Elsevier

NeuroImage

Volume 31, Issue 3, 1 July 2006, Pages 1380-1388
NeuroImage

The neural implementation of multi-attribute decision making: A parametric fMRI study with human subjects

https://doi.org/10.1016/j.neuroimage.2006.01.017Get rights and content

Abstract

Decision making is not a unitary entity but involves rather a series of interdependent processes. Decisions entail a choice between two or more alternatives. Within the complex series of decisional processes, at least two levels can be differentiated: a first level of information integration (process level) and a second level of information interpretation (control level), leading to a subsequent motor response or cognitive process. The aim of this study was to investigate the neural network of these decisional processes. In a single trial fMRI study, we implemented a simple decision-making task, where subjects had to decide between two alternatives represented on five attributes. The similarity between the two alternatives was varied systematically in order to achieve a parametric variation of decisional effort. For easy trials, the two alternatives differed significantly in several attributes, whereas for difficult trials, the two alternatives differed only in small details. The results show a distributed neural network related to decisional effort. By means of time course analysis different subprocesses within this network could be differentiated: regions subserving the integration of the presented information (premotor areas and superior parietal lobe) and regions subserving the interpretation of this information (frontolateral and frontomedial cortex, anterior insula, and caudate) as well as a region in the inferior frontal junction updating task rules.

Introduction

A decision arises when a person is consciously aware of two or more possible alternative behaviors (be it thoughts or actions), only one of which can or should be performed. This person (the decision maker) is thus forced to choose one option out of a set of alternatives. The process of making a decision is a sequence of subprocesses, e.g., evaluating specific aspects of each alternative, constructing a mental representation of the decision situation, or judging the involved uncertainty (Paulus et al., 2005). These subprocesses are at least to some degree interdependent. Most imaging studies investigating decision making have focused on subprocesses, like uncertain decision processes (Blackwood et al., 2004, Paulus et al., 2001, Volz et al., 2003), reward (Haruno et al., 2004, Rogers et al., 2004, Bush et al., 2002, Bechara et al., 1996, Bechara et al., 1997, Bechara et al., 2000), risk taking (Paulus et al., 2003a, Paulus et al., 2003b), ethical decision making (Heekeren et al., 2003), moral judgments (Heekeren et al., 2005, Moll et al., 2005), economic decisions (Sanfey et al., 2003), monetary gains (Gehring and Willoughby, 2002), or personal choice (Turk et al., 2004). An overview article by Krawczyk (2002) on the neural basis of human decision making shows that most research is done on the specific aspect of reward/punishment, emotion, and environmental adaptiveness. Very few functional imaging studies have investigated decision making as the process of relating several independent sources of variance, i.e., in the context of relational integration.

Functional magnetic resonance imaging (fMRI) studies have used mainly gambling situations as decision tasks. As recent research indicates, gambles omit relevant aspects of real-life decision making, such as active risk management (Huber, 2002). In order to get a better understanding of the complex neural network of human decisional processes, one should meet two goals: (i) the experimental situation ought to resemble real-life decision tasks (Ford et al., 1989); (ii) decision making should be stripped down to its bare essentials (Shadlen and Newsome, 1996).

The present study is a first attempt with a basic multi-attribute decision task, in which alternatives are evaluated according to several attributes. Uncertainty and risk could be introduced into a multi-attribute task, but in its basic version, these aspects are not involved. A typical example of such a multi-attribute task is the decision situation of a person who wants to rent an apartment. The decision maker may take into account the location, the quietness, the number of rooms, how luxurious the apartment is, the distance to the person's working place, whether there are shops and restaurants nearby, etc. Different attributes usually carry different importance or weight. If our decision maker is light sleeper, quietness of the apartment would get a high weight, if not, quietness may get only a small weight.

A multi-attribute decision situation can be represented by an Alternative by Attribute matrix, with alternatives as columns and attributes as rows. Table 1 shows the general scheme of such a multi-attribute task as used in our study. In making a decision, the decision maker usually combines her or his evaluations of the aspects and the weights of the attributes to an overall decision. Several theories have been developed to describe multi-attribute decisions, for example, the additive utility model (Slovic et al., 1977), simple heuristics (Payne et al., 1993, Rieskamp and Hoffrage, 1999, Svenson, 1979), or elementary information processing operators (Huber, 1989, Payne et al., 1993). In the present study, details of these theoretical approaches are not relevant.

