Decision Making in Human Performance
Many serious accidents in which human error has been involved can be attributed to faulty operator decision making: The decision to launch the Challenger Space shuttle, which later exploded because cold temperature at launch time destroyed the seals is one example; another is the 1987 decision of personnel on board the USS Vincennes to fire on an unknown aircraft,
which turned out to be a civilian Iranian transport rather than a hostile fighter (U.S. Navy, 1988). However a contrasting tragic decision was made by those on board the USS Stark cruising in the Mediterranean a year before, not to fire on an approaching target which turned out to be hostile and launched a missile which cost several lives on board the Stark.
Types of Decision Environments
We usually consider the uncertainty of the decision as involving risk. The decision to purchase one of two possible vehicles is generally low risk if one has done advanced research on product quality, since the probable outcomes of one purchase or the other are known. But the decision to proceed with a flight in uncertain weather may have a high amount of risk, since it is difficult to predict in advance what impact the weather will have on the safety of the flight.
Time:. Time plays at least two important roles in influencing the decision process. First, we may contrast “one shot” decisions like the choice of a purchase, with evolving decisions like those involved in treating an uncertain disease, in which test is followed by medication which may be followed by further tests and further treatment. Secondly, time pressure has a critical influence on the nature of the decision process
Familiarity and Expertise: As we discuss later, experts can often look at a decision problem and intuitively, nearly instantly pick the correct choice, whereas novices may ponder the problem for some time, and perhaps make a poor choice
Two different schools of decision analysis and research, naturalistic decision making (Zsambok & Klein, 1997, high skill and expertise: system 1) and the heuristics/ biases approach to decision making (system 2; Kahneman & Klein, 2009).
The study of rational or normative decision making (e.g., Edwards, 1987) has focused its efforts on how people should make decisions according to some optimal framework; for example, one that will maximize the expected gain or minimize the expected loss.
The cognitive or information processing approach to decision making focuses more directly on the sorts of biases and processes that reflect limitations in human attention, working memory, or strategy choice, as well as focuses on common decision routines—known as heuristics—that work well most of the time, but occasionally lead to undesirable outcomes.�Less emphasis here is placed on departures from optimal choice per se, and more on understanding the causes of such biases in terms of the structure and limits of the human as an information processing system.
Human Information Processing
The model depicts a series of processing stages or mental operations that typically (but not always) characterizes the flow of information as a human performs tasks. Consider as an example the task of driving toward an intersection. As shown on the left of Figure 1.1, events in the environment are first processed by our senses—sight, sound, touch, etc.—and may be held briefly in short term sensory store (STSS) for no more than about a second. Thus the driver approaching the intersection will see the traffic light, the flow of the environment past the vehicle, and other cars, and may hear the radio and the conversation of a passenger. But sensation is not perception, and of this large array of sensory information only a smaller amount may be actually perceived (e.g., perceiving that the light has turned yellow). Perception involves determining the meaning of the sensory signal or event, and such meaning is, in turn, derived from past experience (a yellow light means caution). As we see below, this past experience is stored in our long term memory of facts, images, and understanding of how the world works. After perception, our information processing typically follows either (or both) of two paths. At the bottom, perceiving (understanding) a situation will often trigger an immediate response, chosen or selected from a broader array of possibilities. Here the driver may choose to depress the accelerator or apply the brake, a decision based on a variety of factors, but one that must be made rapidly. Then, following response selection, the response is executed in stage 4 of our sequence in a manner that not only involves the muscles, but also the brain control of those muscles.
But perception and situation understanding do not always trigger an immediate response. Following the upper path from perception, the driver may use working memory to temporarily retain the state of the light (yellow), while scanning the highway and the crossing road ahead for additional information (e.g., an approaching vehicle or a possible police car). In fact, in many cases an overt action does not follow perception at all. As you sit in lecture you may hear an interesting fact from the lecturer, but choose not to take notes on it (no response selection and execution), but rather to ponder it, rehearse it, and learn it. That is, to use working memory to commit the information to long term memory, for future use on an exam or in applications outside of the classroom. Thus, the function of working memory is not just to store information, but also to think about it: the process of cognition.
