Experienced Utility and Objective Happiness: A Moment-Based Approach
Daniel Kahneman Princeton University
Chapter 37, pp. 673-692, in: D. Kahneman and A. Tversky (Eds.) Choices, Values and Frames New York: Cambridge University Press and the Russell Sage Foundation, (2000)
Experienced Utility and Objective Happiness: A Moment-Based Approach Daniel Kahneman1 Princeton University The concept of utility has carried two different meanings in its long history. As Bentham (1789) used it, utility refers to the experiences of pleasure and pain, the ??sovereign masters?? that ??point out what we ought to do, as well as determine what we shall do.?? In modern decision research, however, the utility of outcomes refers to their weight in decisions: utility is inferred from observed choices and is in turn used to explain choices. To distinguish the two notions I refer to Bentham??s concept as experienced utility and to the modern usage as decision utility. Experienced utility is the focus of this chapter. Contrary to the behaviorist position that led to the abandonment of Bentham??s notion (Loewenstein, 1992), the claim made here is that experienced utility can be usefully measured. The chapter presents arguments to support that claim, and speculates about its implications. This essay has three main goals: (1) to present a detailed analysis of the concept of experienced utility and of the relation between the pleasure and pain of moments and the utility of more extended episodes; (2) to argue that experienced utility is best measured by moment-based methods that assess the experience of the present; (3) to develop a moment-based conception of an aspect of human well-being that I will call ??objective happiness.?? The chapter also introduces several unfamiliar concepts that will be used in some of the chapters that follow. Pleasure and pain are attributes of a moment of experience, but the outcomes that people value extend over time. It is therefore necessary to establish a concept of experienced utility that applies to temporally extended outcomes. Two approaches to this task will be compared here. (i) The memory-based approach accepts the subject's retrospective evaluations of past episodes and situations as valid data. The remembered utility of an episode of experience is defined by the subject??s retrospective global assessment of it. (ii) The moment-based approach derives the experienced utility of an episode from real-time measures of the pleasure and pain that the subject experienced during that episode.
Moment-utility refers to the valence (good or bad) and to the intensity (mild to extreme) of current affective or hedonic experience. The total utility of an episode is derived exclusively from the record of moment-utilities during that episode. The main novelty of the treatment proposed here is that it is thoroughly momentbased. Section 2 reviews some of the evidence that raises doubts about the validity of memory-based assessments. Section 3 presents the conditions that must be satisfied to
I thank Shane Frederick, Barbara Fredrickson, Laura Gibson, David Laibson, Nathan Novemsky, David Schkade, Cass Sunstein, Richard Thaler, Anne Treisman, and Peter Wakker for helpful comments. I also thank Peter Wakker for allowing me to use ideas and sentences that we had fashioned together, and Peter McGraw for extremely valuable assistance and useful suggestions.
permit an assessment of the total experienced utility of episodes from the utilities of their constituent moments. Section 4 introduces a moment-based concept of objective happiness, and examines the feasibility of its measurement. Section 5 exposes the ambiguity of a central idea of the well-being literature -- the hedonic treadmill -- and discusses how measures of objective happiness could contribute to the resolution of that ambiguity. A research agenda and some major objections are discussed in Section 6.
Figure 1: Pain intensity reported by two colonoscopy patients
Patient A 10 8 6 4 2 0 0 10 20
Patient B 10 8 6 4 2 0 0 10 Time (minutes)
The main concepts of the present treatment are illustrated by Figure 1, which is drawn from a study of immediate and retrospective reports of the pain of medical procedures (Redelmeier and Kahneman, 1996). Patients undergoing colonoscopy were asked every 60 seconds to report the intensity of their current pain, on a scale where 10 was ??intolerable pain?? and 0 was ??no pain at all??. These ratings were used to construct the profiles of moment-utility shown in the figure. The patients later provided several measures of the remembered utility of the procedure. They evaluated the entire experience on a scale, and they also compared it to a standard set of aversive experiences.