We make the general assumption that the decision process consists of the application of one or more decision heuristics (e.g., Lexicographic heuristic, Weighted Pros, cf. e.g., Huber, 1989, Rieskamp and Hoffrage, 1999, Svenson, 1979). A decision heuristic consists of a sequence of subprocedures that can be modeled as elementary cognitive operators. For example, the Lexicographic Heuristic consists of the following subprocedures (for a more detailed description in terms of elementary cognitive operators, see Huber, 1989): (i) weighting the attributes and selecting the most important one; (ii) evaluating and comparing the alternatives on this most important attribute; (iii) choosing the alternative that is better; and (iv) if no alternative is better, eliminating the most important attribute and restarting at item (i).

Cognitive process models of decision making, involving heuristics or elementary information processing operators, postulate at least two levels (or components) of the process: a process level and a control level. We regard the distinction of these two levels as being essential for the way decision making is reflected in the brain's activation pattern.

At this level, the subprocedures of the heuristic are performed, for example: evaluation of the alternatives features, the weighting of attributes, but also the concatenation of aspects (e.g., the amount of rent and the extras can be concatenated as ‘costs’) or the trade-off of aspects (e.g., the higher rent of apartment A is compensated for by its better location). At this level, information search is also performed (e.g., what is the size of apartment A).

Let us consider the involvement of the process level, when the decision maker gets information about both alternatives sequentially, for one attribute after the other. Thus, the decision maker would acquire first information concerning aspects A1 and B1 (see Table 1), next about aspects A2 and B2, and so on. The subprocedures on the process level have to be activated for every presented attribute, starting with the first two aspects till the whole information is displayed.

This level constitutes a meta-level for the decision process. It governs the activities on the process level. It selects a decision heuristic and coordinates the subprocedures on the process level when performing a heuristic. The selected heuristic determines how the evaluations on different attributes are integrated. In the Weighted-Pros heuristic, for example, the weights of the attributes are relevant in the integration. The integration of the weights of different attributes is performed by a subprocess within the heuristic. The control level decides whether the alternative favored at the moment is distinctly better than the others (e.g., Svenson, 1993) and thus whether the decision process can be terminated or has to be continued. For example, in a situation with several alternatives, first a heuristic is used that enables a fast reduction of the set of alternatives by inspecting only one or two attributes (e.g., lexicographic heuristic). When the set of alternatives has been reduced to a short list, these alternatives are inspected in more detail with the help, for example, of the ‘Weighted-Pros’ heuristic. There are different assumptions about the control level, which the present study will not dwell upon any further (Huber, 1989, Payne et al., 1993, Svenson, 1993).

With respect to the temporal dynamics, the processes on the control level should behave differently than the process level: in the first information acquisition step, the control process should only be activated weakly if at all because the decision maker knows that more information is coming and that a final decision would be premature. Furthermore, after inspecting the alternatives on the first dimension(s), it may not yet be clear what heuristic is best. For example, use of the ‘Weighted Pros’ Heuristic (e.g., alternative A is better on dimension 1 and 2, but B is better on dimension 3) cannot be considered until a preference conflict occurs. Furthermore, a trade-off between dimensions cannot occur if only one dimension has been inspected. Over time, with the presentation of additional information, activation of the control process should increase until a decision is finally reached.

Because of this, different temporal dynamics of the BOLD response are expected in relation to the underlying processes: areas related to the process level should show an early onset of the hemodynamic response, with the BOLD signal steadily increasing. On the other hand, areas related to the control level should show a delayed signal increase. Further, response-related areas should show signal changes only at the time the response is given.

We additionally hypothesize that the amount of neural activation of the process level and control level should be related to the difficulty of the decision task. Difficulty in the present study is modulated via the similarity between alternatives: the more similar the alternatives are, the more difficult it becomes to decide between alternatives. It has been argued that similarity is the primary determinant of the degree of difficulty associated with decision making (Payne et al., 1993, Krawczyk, 2002). Consider two decision situations, where the information in five attributes is presented to the decision maker. In situation (a), alternative A is distinctly better than alternative B in each of the attributes. In this case, the control process can is taxed minimally, as the final choice is obvious and remained unchallenged during the decision process. Consider, in contrast, situation (b): here, in attribute D1, alternative A is slightly better than B, in attribute D2, B is slightly better, in attribute D3 both alternatives are equally good, in attribute D4 alternative B is slightly better than A, and in attribute D5, A is slightly better than B. In this case, we expect a more elaborated control process because none of the alternatives is unambiguously better. The control process may have to meticulously compare the attributes and their weights or come to the conclusion that more information is needed. Thus, the more difficult a decision process is, the greater the anticipated involvement of the process and control levels will be. The measured BOLD response should differentiate between the levels of similarity.

