Heuristics and Public Policy: Decision-making Under Bounded Rationality

Introduction

How do human beings make decisions under certainty, risk, subjective uncertainty, and true uncertainty? ¹

Aiming to clarify debates about both rationality and public policy, we have three goals here. The first is to offer a disciplined, contemporary overview of departures from BRA in human behaviour, with special emphasis on the role of heuristics. As we shall show, recent advances have allowed far more precision and formalization. The second is to demonstrate that although many advances have been made, and far more remains to be learned, the fundamental claims of the original work on heuristics, undertaken by Daniel Kahneman and Amos Tversky in the 1970’s, remain largely intact. Many objections to that work are rooted in fundamental misunderstandings of its purposes and also its central claims. The third is to demonstrate that for purposes of law and policy, libertarian paternalism, or nudging, does not depend on controversial psychological claims. A GPS device is helpful to human beings—no matter how we think about heuristics, and even if we agree that heuristics generally work well, in the sense they are helpful in the contexts in which most people use them.

Our starting point is that for many years the BRA was almost an article of faith in economics. It was the dominant approach taught in most economics departments, with relatively little discussion of the empirical validity of the assumptions that lie behind it or, indeed, any discussions of an alternative. Even now, most of the leading textbooks in microeconomics and game theory, which lay the foundation for the subject, continue to offer little or no empirical motivation for the theoretical models they use.

The behavioural economics revolution in economics has made significant progress in incorporating a more accurate understanding of human behaviour (Dhami, 2016; Gintis, 2009; Kahneman & Tversky, 2000; Thaler, 2015). To an increasing degree, economics courses give attention to behavioural models and findings. It is important to see that much (not all) of behavioural economics still uses the optimization framework, but relaxes almost everything else. What if individuals do not optimize, or are simply not able to optimize? In this case, behavioural economics draws upon the influential work of Tversky and Kahneman (1971, 1974) to illustrate a range of simple rules of thumb (heuristics) that are fast and frugal (in terms of the time and information required) to solve economic problems (Dhami, 2016, Part 7; Kahneman, 2011; Kahneman, Slovic, & Tversky, 1982).

The work of Kahneman, Tversky, and other researchers (abbreviated as KT&O) on heuristics shows persuasively that the behaviour of people is biased relative to the BRA; hence the name, heuristics and biases. That work has demonstrated that the BRA in economics is not tenable, not even in an ‘as if’ sense. Consistent with our first goal, we aim to show that the heuristics used in the KT&O programme are not merely labels. Attempting to go beyond the first decades of work, we explain that most heuristics can be given precise mathematical definitions and are consistent with the bulk of the evidence (see the section ‘The KT&O Approach’). ²

We consider a range of other objections to the KT&O programme, as it is best understood, and find that the criticisms are overstated or simply do not stand up to a fuller scrutiny. This includes the nature of probability and subject errors; the ‘we cannot be that dumb’ critique (see the section “We Cannot Be that Dumb” Critique’); lack of specification of empirical counterparts of the proposed heuristics (‘Empirical Counterparts to the Heuristics’ section); the use of the recognition heuristic to explain an event (the gambler's fallacy) and its negation (the hot hands fallacy) (see the section ‘How Can Heuristics Explain Events A and Not A?’); and the question of an appropriate statistical norm (see the section ‘The Criticism of Appropriate Statistical Norms’). In the ‘Systems 1 and 2’ section, we briefly outline the System 1 and System 2 distinction proposed by Kahneman (2011) and suggest that it is useful in understanding the nature of heuristics used by humans.

The work of Gerd Gigerenzer and others (abbreviated as G&O) on fast and frugal heuristics draws from and overlaps with the KT&O programme, especially insofar as it shows that heuristics generally work well, as Kahneman and Tversky repeatedly emphasized. It claims to find its motivation in the original work of Herbert Simon (1955) that stressed the procedural rationality of solutions to problems (see the section ‘The G&O Programme’). The stated domain of the G&O programme is large worlds (or true uncertainty), although in some cases, notably the priority heuristic, it can also deal with small worlds (or risk and subjective uncertainty). We consider the foundational elements of the G&O programme in the section ‘The G&O Programme’.