To this four-stage + memory model, we add two vital elements, feedback and attention. First, in many (but not all) information processing tasks, an executed response changes the environment, and hence creates a new and different pattern of information to be sensed, as shown by the feedback loop at the bottom. Thus, if the driver applies the accelerator, this will not only increase the perceived speed of the car, but also may reveal new sensory information (e.g., a police car is suddenly revealed waiting behind a sign), which in turn may require a revision of the stop-go response choice. Second, attention is a vital tool for much of information processing, and here it plays two qualitatively different roles (Wickens & McCarley, 2008). In its first role as a filter of information that is sensed and perceived, attention selects certain elements for further processing, but blocks others, as represented in Figure 1.1 by the smaller output from perception than input to it. Thus, the driver may focus attention fully on the traffic light, but “tune out” the conversation of the passenger or fail to see the policeman. In the second role attention acts as a fuel that provides mental resources or energy to the various stages of information processing, as indicated by the dashed lines flowing from the supply of resources at the top. Some stages demand more resources in some tasks than others. For example, peering at the traffic light through a fog will require more effort for perception than seeing it on a clear, dark night. However, our supply of attentional resources is limited, and hence the collective resources required for one task may not allow enough for a concurrent one, creating a failure in multi-tasking
Information Processing Model
Beginning at the left, the decision maker seeks cues or information from the environment.
Selective attention of course plays a critical role in decision making, in choosing which cues to process (of higher perceived value) and which to filter out. Such selection is based on past experiences (long-term memory) and requires effort or attentional resources
The cues that are then selected and perceived now form the basis of an understanding, awareness, or assessment of “the situation” confronting the decision maker (see Chapter 7), a process that is sometimes labeled diagnosis (Rasmussen & Rouse, 1981).
Many decisions are iterative in the sense that initial hypotheses will trigger the search for further information to either confirm or refute them. Troubleshooting a system failure will often trigger repeated tests to confirm or refute possible hypotheses (Hunt & Rouse, 1981). This characteristic defines the important feedback loop to cue filtering, labeled “confirmation” in Figure 8.1. The entire process of cue seeking and situation assessment has been labeled the “front end” of the decision process (Mosier & Fischer, 2010). Following from the front end stages of cue seeking and situation assessment (or diagnosis), the third principle stage in decision making is the choice of an action, described as the “back end” of decision making (Mosier & Fischer, 2010). From long term memory the decision maker can generate a set of possible courses of action or decision options; but if the diagnosis of the state of the world is uncertain (as it is in much decision making), then the possible consequence of the different choices define their risks. Consideration of risk requires the explicit or implicit estimation of two quantities: (1) the probability or likelihood that different outcomes will come to pass and (2) values, the extent to which those outcomes are “good” or “bad.” This is directly analogous to the discussion in Chapter 2, where the decisions made in signal detection theory depended upon both probability and the values (costs and benefits) imposed on different outcomes (hits, false alarms, misses, correct rejections). Thus the physician will probably consider the values and costs of various outcomes before she decides which treatment (do nothing, drugs, surgery) to recommend for a patient’s abnormality of uncertain identity.
Bounded Rationality
An example of bounded rationality in individuals would be a customer who made a suboptimal decision to order some food at the restaurant because they felt rushed by the waiter who was waiting beside the table. Another example is a trader who would make a moderate and risky decision to trade their stock due to time pressure and imperfect information of the market at that time.
Limitations include the difficulty of the problem requiring a decision, the cognitive capability of the mind, and the time available to make the decision. Decision-makers, in this view, act as satisficers, seeking a satisfactory solution, with everything that they have at the moment rather than an optimal solution. Therefore, humans do not undertake a full cost-benefit analysis to determine the optimal decision, but rather, choose an option that fulfills their adequacy criteria.
bounded rationality can be said to address the discrepancy between the assumed perfect rationality of human behaviour (which is utilised by other economics theories), and the reality of human cognition
Rational vs. Naturalistic Decision Making
What is Good Decision Making
First, early decision research of the normative school offered the expected value of a decision as the “gold standard.” That is, the decision that would produce the maximum value if repeated numerous times (Edwards, 1987; see Section 6.1). However, defining expected value depends upon assigning universally agreed upon values to the various possible outcomes of a choice; values are often personal, making this a difficult undertaking. Even if values could be agreed upon, the choice that might be optimal if the decision is repeated time and again with plenty of time for weighing all the cues will not necessarily be optimal for a single choice, particularly one made under time pressure with little time to fully diagnose the situation and consider all possible outcomes (Zsambok & Klein, 1997). Furthermore, for a single decision, the decision maker may be more concerned about, say, minimizing the maximum loss (worst case) rather than maximizing expected long-term gain which after all can only be realized following a longterm average of the outcome of several decisions.
Second, one may say that “good” decisions are those that produce “good” outcomes and bad decisions conversely produce bad outcomes, such as the decision to launch the Challenger space shuttle, to fire on the Iranian Airliner from the USS Vincennes
After all, considering the USS Vincennes case (a “bad” decision), the decision makers on board the ship must also have considered that the decision made a year earlier on board the USS Stark, not to fire upon an approaching contact, turned out also to be “bad,” leading to the loss of life on the Stark. This tendency to label a decision as good or bad only after the outcome is known is sometimes called the hindsight bias.
