See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/262176101 Foundations for Group Decision Analysis Article in Decision Analysis · June 2013 DOI: 10.1287/deca.2013.0265 CITATIONS READS 77 1,820 1 author: Ralph L. Keeney Duke University 235 PUBLICATIONS 32,750 CITATIONS SEE PROFILE All content following this page was uploaded by Ralph L. Keeney on 13 September 2015. The user has requested enhancement of the downloaded file. Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. Decision Analysis Articles in Advance, pp. 1–18 ISSN 1545-8490 (print) ISSN 1545-8504 (online) http://dx.doi.org/10.1287/deca.2013.0265 © 2013 INFORMS Foundations for Group Decision Analysis Ralph L. Keeney Fuqua School of Business, Duke University, Durham, North Carolina 27708, keeney@duke.edu T his paper derives a general prescriptive model for group decision analysis based on a set of logical and operational assumptions analogous to those for individual decision analysis. The approach accounts for each group member’s potentially different frames of their common decision, including different events and different consequences of concern. Assuming that each group member accepts the decision analysis assumptions to evaluate his or her analysis of what the group should do and that the group accepts an analogous set of decision analysis assumptions for the group’s decision, it is proven that the group expected utility for an alternative should be a weighted sum of the individual member’s expected utilities for the alternatives. After each group member does his or her decision analysis of the group’s alternatives, the essence of the group decision analysis is to specify the weights based on the interpersonal comparison of utilities and on the relative importance or power of each individual in the group. Key words: group decision analysis; expected utility; group decisions; decision analysis History: Received on December 20, 2012. Accepted by former Editor-in-Chief L. Robin Keller on January 6, 2013. Published online in Articles in Advance. 1. Introduction rule, such as the candidate with the most votes wins. In this case, there is a collection of individual decisions by the voters that leads, with no specific group action, to a selected alternative. The third class of decisions is social planning or social welfare decisions where an individual planner or organization, after taking judgments and preferences of individuals affected by the decision into account, makes the decision. The model developed here may have relevance to each of these classes of decisions, but such uses are not the topic of this paper. To provide prescriptive guidance for group decisions, the inherent reasons for their complexity must explicitly be addressed in the group decision analysis. These reasons include the following: 1. The members of the decision-making group may view their common decision differently. They may be concerned about different consequences of the alternatives and feel that different uncertain events matter to the choice. 2. Even when group members agree on the relevant consequences and events, they may assign different probabilities to those events and different values to those consequences. 3. Each group member’s perspectives of the decision and his or her evaluations of the alternatives Decision analysis is a methodology and set of procedures for building models to guide decision making. These models are prescriptive in that they provide insights about what alternative to choose to best achieve stated objectives. The foundations for decision analysis, which are a set of logical and operational assumptions (Pratt et al. 1964), are well developed for individual decisions. However, several attempts to extend the appealing logic of decision analysis to group decisions have not succeeded except in special cases. This paper develops a general group decision analysis model. Group decisions, as defined in this paper, are decisions where a group of two or more individuals must collectively select an alternative from a set of two or more alternatives that best satisfies the group’s objectives, and no individual has veto power. This definition rules out three classes of decisions where individuals make decisions that affect groups. One class is negotiations, because the individual negotiators are trying to best satisfy their own objectives rather than the group’s objectives, and each individual has veto power. A second class of decisions ruled out is voting situations where all votes are tabulated to select an alternative according to a prespecified 1 Keeney: Foundations for Group Decision Analysis Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. 2 Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS must both be incorporated in any group decision analysis. This implies that the interpersonal comparison of preferences for the alternatives among members of the group and the relative importance or power of each individual in the group be recognized. Our group decision analysis model is structured from the beliefs and preferences of each of the individual members of the decision-making group. A set of group decision analysis assumptions, analogous to those for individual decisions, are developed for this particular group model. Given the group model and these assumptions, a group can analyze any decision consistent with the principles of decision analysis. 2. Background Literature Decision analysis was initially developed for individual decisions. Its logical foundations involve constructing a frame to model that decision for subsequent analysis based on expected utility theory, which integrates the related concepts of probability and utility. The development of each of these concepts has occurred over time, and each concept has multiple contributors. Ramsey (1926) provided the initial theory for decision making based upon these two concepts. Von Neumann and Morgenstern (1947) developed the formal theory of expected utility, which was based on so-called objective probabilities. Savage (1954) extended expected utility theory to incorporate probabilities based on a willingness to act, which are often referred to as judgmental or personal probabilities. A concise history of these developments is found in the work of Raiffa (1968). The foundational basis for an operational theory of decision analysis for an individual decision maker was fully developed by Pratt et al. (1964). Because my motivation for this paper is to develop a general decision analysis model to logically analyze group decisions, my analysis uses the individual decision analysis framework of Pratt et al. (1964) as a basis. Pratt et al. (1964) frames an individual’s decision for analysis in the following summary: Frame for an Individual’s Decision Problem. An individual decision maker must choose among a set of alternatives An , n = 11 0 0 0 1 N , and wants to choose the best of this set. The decision maker has specified a set of mutually exclusive and collectively exhaustive events Ej , j = 11 0 0 0 1 J , one of which will occur, and a set of Table 1 Decision Analysis Assumptions for Individual Decisions (from Pratt et al. 1964) Principles of consistent behavior IT: Transitivity. As regards any set of lotteries among which the decision maker has evaluated his or her feelings of preference or indifference, these relations should be transitive. IS: Substitutability. If some of the prizes in a lottery are replaced by other prizes such that the decision maker is indifferent between each new prize and the corresponding original prize, then the decision maker should be indifferent between the original and the modified lotteries. Principles for scaling preferences for consequences and judgments concerning events IP: Preferences. The decision maker can scale his or her preference for any consequence c by specifying a number 4c5 such that he or she would be indifferent between (1) c for certain and (2) a lottery giving a probability 4c5 chance at c ∗ and a complementary chance at c o . IJ: Judgments. The decision maker can scale his judgment concerning any possible event Ej by specifying a number p4Ej 5 such that he or she would be indifferent between (1) a lottery with consequence c ∗ if Ej occurs, c o if it does not, and (2) and a lottery giving a probability p4Ej 5 chance at c ∗ and a complementary chance at c o . consequences cnj , n = 11 0 0 0 1 N and j = 11 0 0 0 1 J , that will result if alternative An is chosen and event Ej then occurs. Each consequence includes all things that matter to the decision maker from the choice of an alternative. The assumptions that Pratt et al. (1964) developed to analyze an individual’s decision are given in Table 1. The principles of consistent behavior provide the logical foundation for decision analysis. For interpreting the substitutability principle, it is important to understand that Pratt et al. (1964, p. 356) defined the word prize to mean “either a consequence or the right to participate in another lottery whose payout will be a consequence.” Given these two principles and assumption IP, Pratt et al. (1964) prove that the decision maker’s preferences over consequences c can be represented by a utility function u that is a positive linear transformation of and scaled by u4c 5 = 0 and u4c ∗ 5 = 11 (1) where c and c ∗ provide lower and upper bounds on the decision maker’s preferences for all consequences. From assumption IJ, the p4Ej 5 terms are the probabilities that represent the decision maker’s judgments about the likelihoods of the various events. Individual Decision Analysis