G&O begin with the plausible idea, also found in KT&O, that people may have an adaptive toolbox of heuristics from which they draw heuristics depending on the context and frame of the problem (see the section ‘Ecological Rationality’). In principle, this is eminently plausible, and consistent with the stated objectives of procedural rationality. But theories cannot be judged on plausibility. We attempt to shed more light on the points of disagreement with the KT&O approach, which has come to be known as the great rationality debate in psychology (Dhami, 2016, Section 19.15; Stanovich & West, 2000).

The analysis of the problem of decision-making under true uncertainty is arguably more challenging than under any of the other cases (certainty, risk, subjective uncertainty). We believe, based on the evidence, that the G&O programme has not yet given us a persuasive account of decision-making under true uncertainty. Economics has no optimization benchmark to offer in the case of true uncertainty, which makes it difficult to evaluate the performance of any candidate heuristic (see the section ‘What Is the Benchmark for Comparison Under Risk and Uncertainty’); Gigerenzer (2008) is aware of this issue. The G&O programme has compared the performance of its proposed heuristics against benchmarks claimed to be optimization benchmarks, typically logistic regression or weighted tallying. However, none of these benchmarks under true uncertainty is persuasive, and certainly not recommended by any optimization theory in economics. We also consider the pros and cons of training people in the use of statistics (see the section ‘Training People in Using Statistics’).

The section on ‘On Distinct Domains of Choices in KT&O and G&O’ argues that the domains of choice in the KT&O and the G&O programms are often non-overlapping. KT&O largely considered situations of certainty, risk, and uncertainty (sometimes called the small worlds situation) to which the BRA applies; their aim was to test the BRA. On the other hand, G&O are typically interested in the large worlds situation (true uncertainty), where one cannot list or even imagine the possible outcomes and/or objective/subjective probabilities. ³

The G&O programme often differentiates itself from the KT&O programme on the following grounds in common and in published discourse (Gigerenzer, 2008, 2014; Gigerenzer, Todd, & the ABC Group 1999) that we paraphrase as follows: (1) The KT&O programme suggests that people are fallible, hardwired with defective mental software; and prone to errors; that heuristics are bad; and that the appropriate normative norm of human behaviour is BRA (Gigerenzer, 1996, 2014; Gigerenzer et al., 1999). (2) By contrast, the G&O programme is designed to show that heuristics are good, and do better than optimization methods once ecological rationality is taken into account.

This distinction is unhelpful and inaccurate. KT&O meant to establish the importance of heuristics, not to say that they are good or bad. To the extent that they addressed that topic, it was to say that the heuristics they identified generally worked well, but also led to severe and systematic errors (which is demonstrably true). Their main goal was to test if human behaviour was consistent with the BRA; and they found it was not. But they went further by identifying various classes of heuristics that explain human behaviour in different contexts and frames. The following two passages from Kahneman (2000, p. 682) show just how close the two sides in the debate really are on core issues: (1) ‘Contrary to a common perception, researchers working in the heuristics and biases (HB) mode are less interested in demonstrating human irrationality than in understanding the psychology of intuitive judgement and choice.’ (2) ‘All heuristics make us smart, more often than not...’

We also offer some claims about law and public policy. Welfare economics often assumes that people follow the BRA and make informed choices that are in their best interests; such preferences are termed as normative preferences. However, the evidence from behavioural economics suggests that individuals sometimes do not make choices that are in their best interests (Dhami, 2016; Dhami & al-Nowaihi, 2018; Thaler & Sunstein, 2009). This may be the case when individuals lack information, have limited attention, misperceive risks, or face self-control problems arising from various forms of present-biased preferences (and have imperfect awareness of such problems). Thus, individuals might undersave for retirement; not enroll in pension plans; consume various goods (such as cigarettes) that harm the quality of their life; make decisions in an emotional hot state that they regret later; and procrastinate in making choices.