A third approach to decision quality has been based upon the concept of expertise (Zsambok & Klein, 1997, Kahneman & Klein, 2009; Brehmer, 1981; Shanteau, 1992, see Chapter 7). Since experts in other fields (e.g., chess, physics) are known to produce “good” and sometimes exceptional performance, why not consider that expert decision makers do the same. The problem here is that several analyses of decision making have shown that experts in certain domains do not necessarily make better decisions than novices
We adopt the approach here that, to the extent that all three of the characteristics described above converge then it becomes increasingly easy to discriminate good from bad decision making. But when they do not, then such discrimination is often fruitless, and it is much more appropriate simply to look at the qualitative ways in which different environmental and informational characteristics influence the nature of the processing operations and outcomes of the decision process
We adopt the approach here that, to the extent that all three of the characteristics described above converge then it becomes increasingly easy to discriminate good from bad decision making. But when they do not, then such discrimination is often fruitless, and it is much more appropriate simply to look at the qualitative ways in which different environmental and informational characteristics influence the nature of the processing operations and outcomes of the decision process
Estimating cue: perception
Systematic biases have been observed in perceiving and estimating three other characteristics of the environment: proportions, projections, and randomness.
PROPORTIONS: e.g., faulty versus normal parts on an inspection line. Biases may result from an inherent conservative tendency (“never say never”), or alternatively they may result from the greater salience, noticeability or impact of the single outlying observation (which is, by definition, the infrequent event) in the sea of more frequent events. For example, if I have seen 99 normal parts, then detecting the one abnormal part will make more of an impact on my consciousness than detecting a 100th normal one. Its greater impact could well lead me to overestimate its relative frequency in hindsight, even as the rarity of the abnormal part will make me less likely to detect it in the first place if its abnormality is not salient
The tendency to overestimate the frequency of rare events from description (versus experience, as above) has important implications for choice behavior. For example, people appear to show little difference in behavior (e.g., purchasing lottery tickets) whether odds of an event (winning) is 1/1000, or 1/10,000, and thereby implicitly overestimating the probability of the latter (Slovic & Finucane et al., 2002). They consider both as equal evidence for the possibility of winning rather than as different evidence for the probability of winning.
PROJECTIONS
With regard to projection, humans are not always effective in extrapolating non-linear trends. As shown in Figure 8.2, they often bias their estimates toward the more linear extrapolation of the tangent where the data end (Waganaar & Sagaria, 1975; Wickens, 1992). This parallels the challenges people have in predicting the dynamic behavior of systems to be tracked
Exponentially increasing temperatures will often trigger fire extinguishing efforts, or opening pressure relief valves that will reduce the rate of growth. So the long-term memory of experience will lead the decision maker—accurately—to infer that the rapidly growing quantity will eventually slow its rate of growth.
RANDOMNESS
People do not do a good job in perceiving (or understanding) randomness in the environment (Tversky & Kahneman, 1971). This is best illustrated by the gambler’s fallacy in observing (or acting on) a series of dichotomous events, like coin tosses or wins and losses in a gamble. People tend to think that “random” implies a heavy bias toward alternation between the two outcomes. When generating a random series of say heads and tails, people will tend to avoid generating
Evidence Accumulation. Selective Attention: Cue Seeking and Hypothesis Estimation
We can represent the diagnostic stage of decision making as a process by which the decision maker receives a series of cues, symptoms, or sources of information as shown near the bottom, bearing on the true (or predicted) state of the world, and attends to some or all of these with the goal of using those cues to influence the cognitive belief in one of several alternative hypotheses shown at the top right. In many instances, we can represent this as a “belief scale,” between two alternative hypotheses, H1 and H2, as shown in the figure.
Evidence Accumulation. Selective Attention: Cue Seeking and Hypothesis Estimation
1. Cue diagnosticity formally refers to how much evidence a cue should offer regarding one or the other hypothesis. Thus, if one sees rain drops falling, this is a 100 percent diagnostic cue that it will be raining; on the other hand, a forecast of “a 50 percent chance of showers” is a totally undiagnostic cue for precipitation. The diagnosticity of any cue can be expressed both in terms of its discriminating value (high or low) as well as its polarity (i.e., which hypothesis the cue favors).
2. Cue reliability or credibility refers to the likelihood that the physical cue can be believed. Thus an eyewitness to a crime may state categorically that “the suspect did it” (high diagnosticity); but if the witness is a notorious liar, his or her reliability is low. Collectively, both diagnosticity and reliability can be expressed on scales of 0 to 1.0, and then their product can reflect the information value of a cue. If the decision maker views a cue with an information value =1 (d=1 × r=1), then that single cue is all that needs to be processed to make an error free diagnosis. However, most diagnostic problems have cues with information value less than 1.0, and hence can produce circumstances in which cues conflict. (Consider opposing witnesses for the defense and prosecution in a legal trial.) 3. The physical features of the cue which can make it conspicuous or salient have an important bearing on the selective attention and the subsequent processing that it receives.