Theorem. Given the assumptions for an individual’s decision analysis in Keeney: Foundations for Group Decision Analysis 3 Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS Table 1, it follows that the decision maker should evaluate alternative An using its expected utility U 4An 5 = èj p4Ej 5u4cnj 51 n = 11 0 0 0 1 N 1 (2) where alternatives with higher expected utilities are preferred and èj means to sum over all j. The past attempts to extend the decision analysis logic to group decisions implicitly assume that each of the members of the decision-making group accepts a commonly held frame of their decision problem. Notably, and subtly, it assumes that each member of the group (i) is concerned with the same consequences of the alternatives and (ii) considers the same events to be those relevant to the decision. Thus, the standard approach has assumed that the group accepts the assumptions in Table 1 and searched for a group utility function and a group set of probabilities that can be used in a result analogous to (2) to calculate a group expected utility for each of the alternatives (e.g., Raiffa 1968). The next three paragraphs, following the literature review in Keeney and Nau (2011), summarize the work most relevant to this paper. Hylland and Zeckhauser (1979) investigate the potential usefulness of separately combining the individual’s probability assessments into a group probability assessment and the individual’s utility functions into a group utility function. They also assume that the resulting model should be consistent with weak Pareto optimality, meaning that if all members of a decision-making group prefer one alternative over another, the resulting group decision model must yield a higher expected group utility for the Pareto dominating alternative. In addition, they also assume that aggregations should not be dictatorial, so group probabilities and the group utility function should not be identical to those of any individual in the group. They prove that no group decision model is consistent with these requirements. Seidenfeld et al. (1989, p. 225) succinctly summarized the problem: “An outstanding challenge for Bayesian decision theory is to extend its norms of rationality from individuals to groups.” Their approach is to broaden the group compromises that they consider to include any set of group probabilities and group utilities, which need not be separately constructed from the individual’s probabilities and individual’s utilities, respectively. Their interest is whether there is any set of such group probabilities and group utilities that would yield a group evaluation of alternatives that is consistent with individual’s common preferences over those alternatives. Seidenfeld et al. (1989) also invoke the weak Pareto condition. They then prove that there is, in general, no compromise group utility function and group probabilities for events that can replicate the common preferences of the individual members of the group. Mongin (1995) is interested in the implications for group decisions of a group of individuals who each, and collectively as a group, accept the decision analysis (i.e., Bayesian) assumptions. In addition to combining individual group members’ probabilities over events into group probabilities and group members’ utilities over consequences into group utilities over consequences, he combines group members’ expected utilities for alternatives into group expected utilities for those alternatives. With an additional assumption of strong Pareto optimality, Mongin (1995) proves that there are no possible combinations consistent with the assumptions. Positive results have been developed for special cases. Notably, Harsanyi (1955) proved that when the expected utility assumptions hold for both individuals and the group, Pareto optimality holds, and all members of the decision-making group have common probability distributions to describe possible consequences of the alternatives, then the group’s expected utility for an alternative must be a weighted average of the individual’s expected utilities for that alternative. Harsanyi’s (1955) result is extended to situations where both probabilities of events and the utilities for consequences can differ among members of the decision-making group used by Mongin (1998), Chambers and Hayashi (2006), and Keeney and Nau (2011). Mongin (1998) and Chambers and Hayashi (2006) assumed that the group can specify expected utilities for (a) the same alternatives as the individual’s consider and (b) a Pareto optimality condition. Keeney and Nau (2011) assumed that the group has expected utility preferences over hypothetical acts described by lotteries resulting in vector consequences, where the components are the utilities Keeney: Foundations for Group Decision Analysis Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. 4 perceived by each member of the group, and that the group’s preferences are consistent with an assumption analogous to Fishburn’s (1965) concept of mutual utility independence over these consequences. Because the result in Keeney and Nau (2011) is the same as that for the group decision analysis theorem following in §5, it is useful to summarize the important distinctions between the theories. These distinctions concern the problem formulation, the decision frame, and the assumptions used. The origin for the distinction between the result in Keeney and Nau (2011) and the result here is the problem formulation. The problem formulation addressed by Keeney and Nau (2011) is the standard approach where each group member is concerned with the same set of events and perceives the same consequences for each alternative–event pair. The members may differ in their assignment of probabilities for events and their utilities for the various consequences. The problem formulation of this paper is broader. In addition to assigning different probabilities to specific events and different utilities to specific consequences, it allows different group members to be concerned with different events and to perceive different consequences following from any alternative–event pair. Hence, the problem formulation in Keeney and Nau (2011) is a special case of that addressed here. Corresponding to the earlier distinction, the group decision frame is broader in the current paper. It incorporates each of the member’s decision frames of their group decision by essentially using a union of those decision frames. In Keeney and Nau (2011), the decision frame for the group decision is exactly the same as that of each member for their joint decision. Concerning the assumptions used to calculate an expected utility for the group, Keeney and Nau (2011) first calculated the expected utilities perceived by each of the individual members for the group’s alternatives. To combine these, Keeney and Nau (2011) used a decision frame of a hypothetical group decision problem, where alternatives are defined by objectively provided probabilities over consequences described in terms of a vector of the expected utilities that each member perceives for this hypothetical group decision. Then an assumption analogous to Fishburn’s (1965), and a weak Pareto type of assumption provide the result. Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS The group alternatives in this paper are the real alternatives available in the group decision, and the probabilities are descriptions of the group’s judgments over the perceived real consequences that will occur to the group. To calculate an expected utility for the group, assumptions analogous to the decision analysis assumptions that each member uses to individually analyze the group problem are used. To relate the group preferences to the member’s preferences, an indifference assumption is used. No Pareto assumption is needed, although the result is clearly consistent with Pareto optimality. The current paper has two results, stated in the corollary in §5, that are not relevant to the group decision frame in the Keeney and Nau (2011) paper. With the broader formulation, it is shown that a group probability distribution over events acceptable to all group members is a product of each of the member’s probability distributions over events in their individual frame of the group’s decision and that a group utility function over consequences must be a weighted average of each of the member’s utility functions over the consequences that each member feels are relevant to the group. The main intent of the two papers is also different. The Keeney and Nau (2011) article is a potential contribution to the Bayesian expected utility literature mainly produced by and of interest to economists. The current article is a contribution to the decision analysis literature. The intent is to provide a logical and practical foundation for spreading the application and use of decision analysis to group decisions. To implement a theory and to have it be useful, it is important that the assumptions are stated in terms that pertain to the real decision being faced, the information needed to implement the result addresses the complexities of the decision, and practical procedures exist to assess the information. I believe the foundations for group decision analysis presented here make such a contribution. 