A comparable two-stage model for visual discrimination in the rhesus monkey has been put forward by Shadlen and Newsome, 1996, Shadlen and Newsome, 2001. (For an overview of an animal decision-making model, see Reddi, 2001, Schall, 2002, Sakagami and Tsutsui, 1999, Kim and Shadlen, 1999). A recent fMRI study by Heekeren et al. (2004) showed that this two-stage model also holds for perceptual decision making in human subjects. The superior parietal lobe, related to the integration of information (Prabhakaran et al., 2000) and premotor structures (Heekeren et al., 2004) are hypothesized to be involved to the process level of decision making.

Several decision-making studies have described an involvement of the dorsolateral prefrontal cortex (DLPFC) in processing similarities and relational processing in decision making (Prabhakaran et al., 1997). Inductive, open-ended processing appears to be right lateralized, while explicit rule-based processing tends to be left lateralized (Goel and Dolan, 2000, Elliott and Dolan, 1998). As the present task is mainly based on explicit contextual cues, a left lateralization in the DLPFC activation is expected. Patient studies showed that lesions in the prefrontal cortex result in planning deficits (Goel et al., 1997). We hypothesize that regions in the prefrontal cortex are related to the control level of decision making.

The present study aims at localizing those neural structures involved in a more complex decision-making paradigm. We implemented a simple multi-attribute decision-making task (see Table 1), where subjects had to choose between two alternatives described on five attributes. The similarity between the two alternatives was varied systematically in order to achieve a parametric variation of decisional effort. For easy trials, the two alternatives differed significantly in several attributes whereas for difficult trials, the two alternatives differed only in small details. As mentioned above, for very different alternatives, the choice is obvious, reliable, and fast. If the similarity of the alternatives is high, the decision process becomes more difficult and is suggested to take longer. Along with the changing onset of the hemodynamic signal, this design allows to distinguish between areas involved in the process level and the control level of decision making.

Section snippets

Subjects

15 neurologically healthy subjects (7 female; mean age 26.6 years) were tested. Written consent from all subjects was obtained prior to the scanning session. All subjects had normal or corrected-to-normal vision and were native German speakers. None of the subjects was taking medication at the time of the study. Subjects were instructed prior to the actual experimental session. They were given time to practice the task with its various conditions and the key allocations. Once they felt

fMRI data analysis

The fMRI data were processed with LIPSIA software (Lohmann et al., 2001). This software package contains tools for preprocessing, registration, statistical evaluation, and presentation of fMRI data. Functional data were motion-corrected offline with the Siemens motion correction protocol (Siemens, Erlangen, Germany). To correct for the temporal offset between the slices acquired in one scan, a cubic-spline interpolation was applied. A temporal highpass filter with a cutoff frequency of 1/144 Hz

Behavioral results

Fig. 1 shows the averaged reaction times (RT) for the four similarity levels of the decisions. A significant increase in RT was elicited with increasing similarity of the two presented alternatives (ANOVA with repeated measures; F(3/42) = 25.06; P < 0.001). The RT of the Easy-Level is significantly different from the Medium1-Level (paired t test; P = 0.001), and Medium2-Level differed from the Difficult-Level (P < 0.001), but the Medium1-Level did not significantly differ from the Medium2-Level

Discussion

The present findings demonstrate that our experimental task required a distributed neural network in order to decide between competing alternatives (see Fig. 2 and Table 3). The network is predominantly located in the left hemisphere and incorporates inferior frontolateral, parietal, and posterior medial frontal areas as well as the anterior insula and the caudate nucleus. The dominance of the left hemisphere is in accordance with findings of Goldberg and Podell, 1999, Goldberg and Podell, 2000

Conclusion

In summary, the current investigation showed that a distributed neural network is involved in solving a rather simple decision-making task. This network includes the left and right frontolateral cortex, the superior parietal lobe, the posterior medial frontal cortex, the anterior insula, as well as the caudate nucleus. Whereas the parietal lobe and preSMA are suggested to be related to the integration of information (process level), the lateral prefrontal cortex and the posterior medial frontal

Acknowledgments

The authors would like to thank Mandy Naumann, Simone Wipper and Annett Wiedemann for their support with acquiring the fMRI-data and Shirley-Ann Rüschemeyer for her helpful comments on the manuscript.

References (80)

  • C. Genovese et al.

    Thresholding of statistical maps in functional neuroimaging using the False Discovery Rate

    NeuroImage

    (2002)
  • G.H. Glover

    Deconvolution of impulse response in event-related BOLD fMRI

    NeuroImage

    (1999)
  • E. Goldberg et al.

    Adaptive versus veridical decision making and the frontal lobes

    Conscious. Cogn.

    (1999)
  • H. Heekeren et al.

