) subtracted the comprehension task from the probability task. These two contrasts reveal which brain activations are specific to reasoning, since the activations involved in just comprehending the premises and conclusion (without any reasoning) are removed. Only 8% of these areas were active in both contrasts despite the fact that the two reasoning tasks involved identical stimuli. The common, overlapping activation is in precuneus or posterior cingulate cortex [Brodmann area (BA) 31 The lack of overlap in significant activity detected for the two tasks suggests that probabilistic and deductive reasoning were performed via different underlying neural mechanisms. This inference is consistent with subjects’ introspective reports suggesting that different cognitive processes were used in the two tasks.

Reasoning versus Language Processing

In the next analysis, we focus on the possible contribution of language processing areas to the activations elicited in the reasoning tasks. When the comprehension task is subtracted from the probability task, there is no activity in or near Broca’s area (in BA 44/45), sub-Broca’s area (in BA 47), or Wernicke’s area (in BA 22/21). These three areas are known to contain primary left-hemispheric language sites (Petersen et al., 1989; Mazoyer et al., 1993; Stromswold et al., 1996; Price, 1998). Likewise, none of these areas show activity when comprehension is subtracted from the deduction task. Apparently, the purely linguistic effort required by the two reasoning tasks does not exceed what is necessary to spot semantic anomaly among the premises and conclusion of an argument. The foregoing contrasts with comprehension argue against interpreting the reasoning activations as "spill over" from overloaded language areas, a phenomenon which appears to occur with the comprehension of sentences of increasing complexity (Just et al., 1996).

Hemispheric Interaction with Type of Reasoning

We next sought to characterize the difference in the location of regions activated in the two reasoning tasks and found a relationship between task and cerebral lateralization
When the probability task is subtracted from the deduction (or "logic") task, 65% of all significantly activated voxels in the brain (excluding cerebellum) (P < 0.001) were in the right hemisphere. This contrast also showed lesser activations in left cerebral hemisphere but only in visual areas and certain subcortical structures. When deduction is subtracted from probability, 59% of all significantly activated voxels in the brain (excluding cerebellum) (P < 0.001) were in the left hemisphere. The activation in the right cerebral hemisphere was again only in areas likely to be involved in the control of attention. Visual areas and some subcortical structures also appear.

Specific Sites for the Two Kinds of Reasoning

We now turn to direct contrasts between the two reasoning tasks. In the contrast subtracting probabilistic from deductive reasoning (P < 0.001), the largest activation by far was in right middle temporal cortex (BA 21) (Fig. 2, Table 1). This activation (51, -27, -9) is just below a region homologous to one of the principal language areas of the left hemisphere (Wernicke’s area). Although the exact boundaries of Wernicke’s area are unknown, tasks thought to activate the region elicit peak intensity responses near a region homologous to that activated here. For example, Wernicke’s area has been reported to be at (-48, -32, 6) in a recent report of a study of verb generation (Xiong et al., 2000), and to be at (-54, -41, 8) in a meta-analysis of four papers on word reading (Fiez and Petersen, 1998).

Another major focus was detected in right inferior frontal gyrus (BA 44) (Fig. 3). This activation (53, 16, 17) is adjacent to a region homologous to the other principal left language area (Broca’s area). Again, although the precise boundaries of Broca’s area are unknown, tasks which appear to activate it produce peak intensity responses near a region homologous to that activated here. (See Becker et al., 1994; Bookheimer et al., 1995; Braun et al., 1997; Buckner et al., 1995; Hirano et al., 1996; Hirano et al., 1997; Petersen et al., 1988; Petrides et al., 1993; Petrides et al., 1995; Price et al., 1994).

Figure 2-8. Grand mean PET-rCBF increases (P < 0.001) for deduction minus probabilistic reasoning (in green-blue z-score scale) and probabilistic reasoning minus deduction (in yellow-red z-score scale) overlaid onto subjects’ mean anatomical MR images (greyscale) in coronal planes.
Figure 2: Deduction-specific rCBF increases in right middle temporal gyrus (BA 21, arrow).
Figure 3: Deduction-specific rCBF increases in right inferior frontal cortex (BA 44) and probability-task-specific rCBF increases in left inferior frontal cortex (BA 47).
Figure 4: Deduction-specific rCBF increases in right basal ganglia (caudate nucleus, arrow).
Figure 5: Deduction-specific rCBF increases in right amygdala; there were no probability-specific activations in amygdala.