3. Frame of the Group Decision Problem Results described in §2 are important and insightful, but they have not included a comprehensive breakthrough that provides the foundations to extend Keeney: Foundations for Group Decision Analysis 5 Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS and implement the decision analysis logic to group decisions. Thus, if decision analysis is going to be extended to group decisions, it cannot follow the approach of framing the group decision as the individual decision is framed and then trying to construct both a group utility function over consequences and group probabilities for events. A different approach must be followed. This section describes a different and broader approach to extend decision analysis to group decisions. It incorporates the different frames that group members may have for their common decision, utilizes the logic embedded in the decision analysis principles for individual decisions to analyze the group’s decision, and produces a result that can be applied to group decisions using decision analysis techniques and procedures. To illustrate different frames, consider a relatively simple decision concerning two people planning to have dinner together at a restaurant. One individual wants to have a substantial discussion at a convenient location (e.g., not far away and with easy parking), and the other individual is interested in great food at an exciting location. Because they have different objectives, they will be concerned about different consequences of their joint decision. This will also result in different events being relevant to each individual, as the first individual would be concerned with how quiet the restaurants are and about travel times to the restaurants, whereas the second individual may be concerned with whether the restaurants have exotic fusion dishes and a trendy clientele. The decision process for a more complex decision may typically involve several group meetings over time with individual effort between meetings, as each individual essentially must evaluate the desirabilities of the alternatives for the group. Subsequently, the group must collectively decide which alternative to choose. Hence, it is useful to conceptualize this process as two distinct sequential stages. In stage 1, each group member frames the group’s common decision and evaluates alternatives using the individual decision model illustrated in §2. In stage 2, the group uses the member’s evaluations to collectively evaluate alternatives. This could be done using a discussion, a voting mechanism, or the more systematic group decision analysis discussed here. Figure 1 Basic Individual Decision Problem (a) Frame of individual decision E1 (b) Model of individual decision with all elements p (E1) c11 A1 u(c11) A1 E2 E1 p (E2) c12 p (E1) c21 A2 u (c12) u (c21) A2 E2 c22 p (E2) u (c22) To clearly describe the two stages of the group decision problem, it is first useful to illustrate the steps in an individual decision analysis model. Figure 1(a) illustrates the frame of an individual decision, for the simplest case of two alternatives and two events, in terms of a decision tree structured directly from the statement of the individual’s decision problem. It has three elements: alternatives, events, and consequences. Figure 1(b) presents a model of that decision that incorporates the decision analysis principles for scaling preferences and judgments. These introduce two new elements into the model: probabilities of events and utilities of consequences. At the beginning of stage 1 for any group decision, the individual members of the group may not agree on any of the elements of the group decision. The members may view their joint decision as having different alternatives, events, consequences, probabilities, and utilities in terms of Figure 1(b). During the meetings, discussions, and information gathering of stage 1, each member develops his or her own model of the group’s decision that may account for the ideas of other members. Fortunately it should be relatively easy for a group to agree on the set of alternatives, and, as will be shown, that is the only agreement necessary to conduct a group decision analysis. If individuals initially identify different sets of alternatives for a joint decision, the union of the individual alternatives in each of the member’s sets is the appropriate set of alternatives to use for the group decision. If some members feel that specific alternatives suggested by others are inadequate, they have Keeney: Foundations for Group Decision Analysis Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. 6 the possibility and responsibility to evaluate those alternatives as poorly as they feel appropriate. However, even after substantial discussions, the group members may not agree on either events or consequences that are relevant to the group decision. Accounting for this reality is distinct from essentially all of the previous work on group decision analysis, which either explicitly or implicitly assumes that all group members recognize the same events and perceive the same consequences. Furthermore, even if group members happen to agree on the events and consequences, they may not and need not agree on their judgments of the probabilities of those events or their utilities for the consequences. By the end of stage 1, each group member has completed his or her analysis of the group’s alternatives, and stage 2 begins. Because each individual’s analysis has taken into account all of the information relevant to the group decision (i.e., alternatives, events, consequences, probabilities, and utilities) from that individual’s perspective, at this stage no group member should alter his or her analysis in any way because of any other group member’s information or analysis. To illustrate the structure of a group decision faced in stage 2, consider the simplest case of a group decision that involves two individual group members I1 and I2 , two common alternatives A1 and A2 , two events for each group member, and a consequence for each alternative–event combination. To keep notation simple, suppose that the events of relevance to I1 and I2 are different and labeled E and F , respectively, and the corresponding consequences are different and labeled c and d. The two members’ separate frames of their joint decision are illustrated in Figure 2. The frame of the group decision should be constructed from the two group members’ frames of their group decision. The logic for this is based on the following principles, which are part of the reason for and consistent with the two-stage decision process described previously. Origin of Judgments Principle. All original judgments must be developed in the minds of individuals. This principle refers to all judgments such as recognizing or creating alternatives, describing consequences, identifying events, specifying probabilities for events, and constructing utilities for consequences. The logic for this principle is that judgments can only Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS Figure 2 Individual Frames for Two-Member Group Decision (a) Frame of individual l1 E1 (b) Frame of individual l2 F1 c11 A1 d11 A1 E2 E1 F2 c12 F1 c21 A2 d12 d21 A2 E2 c22 F2 d22 be developed in a mind, and only individuals have minds. Groups do not have a mind, so all original judgments must occur in the mind of an individual. However, a group of individuals can decide to express group judgments, which leads us to a related principle. Group Judgments Principle. Any group judgments must be constructed by the group based only on the judgments of the individuals in the group. The logic for this principle is as follows: If no group member recognizes some aspect that might be relevant to a group decision, then clearly the group cannot recognize and consider it. If a potentially relevant aspect is recognized by one group member, but neither that group member nor any other member cares about it, the group should not care about it. Succinctly, this principle implies that the group should not care about anything that no member in the group cares about, and the group should care about anything that at least one group member does care about. From these principles, the two group members’ frames of their joint decision in Figure 2 are combined in Figure 3, which illustrates two equivalent frames of the two-member group decision faced in stage 2. In the group decision, there are two alternatives A1 and A2 , four possible joint events represented by (E1 F 5, and eight possible group consequences, represented as vectors of the form (c1 d5 expressing the various possible combinations of consequences in the original decision as perceived by the respective members. From Figure 3(a), it is easy to note that if the F events and d consequences, which are relevant only to member I2 , are eliminated, the group frame reduces to Keeney: Foundations for Group Decision Analysis 7 Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS Figure 3 Equivalent Frames of Two-Member Group Decision (a) Frame with conditional events (b) Model with joint events Group probabilities F1 E1 F2 A1 F1 E2 F2 F1 E1 F2 A2 F1 E2 F2 pG(E1, F1) (c11, d11) pG(E1, F2) (c11, d12) A1 pG(E2, F1) (c12, d11) pG(E2, F2) (c12, d12) pG(E1, F1) (c21, d21) pG(E1, F2) (c21, d22) A2 pG(E2, F1) (c22, d21) pG(E2, F2) (c22, d22) that of member I1 shown in Figure 2(a). This indicates that the group decision frame includes a complete characterization of everything I1 cares about, namely, only E events and c consequences. It follows that if I1 placed any direct value on d consequences, this would involve double counting, because everything I1 cares about is already accounted for with the c consequences. In addition, I1 would not care about the probabilities for F events, as these can have no implications for the