Libertarian paternalism (LP), or more simply ‘nudging’, has been an effort to help human beings to avoid errors while also preserving freedom of choice. A GPS device is an example. It allows people to go their own way, but helps them to arrive at their preferred destination. Most broadly, LP is an effort to increase navigability. It can be useful when people use heuristics that produce errors. When they are working well, policies that are consistent with LP do not distort the choices of those who follow the BRA (or do so minimally), but significantly improve the welfare of those who do not. Consider warnings and reminders, which may overcome the problem of limited attention. Use of default options, an example of an LP policy, has been extremely effective in a large number of domains, and can be more cost-effective than traditional tax/subsidy and direct regulation methods advocated in classical welfare economics (Dhami, 2016, Part 8; Sunstein, 2016, 2017; Thaler, 2015; Thaler & Sunstein, 2009).

As we shall show, endorsement of LP does not depend on a commitment to psychological claims that may be controversial. Most puzzlingly, the perceived adversarial position between the KT&O and G&O approaches, described above, has nonetheless given rise to a critique of the LP approach (Gigerenzer, 2015). The section on ‘Nudges and Nudging’ considers three main elements of this critique. (1) Choice architects (those who enact the LP policy) may not be benevolent; rather, they may be self-interested or malicious. (2) The rationale for nudges lies in people's alleged irrationality. (3) Nudges ignore other policy interventions that might help. None of these criticisms has merit.

The ‘lack of a benevolent policymaker’ criticism applies to any economic policy, not just nudges; this is analyzed in political economy, a well-established field in economics. If policymakers are not be benevolent, the strongest objections should be to mandates and bans, not against nudges, which maintain freedom of choice (in part because of an insistence that policymakers may err). A primary reason for nudging, as opposed to mandates and bans, is precisely the possibility that policymakers may not be benevolent (or adequately informed).

Those who embrace nudges do not speak of irrationality. The rationale for many successful nudges lies in a lack of information, in limited attention, and in self-control problems (and imperfect awareness of such problems). None of these should be controversial, certainly not in the abstract. Objections to LP, based on psychological claims, do not grapple with what LP means in practice (Halpern, 2015; Sunstein, 2013)–that is, with the particular policies that advocates or practitioners of LP have embraced in actual policymaking roles.

The final section concludes.

In this section, we give a simple taxonomy of situations that have played an important part in the debate on the appropriate heuristics to use. ⁵

Let an individual take some action a (e.g., buy a stock) that leads, in the future, to a finite set of outcomes, say, x 1 , … , x n (say, returns on the stock) with respective probabilities p 1 , p 2 , … , p n that are non negative and add up to 1. In the simplest possible setting, where individuals play a game against nature (i.e., no strategic interaction is involved), economists are typically interested in the following four situations: certainty, risk, uncertainty and ambiguity.

Under certainty, the individual makes choices among outcomes that are certain to occur in the future. For instance, the returns on bonds (rather than stocks) are fixed and certain. Which of a set of bonds should the decision maker choose?

Example 1. (Ellsberg paradox) Suppose you have two urns, a known urn U and an unknown urn K. In Urn K are 50 Red and 50 Green balls. In Urn U are 100 balls that are either red or green but the proportions are unknown. You are promised a prize of $10 if a red ball comes up. You assign an objective probability of 0.5 of a red ball being drawn from Urn K. Also, using the principle of insufficient reason, you may assign a probability 0.5 of a red ball being drawn from Urn U. Yet, when most people are asked which urn they will choose to bet on, they prefer Urn K to Urn U, despite assigning equal probabilities of drawing a red ball from each. It is as if each urn is a different ‘source’ of uncertainty. This is known as the Ellsberg paradox.