Two other activated foci of moderate cluster size and intensity in this contrast were in right caudate nucleus (Fig. 4) and right amygdala (Fig. 5).

A different picture emerges of the areas specific to probabilistic reasoning. When deductive reasoning is subtracted from probabilistic reasoning, there are a number of large and intense activations in the left hemisphere areas of inferior frontal (BA 47) [Fig. 3 and 5, Table 1] and insular cortex, as well as posterior cingulate (BA 31) [Fig. 6], parahippocampal (BA 36) [Fig. 7], medial temporal (BA 35), and superior and medial prefrontal (BA 9) cortex.

Figure 6: Probability-task-specific rCBF increases in left inferior frontal cortex (BA 47).
Figure 7: Probability-task-specific rCBF increases in left posterior cingulate cortex (BA 31).
Figure 8: Probability-task-specific rCBF increases in left parahippocampal cortex (BA 36).

Subtraction of logic from probability also revealed left hemispheric foci of moderate cluster size and intensity in subgyral lateral frontal (BA 6) and temporal (BA 35) cortex, midbrain, and paracentral cortex (BA 5). The same contrast revealed right hemisphere foci, mainly in anterior cingulate (BA 24), globus pallidus, and uncus (BA 28). Most of these areas are related to attentional or visual functions. Again, there were bilateral posterior cerebellar foci.

Overall Activation and Task Difficulty

In order to further quantify the differences between deductive and probabilistic reasoning, we examined the amount of overall activity for each task. In total, there was a third more logic-specific activation foci than probability-specific ones in the direct contrasts between deduction and probability. The greater extent of deductive activations is unlikely to result from more vigorous use of the system underlying deduction compared to probability because eight of the ten subjects judged probability to be the most difficult of the three tasks in a post-experimental questionnaire. (Likewise, 17 out of 23 subjects in our pilot study judged probability to the most difficult of the three tasks.)

Discussion

In the following discussion, we offer preliminary hypotheses about the neurobiology of deductive and probabilistic reasoning. To underline the provisional nature of our conclusions, let us affirm at the outset that the results of the experiment we described do not allow us to infer with certainty either the specific function of the activated areas, or even whether particular activations are essential to reasoning or just incidental.

Dissociated Activations for the Two Types of Reasoning

A variety of brain areas were activated by the deduction and probability tasks. A few were active for both kinds of reasoning but many more were active uniquely for one or the other. Our results thus give evidence for a dissociation between the brain areas specifically activated for deductive versus probabilistic reasoning with propositional arguments. The dissociation we observed supports psychological theories that enforce a partial separation between the two reasoning processes (e.g., Braine, 1978). By the same token, the findings suggest that it is inaccurate from the point of view of functional neuroanatomy to claim that humans judge logical truth as a limiting case of probability assessment, i.e., that they use the same cognitive processes for deduction and probabilistic reasoning, as has been suggested (Johnson-Laird, 1994; Johnson-Laird et al., in press). We now offer further remarks about each form of reasoning.

Probabilistic Reasoning

Recall that in probabilistic reasoning, assessing the likelihood of the conclusion requires integrating information that goes beyond what is available in the premises (otherwise, the reasoning would bear on deductive validity instead of probability). Consistent with this requirement, probabilistic reasoning activated a set of brain areas that appear to be involved in recalling and evaluating a range of world knowledge. For example, there was strong activation in left inferior frontal areas, which have been implicated in the retrieval of semantic information as well as the use of working memory (Demb et al., 1995; Petrides, 1995; Rushworth et al., 1997; D’Esposito et al., 1998). The posterior cingulate was also activated during the probability task. The function of the posterior cingulate is still debated but it has been associated with either attention or long term episodic memory (Petrides et al., 1993; Grasby et al., 1993; Price et al., 1994, Shallice et al., 1994). In addition, probabilistic reasoning activated parahippocampal and medial temporal areas. These regions have been convincingly associated with declarative, semantic memory (Damasio et al., 1996; Brewer et al., 1998; Wagner et al., 1998). Finally, responses were observed in medial prefrontal cortex, which may be involved in executive attention (Petrides et al., 1993; Baker et al., 1996; Prabhakaran et al., 1997; Waltz et al., 1999).