judgments of I1 about the chances of different c consequences occurring. The analogous situation applies for I2 . To analyze the two-member group decision in Figure 3(b), the group needs joint probabilities for all possible combinations of (E1 F 5 events, represented by pG 4Er 1 Fs 5, and group utilities for each of the possible (c1 d5, represented by uG 4c1 d5. The decision frame for the M-member group decision is constructed from each member’s individual frames for their group decision analogous to the simple two-member example in Figure 3. An individual’s decision frame for their group decision is shown in Figure 4 and is defined as follows. Frame for a group member’s decision. Group member Im , m = 11 0 0 0 1 M, must evaluate the set of alternatives Group utilities uG(c11, d11) uG(c11, d12) uG(c12, d11) uG(c12, d12) uG(c21, d21) uG(c21, d22) uG(c22, d21) uG(c22, d22) An , n = 11 0 0 0 1 N , one of which the group will choose. The group member has specified a set of mutually exclusive and collectively exhaustive events Emj , j = 11 0 0 0 1 Jm , one of which will occur, and a set of consequences cnmj , n = 11 0 0 0 1 N and j = 11 0 0 0 1 Jm , that will result if alternative An is chosen and event Emj then occurs.1 Each of the group member’s decision frames represent their personal perspective of the group decision at the time a group decision will be made. Their individual decision frames allow them to evaluate all of the group’s alternatives for input to that group decision. Figure 5 illustrates the group decision frame. The events of interest to the group are members of a set E of mutually exclusive and collectively exhaustive events Eg , g = 11 0 0 0 1 G of the form 4E1j 1 0 0 0 1 Emj 1 0 0 0 1 EMj 5, where Emj , j = 11 0 0 0 1 Jm are the events specified by member Im . The consequences of 1 Because the events relevant to member Im are different in general from those relevant to any other group member, it may be more appropriate to put a subscript m on all subsequent j notations. Because this would result in subscripts to subscripts, and because an adjacent subscript m always appears with the subscript j, the second subscript will not be used. Keeney: Foundations for Group Decision Analysis Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. 8 Figure 4 Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS This group frame includes every group member’s frame for the group decision. To see this, recall that group member Im does not care about any events other than Emj or about any aspects of the consequences other than those of the form cnmj . If the events and consequences without the subscript m, which are those not of concern to Im , are eliminated from Figure 5, the problem frame reduces to that for Im shown in Figure 4. Decision Frame of Individual Group Member Im for the Group Decision Em1 A1 c1m1 Emj An cnmj 4. EmJ AN cNmJ interest to the group are described by a vector c = 4c1 1 0 0 0 1 cm 1 0 0 0 1 cM 5, where cm is a consequence that member Im , m = 11 0 0 0 1 M perceives for the group. Frame of the group decision. A decision-making group of M members, M ≥ 2, must choose from a set of alternatives An , n = 11 0 0 0 1 N , one of which will be chosen. One of the mutually exclusive and collectively exhaustive events Eg , g = 11 0 0 0 1 G will occur, and a consequence cng , n = 11 0 0 0 1 N , g = 11 0 0 0 1 G will result if alternative An is chosen and event Eg then occurs. Figure 5 The Group Decision Analysis Assumptions To analyze the group decision in Figure 5, a group utility function over the consequences (c1 1 0 0 0 1 cM 5 and a joint probability distribution over the relevant events are needed. The group decision analysis assumptions in Table 2, which are analogous to those for individual decisions in Table 1, provide the logical basis to derive a solution to the group decision. Note that the assumptions of scaling are existence assumptions rather than constructive assumptions as in Table 1 for the individual’s decision analysis. The group utility function is constructed from the individual group member’s utility functions, and the group probability distribution is constructed from the individual group member’s probability distributions. Both of these are done as part of the proof of the group decision analysis theorem in §5. The reasonableness of the assumptions prior and the resulting Group Decision Frame for the Group Decision A1 E11 An E1j Em1 Emj EM1 (c111,…,c1m1,…,c1M1) EMj (cn1j,…,cnmj,…,cnMj) AN E1J EmJ Group probabilities EMJ (cN1J,…,cNmJ,…,cNMJ) Group utilities Keeney: Foundations for Group Decision Analysis 9 Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS Table 2 Decision Analysis Assumptions for Group Decisions Principles of consistent behavior GT: Group transitivity. As regards any set of lotteries among which the decision-making group has evaluated its feelings of preference or indifference, these relations should be transitive. GS: Group substitutability. If some of the prizes in a lottery are replaced by other prizes such that the decision-making group is indifferent between each new prize and the corresponding original prize, then the decision making group should be indifferent between the original and the modified lotteries. Principles for scaling preferences for consequences and judgments concerning events GP: Group preferences. The decision-making group can represent its preferences over the consequences (c1 1 0 0 0 1 cM 5 in terms of a group utility function. GJ: Group judgments. The decision-making group agrees that there exists a representation of its judgments about the possible occurrence of any combination of the events (E1 1 0 0 0 1 EM 5 in terms of a joint probability distribution function. utility function and group probability distributions are discussed in §6 after presenting the main result. To relate the individual’s preferences to the group’s preferences, we need one additional identity assumption defined as follows. Group Identical Indifference. If some of the prizes in a lottery are replaced by other prizes such that the probabilities of all consequences relevant to each group member in the original and modified lotteries are identical, then the decision-making group should be indifferent between the original and modified lotteries. Note that the group identical indifference (GII) assumption is less restrictive than a Pareto assumption that would require that the group must be indifferent between lotteries when each of the individual group members was indifferent. 5. Group Decision Analysis Theorem The main result of this paper is the following theorem. Its prescriptive usefulness and procedures to use it are discussed in the following sections. Group Decision Analysis Theorem. Given a decision-making group of M ≥ 2 members facing a decision with alternatives An , n = 11 0 0 0 1 N , and assuming that (a) each member accepts the decision analysis assumptions for individual decision-making in Table 1 for his or her analysis of the group decision in Figure 4, (b) the decision-making group accepts the group decision analysis assumptions in Table 2 for their group decision in Figure 5, and (c) the decision-making group accepts the group identical indifference assumption, then the group expected utility of any alternative An , denoted UG 4An 5, is UG 4An 5 = èm wm Um 4An 5 = èm wm 4èj pm 4Emj 5um 4cnmj 551 (3) where pm 4Emj 5 is member Im ’s1 m = 11 0 0 0 1 M1 probability for event Emj 1 um 4cnmj 5 is member Im ’s1 m = 11 0 0 0 1 M1 utility for consequence cnmj 1 Um 4An 5 is member Im ’s1 m = 11 0 0 0 1 M1 expected utility for alternative An 1 and the wm , m = 11 0 0 0 1 M are scaling factors that sum to one, where 0 ≤ wm ≤ 1. The proof involves several steps. From assumption (a), each Im can do his or her own analysis of the group’s decision, which provides the individual’s probabilities and utilities for the group’s decision. Given assumption (b), the group can analyze its decision, depending on the group’s probability distribution over all possible combinations of events and the group’s utility function over consequences perceived by each of the group’s members. Then, using assumption (c), it is shown that the group preferences for alternatives depend only on the marginal probability distributions over the set of events, from which it follows that the group utility function uG must be of the additive form uG = èm wm um , where wm are nonnegative weighting factors. Hence, the expected utility of each real alternative can then be calculated as a weighted sum of the expected utilities of that alternative from each individual’s perspective. Proof. Given assumption (a), from (1), each group member Im has a utility function that can be scaled by um 4cm 5 = 0 and um 4cm∗ 5 = 11 m = 11 0 0 0 1 M1 (4) Keeney: Foundations for Group Decision Analysis Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. 