First, selective attention must be deployed to process the different cues, ideally giving different weight according to their information value. Second, the cue values—raw perceptual information—must be integrated, analogous to the bottom up processing of perceptual features in pattern recognition. Third, expectancies or prior beliefs may play a role in biasing one hypothesis or belief to be favored over the other, analogous to the way that expectancies stored in long term memory influence the top down processing in perceptual pattern recognition and signal detection, Fourth, an operation that is not paralleled by those in perceptual pattern recognition, is the iterative testing and retesting of the initially formed belief, to attain the final belief which is the basis for choice.
Having established the role of reliability and diagnosticity in determining the information value of a cue, we are then in a position to establish the optimal degree of belief in one hypothesis or another on the basis of multiple cues.
The process of attending to and integrating multiple cues typically located at different places and/or delivered at different times along various sensory channels presents a major challenge to human selective attention and hence can be a source of four major vulnerabilities, as we discuss below.
INFORMATION CUES ARE MISSING: A decision maker may not have all the information at hand to make an accurate diagnosis. An operator’s judgment to turn on a faulty piece of equipment cannot be blamed if the operator was not informed by maintenance personnel of the equipment failure. But thwarting this process is the fact that present cues can be perceived, while realizing the existence of absent cues depends upon memory. Thus, the effective planner of a mission will attempt to obtain, and rely on, only the most recent weather data, and if the available forecast is outdated may postpone a decision until a weather diagnosis can be made only on the most recent data.
CUES ARE NUMEROUS: INFORMATION OVERLOAD: As we have noted, when the information value of any cue is known to be 1.0 (both reliability and diagnosticity = 1.0), then other information need not be sought. But this is rarely the case, and so effective diagnosis will rely upon integrating multiple cues.The operators monitoring any nuclear plant in the face of a major failure may be confronted with literally hundreds of indicators, illuminated or flashing. Which of these should then be attended first, as the operator then tries to form a diagnosis as to the nature of the fault. When several different information sources are available, each with less-than-perfect information value, the likelihood of a correct diagnosis can increase as more cues are considered. In practice, however, as the number of sources grows beyond two, people generally do not use the greater information to make proportionately better, more accurate decisions. When few cues are initially presented, this filtering is unnecessary. When several sources are present, however, the filtering process is required, and it competes for the time (or other resources) available for the integration of information. Thus, more information leads to more time-consuming filtering at the expense of diagnostic quality.
CUES ARE DIFFERENTIALLY SALIENT: the salience of a cue, its attention-attracting properties or ease of processing, can influence the extent to which it will be attended and weighted in information integration. salient information cues and are likely to be given greater weight, particularly under time pressure
These findings lead us to expect that in any diagnostic situation, the brightest flashing light or the meter that is largest, is located most centrally, or changes most rapidly will cause the operator to process its diagnostic information content over others: the salience bias.
. It is important for a system designer to realize, therefore, that the goals of alerting (high salience) are not necessarily compatible with those of diagnosis in which salience should be directly related to the information value of the cue in making a diagnosis, not just in detecting a fault.
An extreme case of low salience relates to the absence of a cue. There are often circumstances in which a hypothesis can gain credibility on the basis of what is not seen as well as what is seen. For example, the computer or automotive troubleshooter may be able to eliminate one hypothesized cause of failure on the basis of a symptom that is NOT observed.
PROCESSED CUES ARE NOT DIFFERENTIALLY WEIGHTED BY INFORMATION VALUE:
\ While people will tend to overprocess cues of greater salience, there is also good evidence that people tend to overprocess cues of lesser information value relative to those of greater value.
People do not effectively modulate the amount of weight given to a cue based upon its information value, whether the latter is influenced by diagnosticity or reliability. Instead, they tend to treat all cues as if they were more or less of equal value (Cavenaugh, Spooner, & Samet,
1973; Schum, 1975). This as-if heuristic thereby reduces
the cognitive effort which would otherwise be required to
consider differential weights when integrating information.
Numerous examples of the as-if heuristic can be identified,
downweighting differences in information value. As one example
, Griffin and Tversky (1992) found that evaluators, forming
impressions of an applicant on the basis of
letters of recommendations, tended to give
more weight to the tone or enthusiasm of the
letter (a salient feature) than to the credibility
or reliability of the source (the letter writer)..