    Influence of bodily harm on neural correlates of semantic and moral decision-making

    NeuroImage

    (2005)
  • D. Krawczyk

    Contributions of the prefrontal cortex to the neural basis of human decision making

    Neurosci. Biobehav. Rev.

    (2002)
  • G. Lohmann et al.

    Lipsia—A new software system for the evaluation of functional magnetic resonance images of the human brain

    Comput. Med. Imaging Graph.

    (2001)
  • J. Neumann et al.

    Within-subject variability of BOLD response dynamics

    NeuroImage

    (2003)
  • M.P. Paulus et al.

    Prefrontal, parietal, and temporal cortex networks underlie decision-making in the presence of uncertainty

    NeuroImage

    (2001)
  • M. Paulus et al.

    Decision making by methamphetamine-dependent subjects is associated with error-rate-independent decrease in prefrontal and parietal activation

    Biol. Psychiatry

    (2003)
  • M. Paulus et al.

    Increased activation in the right insula during risk-taking decision making is related to harm avoidance and neuroticism

    NeuroImage

    (2003)
  • M.P. Paulus et al.

    Trend detection via temporal difference model predicts inferior prefrontal cortex activation during acquisition of advantageous action selection

    NeuroImage

    (2004)
  • M.P. Paulus et al.

    Superior temporal gyrus and insula provide response and outcome-dependent information during assessment and action selection in a decision-making situation

    NeuroImage

    (2005)
  • V. Prabhakaran et al.

    Neural substrates of fluid reasoning: an fMRI study of neocortical activation during performance of the Raven's progressive matrices test

    Cogn. Psychol.

    (1997)
  • B. Reddi

    Decision making: the two stages of neuronal judgement

    Curr. Biol.

    (2001)
  • R. Rogers et al.

    Distinct portions of anterior cingulate cortex and medial prefrontal cortex are activated by reward processing in separable phases of decision-making cognition

    Biol. Psychiatry

    (2004)
  • O. Svenson

    Process description of decision making

    Organ. Behav. Hum. Perform.

    (1979)
  • D. Turk et al.

    From facial cue to dinner for two: the neural substrates of personal choice

    NeuroImage

    (2004)
  • K. Volz et al.

    Predicting events of varying probability: uncertainty investigated by fMRI

    NeuroImage

    (2003)
  • B. Wicker et al.

    Both of us disgusted in My insula: the common neural basis of seeing and feeling disgust

    Neuron

    (2003)
  • K. Worsley et al.

    Analysis of fMRI time-series revisited—Again

    NeuroImage

    (1995)
  • E. Zarahn et al.

    Empirical analyses of BOLD fMRI statistics: I. Spatially unsmoothed data collected under null-hypothesis conditions

    NeuroImage

    (1997)
  • S. Zysset et al.

    Color–word matching Stroop task: separating interference and response conflict

    NeuroImage

    (2001)
  • A. Bechara et al.

    Failure to respond autonomically to anticipated future outcomes following damage to prefrontal cortex

    Cereb. Cortex

    (1996)
  • A. Bechara et al.

    Deciding advantageously before knowing the advantageous strategy

    Science

    (1997)
  • A. Bechara et al.

    Emotion, decision making and the orbitofrontal cortex

    Cereb. Cortex

    (2000)
  • Y. Benjamini et al.

    Controlling the False Discovery Rate: a practical and powerful approach to multiple testing

    J. R. Stat. Soc.

    (1995)
  • Benjamini, Y., Yekutieli, D., 2001. The Control of the False Discovery Rate in multiple testing under dependency....
  • J. Binder et al.

    Functional Neuroimaging of Language

  • S. Bookheimer

    Functional MRI of language: new approaches to understanding the cortical organization of semantic processing

    Annu. Rev. Neurosci.

    (2002)
  • G.M. Boynton et al.

    Linear systems analysis of 20 functional magnetic resonance imaging in human V1

    J. Neurosci.

    (1996)
  • Cited by (60)

    • Decision Making and Abductive Reasoning

      2018, Reasoning: The Neuroscience of how we Think
    • Elucidating the neural correlates of related false memories using a systematic measure of perceptual relatedness

      2017, NeuroImage
      Citation Excerpt :

      As predicted, the right mPFC also showed increased activation with respect to target/lure similarity. Previous studies have related activity within this region to task difficulty in decision-making (e.g., Volz et al., 2005; Zysset et al., 2006). False memory studies have also consistently seen activity in mPFC, associating it with greater reliance on evaluation and monitoring processes necessary when making memory decisions related to critical lures (Atkins and Reuter-Lorenz, 2011; Dennis et al., 2014, 2012; Garoff-Eaton et al., 2007; Iidaka et al., 2012; von Zerssen et al., 2001).

    View all citing articles on Scopus
    View full text