10 Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS where any two of the cm consequences need not be, but may be, the same, and the same circumstance holds for the cm∗ . Also, each group member has a set of probabilities pm 4Emj 5 for all possible events that he or she considers can influence the eventual cm . Hence, each group member should evaluate alternatives using his or her own expected utility Um 4An 5 = èj pm 4Emj 5um 4cnmj 51 n = 11 0 0 0 1 N 1 m = 11 0 0 0 1 M0 uG 4c1 1 0 0 0 1 cM 5 = èm uG 4cm 5 = èm wm uGm 4cm 51 (5) A similar result follows from assumption (b). The group should evaluate alternatives using UG 4An 5 = èE pG 4E1j 1 0 0 0 1 EMj 5uG 4cn1E 1 0 0 0 1 cnME 51 (6) where E = 4E1j 1 0 0 0 1 Emj 1 0 0 0 1 EMj 5 is a member of the set E of all possible combinations of events; cnmE is the consequence relevant to Im given alternative An is chosen and E occurs; pG is the group probability of any group consequence (cn1E 1 0 0 0 1 cnME 5 given alternative An is chosen and E occurs; and uG is the group utility for consequence (cn1E 1 0 0 0 1 cnME 5. Because of how pG 4E1j 1 0 0 0 1 EMj 5 in (6) is constructed (see Figure 3), its marginal probability distribution over Emj must be pm 4Emj 5, m = 11 0 0 0 1 M, which provides the marginal probabilities over the (cn1E 1 0 0 0 1 cnME 5 for alternatives An , n = 11 0 0 0 1 N . Now define qG 4E1j 1 0 0 0 1 EMj 5 = çm pm 4Emj 5, so qG is a joint probability distribution assuming probabilistic independence of all Emj , m = 11 0 0 0 1 M. Consider two alternatives Ap and Aq described by different joint probability distribution functions pG and qG . As these have the same marginal probabilities over cm , m = 11 0 0 0 1 M, from the group identical indifference assumption, the group members each consider alternatives Ap and Aq to be identical, so the group must be indifferent between alternatives Ap and Aq . This has two important implications. First, a group probability distribution over events E that all members can agree on is pG 4E1j 1 0 0 0 1 EMj 5 = qG 4E1j 1 0 0 0 1 EMj 5 = çm pm 4Emj 50 sense,” meaning that the group is indifferent between lotteries over multiple attributes (i.e., referring to the different individual’s perceived consequences as different attributes in this case) when the marginal probability distributions over each attribute are identical. From Fishburn’s (1965) Theorem 2, it follows that the group utility function uG must be additive, so (7) Second, the group’s preferences for alternatives must only depend on the marginal probability distributions over those consequences. This is Fishburn’s (1965, p. 38) condition of “independence in the utility (8) where we can scale uG by uG 4c1 1 0 0 0 1 cM 5=0 and ∗ uG 4c1∗ 1 0 0 0 1 cM 5 = 11 (9) uG 4cm 5 is defined as uG 4c1 1 0 0 0 1 cm−1 1 cm 1 cm+1 1 0 0 0 1 cM ), m = 11 0 0 0 1 M, uGm 4cm 5 is defined by uG 4cm 5 = wm uGm 4cm 51 m = 11 0 0 0 1 M1 (10) uGm is scaled by uGm 4cm 5 = 0 and uGm 4cm∗ 5 = 11 m = 110001M1 (11) and the wm are nonnegative scaling factors where ∗ 0 ≤ wm ≤ 1, m = 11 0 0 0 1 M. Evaluating (c1∗ 1 0 0 0 1 cM 5 with (8) and using (9) and (11) yields w1 + w2 + · · · + wM = 10 (12) Note from (4) that u1 is scaled from 0 to 1, so u1 4c1 5 can be used as a probability. Now, suppose the group faces a choice between (a) an alternative described by a group probability distribution yield ing (c1∗ 1 c2 1 0 0 0 1 cM ) with a probability of u1 4c1 5 or (c1 1 c2 1 0 0 0 1 cM ) with the probability of 61 − u1 4c1 5] and (b) a sure consequence of (c1 , c2 1 0 0 0 1 cM ). The expected utility to I1 of both the lottery and the sure consequence c1 is u1 4c1 5, so I1 is indifferent between these choices. Obviously, each other group member Im , m = 21 0 0 0 1 M, is also indifferent. Hence, by group substitutability, the group must be indifferent. Equating the group utilities yields u1 4c1 5uG 4c1∗ 1c2 10001cM 5+61−u1 4c1 57uG 4c1 1c2 10001cM 5 = uG 4c1 1c2 10001cM 50 (13) Substituting (9) and (10) into (13) yields u1 4c1 5w1 uG1 4c1∗ 5 = w1 uG1 4c1 50 (14) Keeney: Foundations for Group Decision Analysis 11 Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS Now, substituting (11) into (14) yields u1 4c1 5 = uG1 4c1 50 (15) An analogous logic to (15) yields um 4cm 5 = uGm 4cm 51 m = 21 0 0 0 1 M0 (16) Substituting (16) into (8) yields uG 4c1 1 0 0 0 1 cM 5 = èm wm um 4cm 51 (17) which is the form of the group’s utility function over consequences specified by combining the member’s utility functions over the group consequences. Substituting (7) and (17) in (6), the expected utility to the group of an alternative An can be calculated from UG 4An 5 = 6çm pm 4Emj 576èm wm um 4cnmj 57 = èm wm Um 4An 51 n = 11 0 0 0 1 N 0 (18) Substituting into (18) from the individual decision analysis result in (5) yields UG 4An 5 = èm wm Um 4An 5 = èm wm 6èj pm 4Emj 5um 4cmnj 571 n = 11 0 0 0 1 N 1 (19) which completes the proof and indicates that only the marginal probability distributions over cm , m = 11 0 0 0 1 M, are required to compute the group expected utility. Corollary 1. A group probability distribution (6) that all group members can agree upon is pG 4E1j 1 0 0 0 1 EMj 5 = çm pm 4Emj 5 as specified in (7). A group utility function consistent with the group value judgments must be of the additive form uG 4c1 1 0 0 0 1 cM 5 = èm wm um 4cm 5 indicated in (17). As demonstrated in the proof of the theorem, the group probability distribution pG incorporates probabilistic independence among the probabilistic judgments of the members Im for events Emj , m = 11 0 0 0 1 M. It could be that there are probabilistic dependencies among these judgments, but no member of the decision-making group cares about this, because the chosen pG includes all the probabilistic judgments of concern to Im , m = 11 0 0 0 1 M, and the resulting analysis would be identical for any probability distribution acceptable to the group. It is useful to recognize that one could replace the GII assumption in the statement of the Group Decision Analysis Theorem with a weak Pareto indifference (WPI) assumption and the result (3) would still hold. The logic is as follows: A weak Pareto indifference assumption would state that if all of the group members were indifferent between two alternatives, then the group should be indifferent between those alternatives. If two alternatives had identical probabilities over the consequences of relevance to each group member, then all of the group members would be indifferent between the alternatives. Hence, WPI would imply that the group would be indifferent between those alternatives, so the GII assumption would hold, from which the group decision analysis theorem follows. 6. Comments on the Result The purpose of this paper is to provide a logically sound and practical result that can be used to guide prescriptive decision making for groups. To examine logical soundness, it is necessary to appraise the problem formulation, assumptions, and result of the group decision analysis theorem. To examine practicality, it is necessary to appraise how the theory can be implemented. Logical soundness is addressed in this section and implementation is discussed in §8. 6.1. Comments on the Problem Formulation The problem formulation is based on the two-stage decision process described in §3. In the first stage, each individual group member analyzes the group’s decision using the decision analysis framework for individual decision making. This framework has been accepted for more than 40 years as the basis for prescriptive decision making under uncertainty for individuals (Savage 1954, Pratt et al. 1964, Howard 1966, Raiffa 1968). The consequences cm of concern to group member Im in his or her own decision frame of the group decision are meant to include all implications that will matter to that group member from the choice of an alternative. The consequences clearly characterize what utilities are needed from the individual. They Keeney: Foundations for Group Decision Analysis Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. 12 Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS also indicate the events that matter to that individual, namely, those events that would affect the consequences that might occur. And these events characterize what probabilities are needed from the individual. Each individual is the expert on recognizing the consequences of importance in the group decision from his or her own perspective and for providing the corresponding individual probabilities and utilities. During the first stage of any group decision, significant interaction would likely occur among group members. As a result, each group member will be able to account for other members’ perspectives on consequences, events, probabilities, and utilities. At the end of this stage, each group member’s decision frame will have a common set of alternatives, but may be different in terms of all other elements (e.g., events, consequences, probabilities of events, and utilities for consequences). At this point, each group member, and the group as a whole, are agreeing that the specified consequences, events, probabilities, and utilities of group member Im , m = 11 0 0 0 1 M, are fixed and appropriately represent his or her judgments and preferences. In the second stage, the decision frame for the group decision is constructed from all of the group members’ frames, and when viewed in terms of only what is relevant to an individual group member, reduces to that individual’s frame of the decision. This group decision frame, expanded compared to those used in previous research, accounts for the reality that individual group members may perceive that different consequences and different events are relevant to the group’s decision in addition to possible different utilities for the consequences and possible different probabilities for the events. In the special case where all group members are concerned with exactly the same consequences and events, the subscript m for member on consequences cnmj and events Emj could be dropped from the problem formulation. If this were done, the main result (3) reduces to UG 4An 5 = èm wm Um 4An 5 = èm 4wm èj pm 4Ej 5um 4cnj 551 (20) and holds when members have the same or different probabilities and utilities. In any group