There is fifth class of extremely important situations on which economic theory has little to offer by way of models or predictions.

5. True uncertainty: A situation of true uncertainty (or Knightian uncertainty) arises when the outcomes and probabilities are unknown or unimaginable and objective/subjective estimates of these outcomes and probabilities are not available. In terms of the Ellsberg paradox in Example 1, true uncertainty would arise, if one does not know how many colours/balls are present in Urn U. The human eye can distinguish between 10 million colours and the number of balls can be very large. So, if you are an economic theorist who believes that this situation can be accommodated within classical subjective uncertainty, then try to imagine the possible colours and the Cartesian product of colours and the number of balls over which you must form a subjective distribution. It is far fetched to imagine that humans could do this sort of thing, even if in principle we could allow for this possibility and the use of subjective uncertainty. As a practical matter, it is better to classify such a situation as one of true uncertainty. There is some overlap between ambiguity and true uncertainty, but true uncertainty is a much broader class.

Example 2 illustrates situations that are potentially consistent with true uncertainty. ⁷

Example 2. True uncertainty seems to be reflected in the judgements of even experienced market participants. ⁸

‘I think there is a world market for maybe five computers.’ (Thomas Watson, 1943, President of IBM)

‘Fooling around with alternating current (AC) is just a waste of time. Nobody will use it, ever.’ (Thomas Edison, 1889)

”Television won't be able to hold on to any market it captures after the first six months. People will soon get tired of staring at a plywood box every night.’ (Darryl Zanuck, 1946, 20th Century Fox).

”Nuclear powered vacuum cleaners will probably be a reality within 10 years.’ (Alex Lewyt, 1955, President of the Lewyt Vacuum Cleaner Company).

‘There is practically no chance communications space satellites will be used to provide better telephone, telegraph, television or radio service inside the United States.’ T.A.M. Craven, 1961, Federal Communications Commission (FCC) commissioner.

‘Remote shopping, while entirely feasible, will flop.’ (Time Magazine, 1966).

‘There's just not that many videos I want to watch.’ (Steve Chen, 2005 CTO and co-founder of YouTube expressing concerns about his company's long-term viability).

In Example 2, Thomas Watson might have, in 1943, foreseen everything possible related to the computer industry in the future (e.g., all possible innovations, supplies, demands, tastes, and prices) for all future time periods and assigned subjective probabilities to all these events. He might have then have used expected utility to make a considered prediction. Some social scientists might be perturbed and amused to note that this would be the standard approach taken in mainstream economics. However, we are very doubtful of this possibility. It is more likely that Thomas Watson faced true uncertainty and chose his best guess. Arguably, some of the most important problems that decision makers face belong to the domain of true uncertainty. Economics does not have clear predictions to offer in these cases. Yet, people often do make decisions in such cases. How do people make these decisions? Indeed, heuristics may offer the most critical insights in dealing with these problems.

The KT&O Approach

When economists deal with situations 1-4 above, they typically use the BRA; see our opening remarks. Is the behaviour of people consistent with the BRA? Even if people do not literally follow the BRA, do they behave ‘as if’ their behaviour is consistent with the BRA? The second question is motivated by Milton Friedman's suggestion to evaluate economic theories by the accuracy of their predictions, not the realism of their assumptions. Consistency of human behaviour with the BRA is taught and accepted as an article of faith in most leading economics programmes in the world, an approach that has been followed in some other social sciences and in management programmes.