In the context of Figure 8.3, information regarding reliability can be said to be less accessible to cognition than the actual diagnostic content of that information, and hence is ignored (Kahneman, 2003)
. It is important for a system designer to realize, therefore, that the goals of alerting (high salience) are not necessarily compatible with those of diagnosis in which salience should be directly related to the information value of the cue in making a diagnosis, not just in detecting a fault
Expectations in Diagnosis: The Role of Long-Term Memory
REPRESENTATIVENESS: The foundation of the representativeness heuristic (Kahneman & Frederick, 2002; Tversky & Kahneman, 1974) is that cues for a diagnostic state are often correlated. Thus, for example, bad weather is diagnosed by both clouds and low pressure. When making a diagnosis, people tend to match the observed case pattern against one of a few possible patterns of symptoms (one for each diagnosis) learned from past experience and stored in long-term memory. If a match is made, that diagnosis is chosen.
There is nothing really wrong with following this heuristic except that people tend to use representativeness without adequately considering the base rate, probability, or likelihood that a given hypothesis or diagnosis might actually be observed.
In a manner similar to the failure to differentially weight cues, discussed above, Christenssen-Szalanski and Bushyhead (1981) have observed that physicians are insufficiently aware of disease prevalence rates (base rate) in making diagnostic decisions.
For example, following the representativeness heuristic, a physician observing a patient who matches four out of five symptoms typical of disease X, and three out of five typical of disease Y will be likely to diagnose disease X as being most representative of the patient’s symptoms, even if X occurs very rarely in the population, compared to disease Y.
Representativeness may be thought to reflect another example of the distorting effects of salience or accessibility in decision making.
However, this mental representation of probability may also be imperfect, as reflected in the second important heuristic in evidence consideration, the availability heuristic.
THE AVAILABILITY HEURISTIC: This heuristic can be employed as a convenient means of approximating prior probability, in that more frequently experienced events or conditions in the world generally are recalled more easily
Availability and accessibility are closely related to the phenomenon of a attribute substitution (Kahneman, 2003) in which certain highly accessible mechanisms get substituted by the intuitive (type 1) decision system, for more effort-demanding mechanisms employed by the analytic (type 2) system when resources are scarce.
As one simple example, when people make choices in a gamble, they are often heavily influenced by the probability of winning or losing between two options, rather than the expected value of the two options.
Interestingly, representativeness (the pattern of data) and availability (estimating frequency of hypothesis) are two commodities that are integrated together in the Bayesian approach to optimal decision making
Belief Changes Over Time
As we have noted, many diagnoses are not the short, “one shot” pattern classifications, but rather take place over time as an initial tentative hypothesis may be formed, and more evidence is sought (or arrives) to confirm or refute it. In this process of refining beliefs over time, we can identify two important characteristics that can sometimes work against the most accurate estimate of the “truth”: the anchoring heuristic and the confirmation bias.
ANCHORING HEURISTIC: Describes how, when cues bearing on a hypothesis, or information sources bearing on a belief arrive over time, the initially chosen hypothesis tends to be favored, as if we have attached a “mental anchor” to that hypothesis and do not easily shift it away to the alternative. If evidence a favors hypothesis A and b favors B, then receiving the evidence in the order should lead to a favoring of A, but receiving it in the order will favor B. It should be noted that while anchoring represents a sort of primacy in memory, there is also sometimes a recency effect in cue integration, in that the most recently encountered of a set of cues may, temporarily, have a strong weighting on the diagnosis (Rieskamp, 2006). Thus the lawyer who “goes second” in presenting closing arguments to a jury may well leave the jury with a bias toward that side, in making their judgment of guilt or innocence.
THE CONFIRMATION BIAS: Describes a tendency for people to seek information and cues that confirm the tentatively held hypothesis or belief, and not seek (or discount) those that support an opposite conclusion or belief.
Arkes and Harkness (1980) demonstrated the selective biasing of memory induced by the confirmation bias. People have less cognitive difficulty dealing with positive information than with negative information (Clark & Chase, 1972, see Chapter 6), and with the presence of information (a present cue that supports what you already believe) than the absence (the absence of a cue which, if present would support your belief), also reflecting cognitive effort.
There is a motivational factor related to the desire to believe. The high value that people place on consistency of evidence leads them to see all (or most) evidence supporting one or the other belief, and that belief is usually the one initially formulated.
A second motivational factor results when people focus more on the consequences of the logical choice of action that would follow from the initially favored hypothesis, rather than the truth of that hypothesis itself.
DECISION FATIGUE: Repeated decisions can often lead to decreased effort invested in accuracy and analysis. This phenomenon was illustrated dramatically in an analysis of parole board decisions carried out by Danzigera, Levav, and Pesso (2011), who observed that the probability of granting parole declined from 75 percent early in the morning, down to approximately 25 percent later in the day.