decision, whether or not a model is used to provide insight for making the decision, group members may strategically misrepresent information about their judgments or preferences. Obviously, the quality and content of the insights potentially available from any analysis using the group decision analysis model depends on whether members strategically misrepresent information. If they do not, the potential insights guide the choice for the group based on balancing perspectives of the members. If a member does misrepresent information, a potential insight is recognition that this is the case. Often, members of the group would know each other, so they can recognize when another member’s judgments are out of line with available information or preferences are not consistent with previously stated views. Based on the circumstances, members can appraise whether this is because of strategic misrepresentation or a misunderstanding of some of the complexities of the decision. In applications, group choices about sharing the member’s judgments and preferences should be made by considering the transparency of the model, the influence of one member on others, and strategic misrepresentation. As with any application of a model, the group must subsequently decide whether and how the results of the model should be used in providing insight to help make the group decision. The relative desirability of alternatives from each decision maker’s perspective is described by his or her expected utilities of those alternatives. The interpersonal comparison of utilities and the relative importance of the different members of the decisionmaking group are both incorporated in the scaling factors denoted by wm in (3). The decision-making group must collectively decide on the values of those scaling factors as described in §8. 6.2. Comments on the Group Decision Assumptions Because each group member has transitive preferences, if the group has expressed its preferences for certain lotteries, then it seems reasonable to assume that the group should wish to make decisions consistent with the group transitivity assumption GT. Note that this assumption does not address situations where the group has not expressed preferences for a lottery, nor does it restrict in any way how group preferences are constructed. Keeney: Foundations for Group Decision Analysis 13 Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS The reasonableness of the group substitutability assumption GS follows the same logic. Because each group member accepts substitutability, if a group replaces prizes in a lottery with new prizes deemed to be equal in value to the original prizes, this should not change the group preferences from that for the original lottery. This assumption does not specify how the group determines that prizes are indifferent, but concerns only group preferences when substitute prizes are indifferent to original ones. The group judgments assumption GJ is that there exists a group judgment about the possible occurrences of all event combinations (E1 1 0 0 0 1 EM 5 in terms of a joint probability distribution. The reasonableness of this assumption is supported by the following. First, as demonstrated in the proof of the group decision analysis theorem, only the marginal probability distributions over each of the Em , m = 11 0 0 0 1 M are needed in the proof. Second, from the structure of the group decision problem, each group member Im only cares about Em and does not at all care about Er 1 r 6= m. Thus, any joint probability distribution over (E1 1 0 0 0 1 EM 5 that has the marginal probability distributions provided by Im for each Em , m = 11 0 0 0 1 M should be acceptable to all group members. Hence, the group probability distribution in (7) constructed from the group member’s probability distributions is an appropriate group probability distribution for the decision. The group preferences assumption GP is a group commitment to fulfill its collective responsibility to make a decision in a reasonable manner and not inadvertently select a significantly inferior alternative. Because each member accepts the decision analysis logic for their individual evaluation of alternatives, whereas preserving the possible distinctions in events, probabilities, consequences, and utilities, it is reasonable to assume a group utility function exists as long as it is of a form that does not require agreements on probabilities and utilities that do not reflect reality. Also, because only group member Im is concerned about consequences cm , it makes sense to use the utility function of Im over cm as the group’s utility function over cm . Consistent with assumption GP, the group utility function that represents the group’s preferences must be of the additive form (17) with scaling factors specified (discussed in §8). The logic for the group identical indifference GII assumption is as follows. The original and modified lotteries referred to in the assumption determine possible consequences of two alternatives. From any group member’s perspective, the probabilities of the possible consequences of the two alternatives are identical. Thus, consistent with the group judgment principle, the group as a whole should consider them to be identical. The group should evaluate two identical alternatives, which should be perceived as the same alternative, as equivalent, so they should be indifferent between the two alternatives. 6.3. Comments on the Group Decision Analysis Result To contrast use of the group decision analysis theorem with previous work, four cases will briefly be considered. Case 1, the most general case, is when the only thing in common in the members’ frames of the group decision is the set of alternatives. Members have different consequences, events, probabilities, and utilities. In this case, there is nothing but the expected utilities of the alternatives that could be combined to evaluate the alternatives from the group’s perspective, which is the result of the group decision analysis theorem. Case 2 is a special case of case 1, where the members agree on the consequences and events, but may disagree on the probabilities for these events and the utilities of the consequences. This frame is the one that has been implicitly assumed in most of the previous investigations of foundations for group decisions. Consistent with the group decision analysis theorem, no agreement on probabilities of events or on utilities of consequences is required. The member’s probabilities and utilities are used in the member’s decision analyses, and their expected utilities of the alternatives are weighted in (3) to provide a group expected utility. Case 3 is a special situation of case 2, where the members also agree on the probabilities of events, but may disagree on the utilities of the consequences. In this case, the result is the same as that found by Harsanyi (1955) using a Pareto optimality assumption in addition to individual and group decision analysis assumptions. Keeney: Foundations for Group Decision Analysis Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. 14 Case 4 is the most restrictive situation of a group decision, where the group members agree on everything: consequences, events, probabilities, and utilities. In this case, the group analyzes the decision exactly as each of the group members analyze the decision, and the expected utilities of alternatives for the group are the same as those calculated in each member’s analysis. Although not assumed, the result of the group decision analysis theorem is consistent with weak Pareto optimality, which is a basic assumption in much of economic literature on the group decision making, including the works of Harsanyi (1955), Hylland and Zeckhauser (1979), Seidenfeld et al. (1989), and Mongin (1995). Clearly, when each group member prefers an alternative Ai to an alternative Ak , then each of their expected utilities Um 4Ai 5 must be greater than Um 4Ak 5, so by (3), the group expected utility UG 4Ai 5 must be greater than UG 4Ak 5, therefore the group must prefer Ai to Ak . There are no restrictions in this group decision framework about whether group members can have common or overlapping events or consequences. For a group consequence that is described by (c1 1 0 0 0 1 cm 1 0 0 0 1 cM 5, there may be common aspects in each cm . On a corporate acquisition decision, each group member may include the profit due to any acquisition as part of the consequences he or she considers relevant to the group decision. Hence, in each cm , m = 11 0 0 0 1 M, that profit is included. Overall, the acquisition profit is included M times in the group’s consequence. However, because cm is a complete summary of what matters to Im , um only considers cm , so there is no double counting of preferences. The point is particularly clear in the special situation where each group member is only concerned with profit and has a utility function over profit. Because the sum of the group member’s weights is one, the group’s weight on profit is one, so there is no double counting. The group decision analysis theorem allows for a dictatorial group utility function, meaning one where one scaling factor wm = 1, and all of the others are zero. Because the group must collectively specify all of the scaling factors, in practice, we would usually expect each of the wm to be positive. Any group will essentially decide how to utilize the member’s Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS evaluations of alternatives, and could choose to follow a single individual’s analysis if they collectively decided to do so. Because of this, there seemed to be no reason to exclude the dictatorial possibility by introducing another assumption. 