The total utility associated with each patient??s colonoscopy has a different nature. Unlike moment-utility and remembered utility, it is not an expression of a subjective feeling or judgment. Total utility is an objective assessment of the statistics of a utility profile. The cases of patients A and B also illustrate the contrast between remembered
utility and total utility. It is immediately apparent from inspection of Figure 1 that patient B had a worse experience than patient A,2 and this impression will be confirmed by the analysis of total utility in section 3 (see Figure 2). However, patient A in fact retained a worse evaluation of the procedure than patient B. In this case, as in many others, remembered utility and total utility do not coincide, and outcomes will be ranked differently depending on whether experienced utility is assessed by a memory-based or by a moment-based method. 2. Memory-Based Assessment: Remembered Utility Anyone who has cared for an elderly relative whose memory is failing has learned that there is a crucial difference between two ostensibly similar questions. The question ??How are you now??? may elicit a confident and cogent answer while the question ??How have you been??? evokes only confusion. This distinction is rarely drawn in other settings. We normally expect people to know how they have been as well as they know how they are. Memory-based evaluations of experience and reports of current pleasure and pain are treated with equal respect in routine conversations -- but the respect for memory is less deserved. Studies of the psychology of remembered utility are reviewed in detail in the next chapter. The main conclusions of this research are listed below, and illustrated by the colonoscopy study from which Figure 1 was drawn. Duration neglect. No one would deny that it is generally better for a colonoscopy to be short than to be long. At least in principle, then, the duration of a colonoscopy is relevant to its overall utility. However, memory-based assessments do not generally conform to this principle. For example, the colonoscopies studied by Redelmeier and Kahneman (1996) varied in duration between 4 and 69 minutes, but the correlation between the duration of a procedure and the patient??s subsequent evaluation of it was only .03. Furthermore, the duration of the colonoscopy had no effect on patients' hypothetical choice between a repeat colonoscopy and a barium enema. Complete, or nearly complete neglect of duration has been found in other studies, using a variety of different research designs. A hypothesis of 'evaluation by moments' is introduced in chapter 38 to explain these findings: it asserts that the remembered utility of an episode is determined by constructing a composite representative moment, and by assessing the utility of that moment. The Peak/End rule. The patients' subsequent evaluation of the procedure was predicted with relatively high accuracy (r = .67) from the average of the most intense level of pain reported during the procedure, and of the mean pain level reported over the last three minutes. Because the Peak/End average was higher for patient A than for patient B, this
On the assumption that the two patients used the pain scale similarly. This issue is discussed further in section 3.
empirical rule predicts -- correctly -- that patient A would retain a more aversive memory of the colonoscopy than would patient B. Strong support for the Peak/End rule was obtained in several other studies, reviewed in detail in the following chapter. Violations of dominance. The Peak/End rule implies a counter-intuitive prediction: adding a period of pain to an aversive episode will actually improve its remembered utility, if it lowers the Peak/End average. For example, several extra minutes at pain level 4 would be expected to improve patient A's global evaluation of the procedure. A clinical experiment with 682 patients undergoing colonoscopy tested this prediction. Half of the patients were randomly selected for an experimental treatment, in which the examining physician left the colonoscope in place for about a minute after terminating the examination. The patient was not informed of the manipulation (Katz, Redelmeier and Kahneman, 1997).3 The extra minute is distinctly uncomfortable, but not very painful. The effect of the experimental treatment was to reduce the Peak/End average for patients, such as patient A, who would otherwise have experienced considerable pain in the final moments of the procedure. As predicted by the Peak/End rule, retrospective evaluations of the procedure were significantly more favorable in the group that experienced the prolonged procedure than in the group that was treated conventionally. Similar violations of dominance were also observed in choices: in one experiment, participants were exposed in immediate succession to two unpleasant sounds of similar composition. One of them lasted for 10 seconds at 78 db; the other consisted of the same 10 seconds at 78 db, followed by 4 seconds at 66db. When given an opportunity to choose which of the two sounds would be repeated later, most participants chose the longer (Schreiber and Kahneman, 2000; [ch. 38]). This choice is odd, because 4 seconds of silence would clearly be preferable to 4 seconds of 66 db noise. In this simple situation, decision utility appears to be determined directly by remembered utility: people choose to repeat the sound they dislike least, and the Peak/End rule determines that. 3. Moment-Based Assessment: Total Utility The evidence reviewed in the preceding section suggested that memory-based assessments of experienced utility should not be taken at face value. The present section introduces a moment-based alternative, in which the total utility of an episode is derived from a temporal profile of moment-utility. The same analysis extends to related episodes separated in time, because utility profiles may be concatenated. For example, the total utility of a Kenya safari should include subsequent occasions of slide-showing and reminiscing. Figure 2a presents the data of Figure 1 in the form of a decumulative function, which shows the amount of time spent at or above each pain level. If the measure of moment-utility
on which it is based satisfies a stringent set of conditions, total utility can be derived from the type of representation illustrated in Figure 2 (Kahneman, Wakker
The ethical justification for the experiment was the observation of poor compliance among patients who have had a painful colonoscopy and are instructed to schedule another.