Tversky and Kahneman (1971, 1974) squarely address these questions in what is now known as the heuristics and biases research programme (we abbreviate this programme by KT&O to also acknowledge the important role of many ‘others’ who have made important contribution).The KT&O programme is one of the most significant achievements in all of social science. ⁹

The most enduring and important contribution of the KT&O programme has been to show that people do not act ‘as if’ they use the BRA. On mathematical optimization in the BRA, in particular, the Nobel laureate Herbert Simon (1978) notes: ‘But there are no direct observations that individuals or firms do actually equate marginal costs and revenues.’ Another Nobel laureate, Reinhard Selten, has argued on many occasions that economic problems are NP-Complete, a term that is borrowed from the computer science literature. The solution to such problems is hard/impossible to obtain; while no analytical solution may obtain, numerical solutions may be possible with high powered computers and sophisticated algorithms. Yet, in the BRA, the person on the street is assumed to solve such problems in his head, in an instant. Despite the enormous significance of the KT&O programme for the BRA, it is hardly taught in economics degrees.

In recent decades, the KT&O approach has been massively influential, particularly in finance and public policy. But it has come under intense criticism from another school within bounded rationality. The work of Gerd Gigerenzer and others, which we abbreviate as the G&O approach, identifies Herbert Simon's bounded rationality approach as its intellectual fountainhead. ¹⁰

The debate between KT&O and G&O has come to be known as the great rationality debate in psychology (Dhami, 2016, Section 19.5; Stanovich and West, 2000). We believe that several aspects of this debate appear to rely on confusions and misunderstanding of the relevant subject matter. We try to clarify several aspects of this debate below. ¹¹

The KT&O Programme with Definitions and Labels

A major criticism of KT&O by G&O is that the relevant heuristics are not stated formally, so it is not clear what they mean, and anything goes (Gigerenzer, 1991, 1996; Gigerenzer & Gaissmaier, 2011). For instance, talking of representativeness and availability, Gigerenzer (1991, p. 102) writes that these are ‘largely undefined concepts and can be post hoc used to explain almost everything.’

G&O have also criticized KT&O for not providing precise labels for the heuristics that they use (Gigerenzer & Brighton, 2009). The term labels appears to us to mean two things: (a) precise definitions of the heuristics and (b) specification of the exact models in which these heuristics are situated. In this section, we consider the first of these meanings, while the section ‘The Criticism of Appropriate Statistical Norms’ deals with the second.

It is reasonable, of course, to ask for precise definitions of heuristics. One of the hallmarks of the use of heuristics within behavioural economics has been an increasingly formal approach that relies on clear definitions of the underlying phenomena. However, this appears not to be widely recognized outside behavioural economics. Detailed treatments are now available that give formal definitions for most of the heuristics; see Dhami, al-Nowaihi, & Sunstein (2018) and Dhami (2016, Part 7). In this section we give a largely verbal description of how these labels may be given, but a technically equipped reader may benefit from the rigorous treatment given in Dhami, al-Nowaihi, & Sunstein (2018).

An Evaluation of the Criticisms of the KT&O Programme

In this section, we consider the criticisms of the KT&O programme, mainly levelled by the G&O programme. Many of these criticisms are discussed in Kahneman and Tversky (1996) and Gigerenzer (1996). Although the original exchange remains worth reading, the literature has greatly progressed over the years. One of our aims is to offer an up to date evaluation of this debate, with the benefit of recent research and findings.

Some of the arguments against KT&O can be dealt with relatively easily and we simply refer the reader to the discussion in Dhami, al-Nowaihi, & Sunstein (2018). These include inferences from single event probabilities; the charge that Kahneman–Tversky ignore ecological rationality; and that subjects may simply be making errors, could be inattentive, or may be suffering from temporary lapses of judgement. We consider the other criticisms below.

Frequency Versus Probability Format and the KT&O Programme

Several critics of the KT&O approach have argued that the biases created by heuristics in the KT&O programme are eliminated or substantially reduced when data is presented in frequency format rather than probability format (Gigerenzer, 1991, 1996, 2008; Gigerenzer, Hell, & Blank, 1988). Our own reading of the evidence is as follows:

1. An entire class of heuristics in the KT&O programme is unaffected by the distinction between frequency and probability formats. This includes the representativeness heuristic (including the gambler's fallacy and the hot hands fallacy), anchoring, availability, hindsight-bias, regression to the mean, confirmation-bias, and the affect heuristic.