7. Relationship to Arrow’s Impossibility Theorem Arrow (1951, 1963) investigated a specific group decision formulation to combine the preferences of individual members of a group for the alternatives to obtain a group preference for those alternatives. In general, he wanted to obtain group preferences P for all An , n = 11 0 0 0 1 N given the individual member’s preferences Pm , m = 11 0 0 0 1 M, so PG 4An 5 = f 4P1 4An 51 0 0 0 1 PM 4An 551 (21) where f is a function. Arrow’s interest was where the preferences P were rankings of the alternatives. He postulated five assumptions and proved that they were inconsistent. Hence, an impossibility theorem resulted, which has been very influential and continues to generate great interest. Suppose one maintains the general formulation (21), but changes the preferences of the individuals and the group to be ratings instead of rankings of the alternatives. Then, if one adopts assumptions analogous to Arrow’s using ratings, specifically the expected utility of alternatives, it has been proven that the group decision analysis result (3) follows (Keeney 1976). Thus, if one assumes that each member of the group can specify the expected utilities of the group’s alternatives, the result of the group decision analysis theorem can also be derived from a different set of logical, but not decision analytic, assumptions. This is understandable if one recognizes that the formulation (21) with expected utilities of alternatives calculated by the individual group members explicitly assumes that group’s preferences for alternatives depend only on the overall ratings (i.e., expected utilities) of the group members, and implicitly assumes that the group’s preferences for alternatives depend only on the individual member’s marginal probability distributions over the consequences of alternatives. Keeney: Foundations for Group Decision Analysis Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS 8. Comments on Implementation To implement the group decision analysis result (3) requires information from each of the group members and from the group as a whole. Each member must verify that the decision analysis assumptions for an individual decision in Table 1 are appropriate for his or her own analysis of the group decision and then conduct that analysis. Because these are each individual decision analyses, the procedures to do this are well developed and understood (e.g., Raiffa 1968, Kirkwood 1997, Clemen and Reilly 2001). The group decision model does not require or allow individuals to assign probabilities to events or utilities to consequences that are not directly relevant to that individual. The group as a whole must verify the appropriateness of the group decision analysis assumptions in Table 2 and the group identical indifference assumption. Verification could occur in a discussion that clarifies the meaning of each assumption and decides on whether the group should make their group decision consistent with that assumption. The group must also collectively construct the scaling factors wm , m = 11 0 0 0 1 M to weight the individual member’s expected utilities in the group evaluation. Specifying these weights is perhaps the most difficult aspect of implementing the group decision analysis framework because (1) it requires value judgments about which the individual members may not agree, and (2) because the sum of the weights must equal one, the larger the weight given to one individual’s evaluations of the alternatives, the smaller the sum of the weights for the other individuals’ evaluations. A brief description of the conceptual foundation for specifying weights will provide a basis for discussing the practical aspects of specifying the weights and using the group decision analysis model. Each individual’s weight depends on two separate factors, the relative importance of that individual in the group and the interpersonal comparisons of group member’s utilities. These two factors can be addressed separately, and their implications are easily combined, as we will see. In many situations, the group would either explicitly or implicitly assume that the importance of all group members is equal; the basis for such a judgment is simply that being a member of the group 15 accords one an equal vote. In other situations, the importance of different members may differ depending on characteristics such as ownership, position, or seniority. For example, if three individuals jointly owned an investment with 40%, 35%, and 25% shares, the relative weights for joint decisions may be the same as these percentages. The interpersonal comparison of utilities is a more complex issue that has long been studied and discussed (e.g., Harsanyi 1955, Luce and Raiffa 1957, Sen 1970, Dyer and Sarin 1979). The essence of the interpersonal comparison of utilities issue is to compare the utility to individual I1 of going from a utility of u1 = 0 to u1 = 1, where u1 = 0 corresponds to the worst consequence for the group from the viewpoint of I1 and u1 = 1 corresponds to the best such consequence, to the utility to individual I2 of going from u2 = 0 to u2 = 1. The group may specify equivalent changes in the utility for I1 and I2 to make this judgment. For instance, the group may feel that a change from u1 = 0 to u1 = 005 is as significant to I1 as a change from u2 = 0 to u2 = 1 is to I2 . It follows that the change from u1 = 0 to u1 = 1 is twice as significant to I1 as a change from u2 = 0 to u2 = 1 is to I2 . This would imply that the ratio of w1 :w2 would be 2:1 if the only relevant concern was the interpersonal comparison of utilities. Suppose this utility comparison was for a decision where I1 owned 40% of an investment and I2 owned 60%; the importance contributions to weights were set at 0.4 and 0.6, respectively. The implication would be that the relative weights should be 2 × 004 for I1 and 1 × 006 for I2 , so the scaling factors would be w1 = 4/7 and w2 = 3/7, respectively. In practice, if there is a reason for having the relative importance of group members be different, this issue will likely come up in group discussions and be explicitly considered. Following the origin of judgments principle, it takes one member to think of this concern and bring it to the group’s attention for consideration. This may result in selecting specific judgments about relative weights as illustrated earlier. If no member brings up the issue of the relative importance of group members for the decision, then it will likely be implicitly assumed that all members should be treated as equal in this respect. For some decisions, the interpersonal comparison of utilities factor may also be unconsciously handled Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. 16 by implicitly assuming that the changes of utilities from um = 0 to um = 1 are equally significant for all Im . If this issue is brought up explicitly, it may be possible to specify weights using agreed-upon value trade-offs as discussed in the two-person example earlier and the fact that all weights must sum to one. Baucells and Sarin (2003) investigated minimum agreements among group members that are sufficient to specify the weights on the member’s utilities when the group utility function is additive. If through open discussion the group does not reach an agreement on specifying the set of member’s weights, one of two procedures may be useful. The first is to have each group member construct a set of scaling factors that he or she feels is appropriate. Analysis with each of these sets may provide insights to reach an agreed-upon decision. The second procedure, which may utilize information from the first procedure, is to jointly set bounds on the relative importance of group members and on their interpersonal utility comparisons or directly on the relative scaling factors, both of which lead to bounds on the weights. For example, a group may unanimously agree that the significance to member I1 of going from a utility of u1 = 0 to u1 = 1 is greater than, but not twice as great as, the significance to member I2 of going from u2 = 0 to u2 = 1. This would indicate that the relative weight of w1 to w2 must range between one and two. For a two-member group of equally important individuals, the weight of w1 must be between 1/2 and 2/3, with w2 = 1 − w1 . The group decision analysis using (3) can be done over the collective ranges of the weights that are acceptable to all group members of the decisionmaking group. Depending on how restrictive the weights are and on the specifics of the individual’s expected utilities for the alternatives, the acceptable ranges of weights may be sufficient to provide a unique ordering of the alternatives. In other cases, it is likely that the least desirable alternatives would be eliminated from consideration based on the analysis using the acceptable ranges of weights. If the expected utilities of the individuals have greater positive correlation, given bounds will tend to provide more definitive group evaluations of the alternatives. Keeney: Foundations for Group Decision Analysis Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS For any model to be useful, in addition to being able to obtain the information necessary to implement the model, it must be relevant to some important problems. Bordley (2009) discusses classes of important business decisions where the members of a decision-making group would reasonably identify different