Deduction

In contrast to probabilistic reasoning, deduction does not depend on general knowledge, but only on recognition and use of the logical structure spanning premises and conclusion. The PET data indicate that, in distinction to probabilistic reasoning, deductive inference primarily activated a set of right brain areas, the major ones being proximal to homologues of the left hemisphere structures responsible for language processing (i.e., Wernicke’s area and Broca’s area). No such areas in right hemisphere were active for probabilistic reasoning.

To date, few functions have been attributed to the two principal regions we observed for deduction. All have involved higher-order linguistic tasks such as maintenance of thematic coherence (St. George et al., 1999), discourse management (Beeman and Chiarello, 1998), and interpretation of context (Bottini et al., 1994; Just et al., 1996; Shammi and Stuss, 1999). Perhaps the right brain areas specifically active for deduction are distinct from those supporting such higher-order language functions. This is because thematic coherence and discourse management are likely to be equally required for evaluating identical arguments under probabilistic versus deductive instructions. Yet the foci observed in those areas were present only for deduction. We therefore suggest that the activated areas support logical reasoning. This suggestion is consistent with reports that right frontal patients have specific difficulties making inferences and interpreting propositions joined or modified by logical connectives (Grossman and Haberman, 1987; Beeman and Chiarello, 1998).

Deduction also produced a major focus of activation in right amygdala. The amygdala has been implicated in the processing of emotion, most often fear, aggression, reward, and risk (Hyman, 1998; LaBar et al., 1998, Kahn et al., 2000). In particular, it may play a role in learning to avoid neutral stimuli that are paired with aversive events (Sananes and David, 1992; Romanski et al., 1993; Cahill et al., in press). The amygdala is suspected more generally of emotionally tagging learned associations (McGaugh et al., 1996), for example, recognizing the value of stimuli that predict positive reinforcers (Cador et al., 1989; Everitt et al., 1989).

The activation of the right amygdala (but not the left) in the deduction minus probability contrast suggests its connection to deductive reasoning, which activated predominantly right hemisphere structures. We speculate that an emotional basis for such activation is provided by the "Aha!" phenomenology reported for the deduction task in post experimental interviews. Indeed, 70% of our PET subjects noted sudden insight during the deduction task whereas 80% described probabilistic reasoning as involving the gradual stabilization of judgment. Consistent with the lack of an "Aha!" during probabilistic reasoning, no activation was detected in either the left or right amygdala in the subtraction of deduction from probability. These observations reinforce the hypothesis that the two kinds of reasoning are fundamentally different. Deductive reasoning might thus involve a pleasurable release from tension as the subject suddenly perceives the logical status of an argument (as in other problem-solving settings; Davidson, 1995). Such an interpretation is consistent with studies that suggest a role of such structures as amygdala in interactions between emotion and higher cognitive processes like decision making (e.g., Damasio, 1994). Of course, the introspective "Aha" is not guaranteed to signal a significant brain event. Its correlation with deductive (but not probabilistic) reasoning and with right (but not left) amygdala activation nonetheless reinforces the hypothesis that deductive and probabilistic reasoning are fundamentally different at the neurological level.

In addition, deduction activated medial and dorsolateral prefrontal cortex. These areas (which are still being vigorously explored with a variety of methods) have been associated with executive attention and strategy functions, including controlling or monitoring the contents of working memory (Petrides et al., 1993; Posner and Dehaene, 1995; Baker et al., 1996; Fiez et al., 1996). Moreover, there were deduction-specific responses in temporoparietal and anterior cingulate areas, which have been associated with selective and sustained attention and with response selection (Corbetta et al., 1995; McCarthy, 1996).

There was no significant activation for deductive reasoning in right hemispheric areas associated with visual-spatial processing, i.e., posterior parietal (BA 40) and lateral prefrontal (BA 46) cortex (Kosslyn, 1994; Parsons and Fox, 1998; Carpenter et al., 1999; Dehaene et al., 1999; Farah, 1999). This absence of activation is consistent with the fact that 80% of the PET subjects indicated on the post experimental questionnaire that they did not generate visual diagrammatic representations of the stimulus information when performing the deductive problems. The use of stimuli from propositional logic seems therefore to have served its intended purpose of limiting recourse to diagrammatic strategies when reasoning deductively.