2. The frequency format might reduce biases in the case of some heuristics, particularly in the case of the conjunction fallacy and in applications of Bayes's law. However, significant bias remains in applications of Bayes’ law. After reviewing the relevant evidence, Kahneman and Tversky (1996, p. 585) cautiously noted: ‘Contrary to Gigerenzer's unqualified claim, the replacement of subjective probability judgements by estimates of relative frequency and the introduction of sequential random sampling do not provide a panacea against base-rate neglect.’ The conjunction bias is found to be sufficiently high in a between-subjects treatment (Kahneman & Tversky, 1996). The anchoring heuristic was also found when the information was presented in a frequency format (Tversky & Kahneman, 1973, 1974).

The use of Bayes’ law is critical to the BRA, and it is indispensable in game theory. However, its central role and acceptance in economics is an article of faith; no empirical evidence of conformity with Bayes’ law is taught in economics courses. When attempts have been made to test it in experiments, conformity with Bayes’ law is low, and the typical finding is one of neglect of base rates. Consider the well-known cab problem.

OPEN IN VIEWER

Table 1.

Percentage of responses consistent with Bayes, rule in different empirical studies.

	Information format and judgement domain
Study	Probability	Frequency
Cascells et al., (1978)	18(60)	–
Cosmides & Tooby (1996; Exp. 2)	12(25)	17(25)
Eddy (1982)	5(100)	–
Evans et al., (2000; Expl)	24(42)	35(43)
Gigerenzer (1996b)	10(48)	46(48)
Gigerenzer and Hoffrage (1995)	16(30)	46(30)
Macchi (2000)	6(30)	40(30)
Sloman et al., (2003; Exp1)	20(25)	51(45)
Sloman et al., (2003; Exp1b)	–	31(48)

Source: Barbey and Sloman, 2007.

Example 3. In Tversky and Kahneman (1980), subjects received the following information. There are only two cab companies in the city, Green and Blue; 85 per cent of the cabs are Green. There was an accident last night. A witness comes forward to testify that the cab involved in the accident was Blue. In similar conditions, the reliability of the witness is 80 per cent, that is, the probability that the witness gets it wrong is 20 per cent. What is the probability that the actual cab involved in the accident was Blue? The median and the modal response was 0.8 , which is also the probability with which the testimony of the witness is correct. The statistically correct answer, using Bayes’ rule is 0.414. In particular, individuals appear to underweight base rates, that is, that only 15 per cent of the taxis are actually Blue. For factors that influence the degree to which base rates are taken into account, for example, when data is given a causal or incidental representation, or when people are encouraged to think like statisticians, see Kahneman and Tversky (1996) and Kahneman (2011).

Table 1, which uses information from Barbey and Sloman (2007), shows the percentage of responses that are consistent with Bayes’ rule in a probability format (second column) and a frequency format (third column). There is greater conformity with Bayes’ rule when the information is presented in a frequency format. ¹⁴

14

The Cosmides and Tooby (1996) result is an outlier that Sloman et al. (2003) and Evans, Handley, Perham, Over, & Thompson (2000) could not replicate.

Yet, less than half of the subjects conform to Bayes’ rule in this cross section of studies. While the frequency format reduces base rate neglect, on average about 60 per cent of the subjects across the studies still exhibit base rate neglect. Evans et al. (2000) report that the frequency format has been found, depending on the experiments, to worsen, improve, or leave unchanged the quality of the judgements made.

3. It is possible that humans might have evolved to understand natural frequencies relatively better than percentages (Cosmides & Tooby, 1996; Pinker, 1997). However, most real world economic data is presented in percentage terms (e.g., interest rates on mortgage borrowing). Thus, as a practical matter, one needs to understand human judgement and decision-making when information is presented in a probability format.