events as relevant to the decision. The same logic for events and consequences is relevant to many governmental and public decisions as indicated by the following case. As it happens, the group decision analysis model (3) was essentially applied to a very significant decision years ago. Dyer and Miles (1976) worked with the Jet Propulsion Laboratory to provide an analysis of NASA’s Mariner Jupiter/Saturn 1977 project alternatives. This was a 300 million (in 1977 dollars) project that included two spacecraft trajectories to be launched within a couple of weeks of each other and fly by both Jupiter and Saturn to conduct numerous scientific experiments. There were 10 groups of scientific studies to be conducted on this mission. A different scientific team had responsibility for each group of studies. A Scientific Steering Committee, composed of the leaders of each of the 10 scientific teams, had responsibility for creating and choosing trajectory pairs. Because the scientific experiments concerned distinct fields of study, such as infrared radiation, imaging science, ultraviolet spectroscopy, and cosmic ray particles, the teams naturally had different consequences of concern and different events that could affect those consequences. These characteristics are exactly those assumed in the group decision analysis frame of this paper, and the previous standard group decision analysis frame would not be appropriate for a decision with these characteristics. To evaluate alternatives, Dyer and Miles (1976) used a few different procedures including what they referred to as an “additive collective choice rule” that had intuitive appeal and is essentially the group decision analysis result (3). They had the scientific teams each provide the expected utilities for their respective experiments given each trajectory pair. In addition to other sensitivity analyses, the Steering Committee selected two sets of scaling factors for the additive collective choice rule and used each to lend insight about the relative desirability of the alternatives. Dyer Keeney: Foundations for Group Decision Analysis 17 Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS and Miles (1976) provided details about how the evaluation process helped in both guiding the Steering Committee to select an alternative trajectory pair and in honing that pair of trajectories to improve its value. Their application definitely indicates the potential relevance of the group decision analysis result to important decisions. 9. Summary There have been several attempts over the past halfcentury to extend decision analysis from individual to group decisions. None have resulted in a decision analysis framework to analyze the general case of a group decision. These attempts, including those by Raiffa (1968), Hylland and Zeckhauser (1981), Seidenfeld et al. (1989), and Mongin (1995), each resulted in impossibility theorems. In retrospect, a key contributor to the difficulty in finding a positive solution to the group decision was the implicit assumption in all of these efforts that the group decision had the same frame as an individual decision, which implies that the group members are each concerned with the same set of consequences and with the same set of possible events. As a result, the search for a solution proceeded to specify group probabilities for these events and group utilities for these consequences to produce a group decision analysis. As group members could have different judgments about probabilities of events and different preferences for the consequences, simultaneously combining both different probabilities and different utilities turned out to be problematic. The approach taken in this paper addresses the more realistic general decision problem where group members may have different perceptions, and therefore different decision frames, of their common group decision. The group decision frame explicitly incorporates each member’s frame, so it is broader than any member’s decision frame. This allows each member to incorporate his or her potentially different consequences and events of concern, as well as different probabilities of events and utilities of consequences, into the group decision. Using group decision analysis assumptions, analogous to those for an individual decision and relevant only for the specific group decision constructed from the individual member’s frames, provides a decision analysis solution for group decisions. This solution is that the group expected utility for an alternative is the weighted sum of the individual member’s expected utilities for that alternative. This result incorporates and maintains the integrity of each member’s decision analysis of what he or she feels is in the group’s interest and, in addition, explicitly addresses how the evaluations of the group members should be combined. The result is a logically sound operational framework to conduct a decision analysis of any group decision. Acknowledgments The comments of David Bell of Harvard University and Robert Nau of Duke University were very helpful and much appreciated. References Arrow KJ (1951) Social Choice and Individual Values, (2nd ed. 1963) (John Wiley & Sons, New York). Baucells M, Sarin RK (2003) Group decisions with multiple criteria. Management Sci. 49(8):1105–1118. Bordley RF (2009) Combining the opinions of experts who partition events differently. Decision Anal. 6(1):38–46. Chambers CP, Hayashi T (2006) Preference aggregation under uncertainty: Savage vs. Pareto. Games Econom. Behav. 54(2): 430–440. Clemen RT, Reilly T (2001) Making Hard Decisions with Decision Tools (Duxbury, Pacific Grove, CA). Dyer JS, Miles RF Jr (1976) An actual application of collective choice theory to the selection of trajectories for the Mariner Jupiter/Saturn 1977 project. Oper. Res. 24(2):220–244. Dyer JS, Sarin RK (1979) Group preference aggregation rules based on strength of preference. Management Sci. 25(9):822–832. Fishburn PC (1965) Independence in utility theory with whole product sets. Oper. Res. 13(1):28–45. Harsanyi JC (1955) Cardinal welfare, individualistic ethics, and interpersonal comparisons of utility. J. Political Econom. 63(4):309–321. Howard RA (1966) Decision analysis: Applied decision theory. Hertz DB, Melese J, eds. Proc. Fourth Internat. Conf. Oper. Res. (Wiley Interscience, New York), 55–71. Hylland A, Zeckhauser R (1979) The impossibility of Bayesian group decision making with separate aggregation of the beliefs and values. Econometrica 47(6):1321–1336. Keeney RL (1976) A group preference axiomatization with cardinal utility. Management Sci. 23(2):140–145. Keeney RL, Nau R (2011) A theorem for Bayesian group decisions. J. Risk Uncertainty 43(1):1–17. Kirkwood CW (1997) Strategic Decision Making: Multiobjective Decision Analysis with Spreadsheets (Duxbury Press, Belmont, CA). Luce RD, Raiffa H (1957) Games and Decisions (John Wiley & Sons, New York). Mongin P (1995) Consistent Bayesian aggregation. J. Econom. Theory 66(2):313–351. Mongin P (1998) The paradox of the Bayesian experts and statedependent utility theory. J. Math. Econom. 29(3):331–361. Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to subscribers. The file may not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org. 18 Pratt JW, Raiffa H, Schlaifer R (1964) The foundations of decision under uncertainty: An elementary exposition. Amer. Statist. Assoc. J. 59(306):353–375. Raiffa H (1968) Decision Analysis (Addison-Wesley, Reading, MA). Ramsey FP (1926) Truth and probability. Kyberg HE Jr, Smokler HE, eds. Studies in Subjective Probability (Wiley, New York), Reprinted in 1964. Savage LJ (1954) The Foundations of Statistics (John Wiley & Sons, New York). Seidenfeld T, Kadane JB, Schervish MJ (1989) On the shared preferences of two Bayesian decision makers. J. Philosophy 86(5):225–244. Sen A (1970) Collective Choice and Social Welfare (Holden-Day, San Francisco). Von Neumann J, Morgenstern O (1947) Theory of Games and Economic Behavior, 2nd ed. (Princeton University Press, Princeton, NJ). Ralph L. Keeney is a research professor emeritus at the Fuqua School of Business at Duke University. His education includes a B.S. in engineering from the University of California, Los Angeles, and a Ph.D. in operations View publication stats Keeney: Foundations for Group Decision Analysis Decision Analysis, Articles in Advance, pp. 1–18, © 2013 INFORMS research from Massachusetts Institute of Technology. His research interests are in the areas of decision making and risk analysis. He has applied such work to important personal decisions and as a consultant for private and public organizations addressing corporate management problems, environmental and risk studies, and decisions involving life-threatening risks. Prior to joining the Duke faculty, Professor Keeney was a faculty member in Management and in Engineering at MIT and at the University of Southern California, a research scholar at the International Institute for Applied Systems Analysis in Austria, and the founder of the decision and risk analysis group of a large geotechnical and environmental consulting firm. Professor Keeney is the author of many books and articles, including Value-Focused Thinking, Decisions with Multiple Objectives, coauthored with Howard Raiffa, and Smart Choices, coauthored with John S. Hammond and Howard Raiffa, which has been translated into 15 languages. Dr. Keeney was awarded the Ramsey Medal for distinguished contributions in decision analysis by the Decision Analysis Society and is a member of the U.S. National Academy of Engineering.