and Sarin, 1997). Six conditions are listed below. The first four impose requirements on the measure of moment-utility. The last two conditions are normative in character; they specify how total utility is constructed from moment-utilities. Figure 2a: Decumulative temporal function for pain profiles of Patients A & B
25 20 15 10 5 0 0 2 4 6 8 10 Pain Intensity
Patient A Patient B
Figure 2b: Fictitious decumulative functions representing the objective happiness of two individuals over a period of time
100 80 60 40 20 0 -10 -5 0 5 10 Moment Utility
Inclusiveness. In a moment-based approach the utility profile is a 'sufficient statistic' to determine the experienced utility of an extended outcome. The measure of moment-utility should therefore incorporate all the aspects of experience that are relevant to this evaluation. In particular, a measure of moment-utility should reflect the affective consequences of prior events (e.g., satiation, adaptation, fatigue), as well as the affect associated with the anticipation of future events (fear, hope). Ordinal measurement across situations. The measurement of moment-utility must be ordinal or better. Experiences of different types (e.g., a stubbed toe and a humiliating rebuke) must be measured on a common scale. Distinctive neutral point. The pain scale that was used in the colonoscopy study has a natural zero point. However, the dimension of moment-utility is bipolar, ranging from intensely positive to neutral, and from neutral to intensely negative affect. A distinctive neutral point ("neither pleasant nor unpleasant", "neither approach nor avoid") anchors the scale and permits comparisons across situations and persons.4 As will be seen later, a stable zero is also essential for cardinal measurement of moment-utility on a ratio scale. Interpersonal comparability. The scale must permit comparisons of individuals and groups. The next section shows that this requirement may be more tractable than is commonly thought. The next two requirement are of a different nature. They involve normative assumptions about the nature of total utility. The assumptions of separability and timeneutrality are required to justify the transformation of utility
profiles (e.g., Figure 1) into the decumulative format (e.g., Figure 2). The discussion of these assumptions highlights a critical difference between the present analysis and economic models of the utility of sequences of outcomes. These models generally describe outcomes as physical events (see, e.g., chapters 32-33). The analysis of total utility, in contrast, describes outcomes as moment-utilities. Separability: the order in which moment-utilities are experienced does not affect total utility. Order effects are ubiquitous in experienced utility. For example, a strenuous tennis game and a large lunch yield a better experience in one order than in the other, because the enjoyment of the tennis game is sharply reduced when it follows lunch. The condition of separability states that the contribution of an element to the global utility of the sequence is independent of the elements that preceded and followed it. This condition is often violated when the sequences are described in terms of physical events, such as lunch and a tennis game. In a moment-based treatment, however, the elements of the sequence that is to be evaluated are not events -- they are moment-utilities associated with events. Because all the effects of the order of events are already incorporated into moment-utilities, separability can be assumed for these moment-utilities. Separability is necessary for the decumulative representation, which does not preserve order
Some authors consider valence as bi-valent rather than bipolar (e.g., Cacioppo, Gardner, and Berntson, 1999).