CHOICE OF ACTION
The output of decision making must also include a choice of some action.One key feature of this choice, is not relevant for diagnosis but was clearly represented by signal detection theory is the value that the decision maker places on different possible outcomes. One key feature of this choice, is not relevant for diagnosis but was clearly represented by signal detection theory is the value that the decision maker places on different possible outcomes. We consider below, how people “should” and how they do combine information on value and probability to make decisions,
Certain Choice
an array of possible objects (e.g., products) are compared, each with varying attributes. For example the set of personal computers to purchase may vary in their attributes of price, usability, maintainability, warrantee, and so forth. In making such a choice that will maximize the consumer’s overall satisfaction, the decision maker should carry out the following steps:
This decision process is known as a compensatory one, in that a product which may be low on the most important attribute (an expensive computer, when cost is most important), can still be chosen if this deficiency is compensated for by high values on many other attributes of lesser importance. For example the most expensive computer may have far and away the best user interface, the most reliable maintenance record, and the best warrantee, allowing these strengths to compensate for the weakness in price.
For example the rule of satisficing (Simon, 1955) is one in which the decision maker does not go through the mental work to chose the best option, but rather one that is “good enough” (Lehto, 1997), and this is often the strategy employed in real-world naturalistic decision making, when there is time pressure.
A more systematic heuristic that people sometimes employ when the number of attributes and objects is quite large, is known as elimination by aspects. Here, for example, the most important attribute is first chosen, then any product that does not lie within the top few along this attribute (aspect) is eliminated from consideration, and then the remaining products are evaluated by comparing more of the aspects of the remaining few objects. As a heuristic, this technique will easily reduce the cognitive effort of needing to compare all attributes across all objects. And it will usually prove satisfactory, only failing to pick a satisfactory choice if an object that is low on the most important attribute (and hence eliminated) happens to be near the top on all others. Understandably, the EBA heuristic is one that begins to dominate over time, as people suffer the effort depletion of decision fatigue (Tierney, 2011).
Expected Value and Utility Models
Information Processing Model in DM
Choice Under Uncertainty: The Expected Value Model
Many decisions are made in the face of uncertainty regarding their future consequences. Unlike those choices discussed in the previous section in which the consequences of the choice were relatively well known, many decisions are made in the face of uncertainty regarding their future consequences. Such uncertainty may result because we do not know the current state of the world; for example a physician may choose a particular treatment, but be uncertain about the diagnosis. Indeed we can often represent decision making under uncertainty as shown in Figure 8.6, by providing the possible states of the world (A, B, C, . . .) across the top of a matrix, each associated with their estimated probability or likelihood, and the possible decision options (1, 2, . . .) down the rows.
Choice Under Uncertainty: The Expected Value Model
While expected value maximization is clear, simple, and objective, there are several factors that complicate the picture when it is applied to most human decisions under uncertainty. First, it is not necessarily the case that people want to maximize their winnings (or minimize their expected losses) over the long run. For example, they may wish to minimize the maximum loss (i.e., avoid picking the option which has a catastrophic negative outcome value). This is, of course, one reason why people purchase fire insurance and avoid the decision option of “no purchase”, even though the expected value of the purchase option is negative in the long run
Second, in many decisions it is not easy to assign objective values like money to the different outcomes. A case in point are decisions regarding safety, in which consequences may be human injury, suffering, or the loss of life. Third, as we discuss in the following section, people do not treat their subjective estimates of costs and values as linearly related to objective values (i.e., of money). Fourth, people’s estimates of probability do not always follow the objective probabilities that will establish long term costs and benefits.
Heuristics and Biases in Uncertain Choice
We consider below first a shortcut or heuristic related to direct retrieval that totally bypasses the explicit considerations of risk, and then the forms of influences of human perception of value and of probability, which have been incorporated in to a theory of choice known as prospect theory.
DIRECT RETRIEVAL: . Choices of action may sometimes be implemented simply on the basis of past experience. If the conditions are similar to those confronted in a previous experience, and an action worked in that previous case, it may now be selected in the present case with confidence that it will again produce a satisfactory outcome. This direct retrieval strategy is a hallmark on naturalistic decision making. s. So long as the domain is familiar to the decision maker, and the diagnosis of the state of the world is clear and unambiguous, the comparative risks of alternatives need not be explicitly considered. Sometimes such an approach may be coupled with a mental simulation (Klein & Crandall, 1995), in which the anticipated consequences of the choice are simulated in the mind, to assure that they produce a satisfactory outcome.
DISTORTIONS OF VALUES AND COSTS: LOSS AVERSION: It is assumed that humans are trying to maximize an expected utility rather than expected value (Edwards, 1987), in which utility is the subjective value of different expected outcomes. Within this context, the important principle of loss aversion specifies that people are more concerned about (greater loss in utility) the loss of a given amount of value, than they appreciate (increase in utility) a gain of the same amount (Garling, 1989; McGraw et al., 2010). This difference is explicitly represented as one important component of the prospect theory of decision making, proposed by Kahneman and Tversky (1984) as shown in Figure 8.7, which relates objective value on the x-axis to subjective utility on the y-axis. To the right, the figure represents the functions for utility gains (receiving money or other valuable items). To the left, it represents the functions for losses. Certain features of this curve nicely account for some general tendencies in human decision making
TEMPORAL DISCOUNTING: Differences between value and utility are also reflected in a phenomenon known as temporal discounting. Here people often make decisions or chose options that maximize the short term gains (an immediate positive experience) rather than postponing them (a delayed utility) for an option that may result in equal or even greater long term gains; this behavior reflects an
implicit belief that the passage of time
“discounts” those gains.