Deduction Across Individuals and Logical Forms

Our results held equally for men and women. Gender invariance is noteworthy in view of possible differences in cerebral organization and lateralization between the sexes (Gur et al., 1999; Kimura, 1999).

Deduction and Language Processing

We now examine the implications of our data for the relation between reasoning and language. To begin, it is worth considering an interpretation of the results that is alternative to the functional hypotheses discussed above. It assumes that comprehension and deduction are similar processes with common neural bases, and that the greater activation in the right hemisphere during deduction reflects the additional memory load of the latter. A straightforward version of this "spill over" hypothesis is that deductive reasoning loads left hemisphere language areas beyond capacity, so other support areas (e.g., right hemispheric ones) are recruited to carry out aspects of the task (see Just et al., 1996, for evidence of such spill over in a different linguistic context).

The hypothesis of spill over is contradicted, however, by the fact that no activations are seen (even at a relaxed statistical threshold) in left language areas when the comprehension control task is subtracted from deduction. The latter contrast involves both location and intensity, and indicates that no excess intensity of activity is observed in left language areas beyond what is required to spot anomalous semantic content during the control task. In other words, the spill over hypothesis predicts at least as much activation in the primary language areas as in the right hemispheric regions that are recruited during overloading, yet, only the right hemisphere regions appear when the comprehension task is subtracted from deduction. It is also of interest that the comprehension task was consistently rated as easiest by our subjects. Hence, spill over during deduction would not be expected to allow left hemisphere activations to be erased by subtraction of the activations for the comprehension task.

On the basis of our data, we are led to conclude that deductive reasoning is localized in brain areas far from the principal language centers in left hemisphere. This raises the possibility that logical competence is largely independent of natural language processing, except for statement decoding. Although there is evidence that right hemisphere areas are involved in many higher-order language functions, as discussed earlier, the primary locus of linguistic analysis is left hemispheric (Petersen et al., 1989; Mazoyer et al., 1993; Stromswold et al., 1996; Price, 1998). Our findings thus contradict the belief often expressed in the cognitive sciences and philosophy that deductive reasoning is derivative to linguistic processing (Quine, 1970; Polk and Newell, 1995). This belief is sustained by the plausible thesis that logic licenses inferential relations among statements, and that statements must be embedded in a structured language if they are to have the kind of grammatical properties that permit deduction (such properties as being conditional in form, having quantifiers with determinate scope, etc.; see Fodor, 1975).

The language that supports deduction, however, need not be a natural language like English, burdened as it is by ambiguity and ellipsis. Deduction (as well as other forms of reasoning) might rather be performed in a format that is antecedent to natural language, the latter being acquired for the purpose of expressing meanings that exist prior to their linguistic expression. If logic and language are independent in this sense, then logic might be available to prelinguistic infants and other animal species --- a prediction that is still without adequate test, in our opinion. Consistent with this hypothesis, the presence of complex reasoning in a profoundly aphasic patient with extensive lesions in left hemisphere language areas is reported in Varley and Siegal, 2000.

Conclusion

We conclude by summarizing the working hypotheses that emerge from the data described in this chapter. We postulate the existence of a logic-specific network in the right hemisphere comparable to the language-specific network in the left. Both involve temporal, frontal, and basal ganglia structures. Just as linguistic rules are encoded in the left hemisphere, deductive rules are encoded in the right. According to our hypothesis, each system allows for the successive transformation of mental representations specific to its function. The two circuits interact when the transformations for deduction are carried out via right hemisphere mechanisms over formal structures retrieved by left-hemisphere language areas. The latter structures would be coarse representations of the sentences from which they are abstracted since only their logical structure would be retained.

We hypothesize that probabilistic judgment is achieved via non-linguistic left hemisphere areas that are involved in the recall and evaluation of world knowledge. Note that in contrast to the reliance of deductive reasoning on coarse linguistic representations, probabilistic judgment must rely on the fine detail of sentences, since every word contributes to overall plausibility. Our hypotheses are thus consistent with the conjecture that right hemisphere regions are specialized for processing relatively coarse aspects of stimuli whereas left hemisphere regions are favored for fine aspects. A variety of empirical findings support this broader conjecture (see Ivry and Robertson, 1998, for a review of the evidence).