information. To appreciate the intuition, consider an individual who receives two unexpected prizes in immediate succession, one of $500, the other of $10,000, then promptly dies, or loses his memory. In evaluating the total utility of these experiences, we recognize that it would be better for the two prizes to arrive in ascending rather than in descending order ?C presumably because the enjoyment of the smaller prize is greater when it comes first. Now imagine that all you know is that just before he died (or became amnesic) an individual had two pleasurable experiences with utilities Ua and Ub, where Ua>> Ub, Would we still think that their order matters? When outcomes are moment-utilities, there is no compelling reason to reject separability. Time neutrality: all moments are weighted alike in total utility. Total utility is a measure on completed outcomes, and is therefore always assessed after the fact. Unlike decision making, in which the temporal distance between the moment of decision and the outcome may matter, the temporal distance between an outcome and its retrospective assessment is entirely irrelevant to its evaluation. Total utility is therefore time neutral. In this important respect, it is unlike decision
utility and remembered utility, which both assign more weight to some parts of the sequence than to others. The decision utility of outcomes that occur late in a sequence is often heavily discounted. In remembered utility, on the other hand, the last parts of a sequence are weighted more than those that came earlier -- a bias that is incorporated in the Peak/End rule. The normative status of both weighting schemes is dubious. If the benefits are obtained before the costs must be paid, discounting of delayed outcomes in decisions favors myopic preferences for options that do not maximize total utility. The overweighting of endings may be equally unreasonable: an experience that ended very badly could still have positive utility overall, if it was sufficiently good for a sufficiently long time (Kahneman, Wakker and Sarin, 1997). Measures of total utility The representation of Figure 2 assumes both that a utility profile can be rearranged at will and that all its parts are weighted equally. Separability and time neutrality are therefore necessary, and together with the assumptions of inclusiveness and ordinal measurement, sufficient for the representation of utility profiles as decumulative temporal distributions. The total utility of episodes is a measure on these distributions. Figure 2 illustrates two representations of temporal distributions of utility, which differ in their ordinates: time is shown in absolute units in panel (a), but in proportional units in panel (b). The representation of panel 2a is appropriate when the duration of the episode is relevant to its evaluation. Thus, it is reasonable to say that the colonoscopy of patient B was worse than that of patient A because it lasted longer. On the other hand, it does not make sense to say that Helen was happier last week than she was last Sunday because last week was longer than last Sunday. The representation of Figure 2b is correct when the duration of the period of evaluation is not relevant to its evaluation. It is the appropriate representation in the assessment of the well-being of individuals and groups, which is discussed in the next section.5
There are situations in which both representations are relevant. The total utility (or happiness) that Alan enjoyed while he was married to Helen may depend on how long they were
As Figure 2a illustrates, the ordinal measurement of moment-utility permits the detection of distributional dominance. By this simple test, patient B had a worse colonoscopy than patient A. The decumulative distribution can also be characterized by non-parametric statistics, such as the median and other fractiles. However, distributional dominance is a blunt measuring instrument, and no single non-parametric index captures all the relevant information contained in a temporal distribution of moment-utilities. Figure 2b presents decumulative
distributions of moment-utility for two individuals, George and Helen. There is no dominance in this comparison, and the medians are close. The main conclusion that can be drawn is that George experienced more extremes of affect than Helen did. Cardinal measurement of moment-utility would be desirable, of course. With cardinal measurement, the most natural index of total utility could be calculated: the temporal integral of moment-utility. The idea has a long history (Edgeworth discussed it in 1881) but it effectively requires a rescaling of moment-utility in terms of physical time, which is difficult to implement. This reasoning is explicit in the use of QALYs (Quality Adjusted Life Years) in medical decision making. QALY's are derived from judgments of equivalence between periods of survival that vary in duration and in level of health. For example, two years of survival at a QALY of 0.5 are equally desirable as one year in normal health (Broome, 1993). A formal analysis of the temporal integration rule was offered by Kahneman, Wakker and Sarin (1997). Their treatment invoked all the assumptions that were discussed in this section, including separability and temporal neutrality. In addition, it introduced an idealized objective observer, who assesses the total utility of utility profiles, such as those of Figure 1. The following axioms specify the logic of this assessment. 1. The global utility of a utility profile is not affected by concatenation with a neutral utility profile. 2. Increases of moment-utility do not decrease the global utility of a utility profile. 3. In a concatenation of two utility profiles, replacing one profile by another with a higher global utility will increase the global utility of the concatenation. The following theorem can be proved: "The three axioms above hold if and only if there exists a non-decreasing ("value") transformation function of moment-utility, assigning value 0 to 0, such that global utility orders utility profiles according to the integral of the value of moment-utility over time." The proof is due to Peter Wakker. The representation of total utility as a temporal integral implies a scale of moment-utility, monotonically related to the original scale, but now calibrated by its relation to duration. For example, suppose that an idealized observer who conforms to the axioms judges that 1 minute of pain that had been rated 7 on the original scale is equivalent to 2 minutes at a rating of 6. On the transformed scale, the value that corresponds to the original rating of 7 will be double the value assigned to a rating of 6. Idealized observers are hard to find, of course, and cardinal scaling of utility is therefore
married before she died in an accident. On the other hand, an assessment of how happy Alan was in his marriage should not be influenced by how long it lasted.