PERCEPTION OF PROBABILITY:
Consistent with these biases in prospect
theory, Kahneman and Tversky (1984)
have suggested a function relating true
(objective) probability to subjective
probability (as the latter is inferred to
guide risky decision making) that is
shown in Figure 8.8.
THE FRAMING EFFECT: In its simplest version, the framing effect accounts for how people’s preference for outcomes and objects change as function of how their description is framed (Tversky & Kahneman, 1981). For example the same ground beef product will seem more attractive if it is described as 80 percent lean than if it is described as 20 percent fat, even though the product is identical in the two descriptions. People will be more likely to choose the beef (over some other meat) with the former description, framed in the positive, than the negative. More seriously, a physician considering treatment of a severely ill patient may have the treatment outcomes listed as a 98 percent chance of survival or a 2 percent chance of mortality. Again, both options describe the same probabilistic outcome. But skilled medical personnel will tend to choose the treatment (over the option, for example doing nothing) more often with the former positive frame than with the 2 percent negative frame (McNeil, Pauker, et al., 1982). A classic example, faced by most of us at some time or another is when we chose between adhering to some time (or cost) consuming safety procedure (a sure loss), or adopting the risk of avoiding the procedure (driving too fast, running the red light, failing to wear safety glasses) because the cost of compliance outweighs our expected benefits of enhanced safety. The importance of these differences between perceived losses and gains is that a given change in value (or expected value) may often be viewed either as a change in loss or a change in gain, depending on what is considered to be the neutral point or frame of reference for the decision making; hence the title of the framing effect.
The Decision to Behave Safely
At least three factors appear to be responsible for the fact that people elevate their estimate of risk above the true “objective” values associated with, for example probability of death. The first is the fact that publicity, for example from the news media, tends to make certain risks more available to memory than others (Combs and Slovic, 1979). Hence we observe the high perceived risks of well publicized events (like a major plane crash or a terrorist bombing). Second, people’s perception of risk is driven upward by what is described as a “dread factor” (uncontrollable, catastrophic consequences, unavoidable), and third, perceived risk is inflated by an “unknown” factor, which characterizes the risk of new technology, such as genetic manipulations and many aspects of automation (Slovic, 1987).
An important way to mitigate risky behavior when it results because the probability of the negative event may be very rare (and hence never personally experienced) even as its negative consequences may be severe, is through “gentle reminders” (Hertwig & Erev, 2009). This technique imposes minor penalties—a gentle reminder—for the risk-producing behavior (e.g., failing to heed a safety precaution) which will be experienced much more frequently than the rare severe consequences.
Effort
Our treatment of decision making up to now has focused most on the extermal drivers of decision making—problem structure, risk, values, and probability—as filtered by human cognition. However, as shown in Figure 8.1, there are two critical inputs to the decision process emanating from the decision maker himself or herself: effort and meta-cognition.
In our discussion of decision fatigue, we emphasized that effective decision making often requires effort. Resource-dependent working memory is necessary to diagnose and evaluate options. Decision making competes for those resources with concurrent tasks (e.g., Sarno & Wickens, 1995; see Chapter 10), and sustained decision making depletes that pool of resources or cognitive effort (Tierney, 2011). Indeed, it has been shown that repeated decision making competes with the effort required for exerting self control in other aspects of life (e.g., resisting temptations; Tierney, 2011)
The effort required and accuracy observed of these two classes of DM strategies is reflected schematically in Figure 8.9, which indeed previews the concept of the performance-resource function. Within this context, effort itself can be viewed as a valuable resource to be conserved.
The contingent model of decision strategies developed by Payne, Bettman, and Johnson, 1993, predicts how different strategies will be chosen, contingent upon the available time (resources).
Meta-cognition and Over-confidence
The issues of anticipated effort and accuracy, and the conscious choice of a decision strategy brings us to the important role of meta-cognition in decision making. What does the decision maker know (or think) about the accuracy of his diagnosis and choice? How does this anticipation influence the choice of strategy and subsequent decision-making behavior (including the choice not to decide at all, as in the case of the parole boards discussed in Section 5.4.3). As Kahneman and Klein (2009) note, this is the role of the type 2 system: to oversee, review, and audit the more automatic decision-making behavior of the type 1 system.