a conceptual exercise rather than a practical procedure. Fortunately, the decumulative representation is adequate for many purposes. The conditions identified by Kahneman, Wakker and Sarin (1997) are sufficient to guarantee this representation, without attempting cardinal measurement and without involving observers. 4. Objective Happiness: Concept and Measure Moment-utility is the building block of the broader construct of experienced utility. It is also the building block for a construct of objective happiness, with which the remainder of this chapter is concerned. Like total utility, objective happiness is to be derived from a distribution of moment-utility (see Figure 2b for an example) that characterizes an individual (George or Helen), a group (Californians, midwesterners, paraplegics), or a setting (the Washington subway, the New York subway). Like total utility, objective happiness is a moment-based concept, which is operationalized exclusively by measures of the affective state of individuals at particular moments in time. In this essential respect, objective happiness differs from standard measures of subjective well-being, which are memory-based and require the subject to report a global evaluation of the recent past. The term ??objective?? is used because the judgment of happiness is made according to objective rules. The ultimate data for the judgment are, of course, subjective experiences. In the special conditions of the clinic or laboratory it is sometimes possible to obtain continuous or almost continuous reports of experienced utility from patients or experimental subjects. Continuous measures are of course impractical for the measurement of objective happiness over a period of time. Sampling techniques must be used to obtain a set of values of moment-utility that adequately represents the intended population of individuals, times and occasions. For example, a study of the objective happiness of Californians should use a sample of observations that reflects the relative amounts of time spent on the freeway and in hot tubs. Techniques for sampling times and occasions have been developed in the context of Experience Sampling Methodology (ESM) (Csikszentmihalyi, 1990; Stone, Shiffman and DeVries, 1999).6 Reporting the sign and intensity of current hedonic and affective experience is not essentially different from the standard psychophysical tasks of reporting color or smell. The report of affect is probably intermediate in difficulty between these tasks: somewhat more difficult than labeling colors, but much easier than describing smells. The worlds of affective experience and of color experience are similar in another important respect: they combine phenomenological richness with a simple underlying structure. A nonintuitive finding of color research is that, in spite of the enormous variety of subjective color experience, the world of color can be represented in a two-dimensional space -- the color circle -- with additional
information provided in a third dimension of luminance. A major result of research on affect is that a similarly simple structure is found in that domain as well. Much of the variation among affective states is captured by specifying
Participants in studies using ESM carry a palmtop computer that beeps at random times during the day. The palmtop computer then displays questions that elicit elicit information about the current setting and about the subject??s present affective state.
their positions in a two-dimensional space, which is defined by the major dimensions of valence (good to neutral to bad) and arousal (from frenetic to lethargic) (Plutchik and Conte, 1997; Russell, 1980; Russell and Carroll, 1999; Stone, 1995; Warr, 1999). As Figure 3 illustrates, the two-dimensional structure permits a distinction between two forms of positive affect?ªexuberant joy or serene bliss?ªand two forms of negative affect?ªagitated distress or apathetic depression.
Figure 3: A representation of affective space
Agitated Distress Joyful Enthusiasm
A significant limitation of the two-dimensional representation of affect is that it does not capture the nature of primary emotions, such as surprise or anger. Another objection to this scheme questions the assumption that valence is a single bipolar dimension. Cacioppo, Gardner, and Berntson (1999) point out that positive and negative affect are processed by different neural systems and may be activated concurrently. They suggest that a three-dimensional representation may be necessary, in which 'good' and 'bad' are independent dimensions. However, the systems are not functionally independent, and there is evidence that they inhibit each other. Lang (1995) has shown, for example, that watching pleasant pictures of food or smiling babies attenuates the startle response to a loud sound, whereas startle is actually enhanced in the presence of disgusting or otherwise awful pictures. For the present purposes, the description of valence as a bipolar dimension can be retained as a useful approximation, even if it is not