It turns out that one of the most critical and enduring influences on meta-cognition is the confidence in assessing ones own diagnosis and judgment. Such confidence is often unrealistically high, as manifest in the overconfidence bias (Nickerson, 1998). In diagnosis, confidence judgments will influence the extent to which we jump into action (choice), rather than seek more evidence, or prepare for the case in which the assessment may have been wrong. In choice, confidence assessments will influence the extent to which we plan for alternative actions (to the extent that we think our chosen action might have been wrong). The average driver estimates him/herself to be within the top 25th percentile of safe drivers (Brehmer, 1981). By definition, if confidence were calibrated, this should be 50 percent. OC is well documented in the planning fallacy (Buehler Griffin & Ross, 2002). Here people are eternally optimistic in their projections of how long it will take (or how many resources will be required) to do something, from achieving a personal goal (like turning in an assignment on time), to completing massive construction projects
like the Denver International Airport or the
Sidney Opera house. Indeed in one study,
students expressed 84 percent confidence
that they would complete an assignment
on time, whereas in fact, only 40 percent
did so (Buehler et al, 2002).
In a pattern reflecting the solid arrow of Figure 8.10, when people rely on progressively more sources of correlated information, they gain confidence (Kahneman & Klein, 2009).
Diagnostic or problem difficulty. Evidence reliability.
Expert vs. Novice Decision Processes
In front end decision making (diagnosis), experts typically manifest recognition primed decision making (RPDM). Here through repeated exposure to the same set of correlated cues, leading to the same state assessment, experts are able to automatically classify the appropriate state, almost the same as the automatic pattern recognition discussed in Chapter 6. Hammond et al. (1987) refer to this as holistic decision making, a function associated with decision system 1 (Kahneman & Klein, 2009).
Furthermore, as we have considered before, practice in decision making does not necessarily make perfect, as it does in other skills. Expertise in some decision-making tasks does not guarantee immunity to certain biases and heuristics (Kahneman & Klein, 2009; Taleb, 2007; Tetlock, 2005).
Feedback is often ambiguous, in a probabilistic or uncertain world. That is, sometimes a decision process will be poorly executed, but because of good luck will produce a positive outcome; at other times, a decision process can follow all of the best procedures, but bad luck produces a negative outcome.
Feedback is often delayed. In many decisions, such as those made in investment, or even prescribing treatment in medicine, the outcome may not be realized for some time.
Feedback is processed selectively. As the decision maker may process feedback from this process, we note that he or she will typically only have available feedback from those who were admitted (and succeeded or failed), rarely learning if the people excluded by his decision-making rule would have succeeded had they been admitted.
Skill-, Rule-, and Knowledge-Based Behavior
Limits of Expertise
Training and Debiasing
Proceduralization and Checklists
Decision Support and Visualization
Group and Team Decision Making
Role of Automation in Decision Making
Automation Bias and Mistrust
Levels of Automation
Human–Automation Teaming
Improving Decision Outcomes
Training debiasing: Research has focused on more targeted practice and instructions to remove or reduce many of the biases discussed above, a technique known as debiasing Some success in reducing the confirmation bias has also been observed by the training strategy of “consider the opposite” (Mussweiler et al., 2000).
Also successful is a kind of training aid designed to provide more comprehensive and immediate feedback in predictive and diagnostic tasks, so that operators are forced to attend to the degree of success or failure of their rules. Finally, in an interesting take on debiasing training, Fischhoff (2002) described the success of some training programs designed to reduce the prevalence of teens engaging in risky behavior (drinking, speeding).
Proceduralization
Displays
Automation and Decision Support Tools
While debiasing is a form of training that often focuses people’s awareness directly on understanding the sources of their cognitive limitations, proceduralization simply outlines prescriptions of techniques that should be followed to improve the quality of decision making (Bazerman, 1998). This may include for example prescriptions of following the decision decomposition steps of diagnosis and choice theory, as shown in Figures 8.5 and 8.6 (Larrick, 2006). Such a technique has been employed successfully in certain real world decisions which are easily decomposable into attributes and values, such as selecting the location of the Mexico City airport (Kenney, 1973), or assisting land developers and environmentalists to reach a compromise on coastal development policy (Gardner & Edwards, 1975).
There is good evidence that effective displays can support the front end of decision processes (cue integration and diagnosis), by assisting the deployment of selective attention (Mosier & Fischer, 2010). For example, Stone, Yates, and Parker (1997) observed that pictorial representations of risk data supported more calibrated risk decisions than do numerical or verbal statements. S
Finally, automation and expert systems have offered promise in supporting human decision making.
such support can be roughly categorized into front end (diagnosis and situation assessment) and back end (treatment, choice, and course-of-action recommendations) support.
Integration Across Chapters
Review and Discussion
Summary of Key Concepts