4. When conformity with Bayes’ rule under a frequency format is claimed to be truly spectacular, it appears confounded, at least in some cases, by framing effects that favour the outcomes in a frequency format; see the discussion in Dhami, al-Nowaihi, & Sunstein (2018) on the use of Bayes's law in the frequency format in Gigerenzer (2008).

Experts and the KT&O Heuristics

There is now widespread evidence that many experts do exhibit biases (Dhami, 2016, Section 19.18); this contributes towards supporting the ecological rationality of heuristics in the KT&O programme. The evidence comes from multiple domains. When making predictions of political events, experts were only slightly more accurate than chance. However, they were better able to generate explanations for their predictions (Tetlock, 2002, 2006). Experts are typically more overconfident than lay people who lack similar experience and more experienced experts are more overconfident (Glaser, Langer, & Weber, 2007; Heath & Tversky, 1991; Kirchler & Maciejovsky, 2002). The realized returns on stocks are within the 80 per cent confidence intervals of the returns predicted by senior finance professionals in only 36 per cent of the cases (Ben-David, Graham, & Harvey 2013). ¹⁶

Mathematical psychologists exhibit the law of small numbers (Tversky & Kahneman, 1973). The perceived riskiness of various hazardous substances by toxicology experts is consistent with the affect heuristic (Slovic, MacGregor, Malmfors, & Purchase 1999). Clinical psychologists underweight base rates relative to a Bayesian calculation (Meehl & Rosen, 1955). A meta study establishes that decision makers who have the relevant expertise in the field suffer from the hindsight bias (Guilbault, Bryant, Brockway, & Posavac 2004). Professional traders in a large investment bank were found to be hindsight-biased (Biais & Weber, 2009).

Evidence supports the important role of anchoring on a given list price by estate agents (Northcraft & Neale, 1987). Evidence of anchoring is also found in legal judgment (Chapman & Bornstein, 1996; Englich & Mussweiler, 2001; Englich, Mussweiler, & Strack, 2006). Judges exhibit the false consensus effect (Solan, Rosenblatt, & Osherson, 2008). Finance professionals also exhibit a false consensus effect and they impute to others their own risk preferences (Roth & Voskort, 2014).

Experts, such as physicians and World Bank staff, exhibit framing effects, often of a similar magnitude to that observed with student populations (Kahneman & Tversky, 1984; McNeil, Pauker, Sox, & Tversky, 1982; World Bank Development Report, 2015). Wholesale car market dealers exhibit limited attention (Lacetera, Pope, & Sydnor 2012). The World Bank Development Report (2015) documents several kinds of biases among its professional staff. These include confirmation-bias, susceptibility to sunk costs, and the influence of framing. It is candid in its assessment of expert-bias (p. 18): ‘This finding suggests that development professionals may assume that poor individuals may be less autonomous, less responsible, less hopeful, and less knowledgeable than they in fact are.’ The World Bank Development Report also suggests potential solutions to the problem of expert-bias. These include dogfooding (experts signing up and playing their own programmes for real) and red teaming (having an adversarial outside team that tests the proposals).

In sum, the evidence shows that many experts do use heuristics and exhibit biases relative to the BRA and that this is not always eliminated by market experience.

The G&O programme

Herbert Simon (1955) distinguished between substantive rationality (maximizing an objective function under constraints as in the typical optimization problem in economics), and procedural rationality (the process/quality of decision making). Simon's insight was that human beings may lack the information and cognitive abilities to solve problems that are typically posed in the BRA. But people do make decisions. Hence, he was interested in the cognitive processes that give rise to such decisions, i.e., procedural rather than substantive rationality. Simon proposed the satisficing heuristic to operationalize procedural rationality, in which individuals set goals or targets, called aspiration levels. The individual then searches for alternatives that give rise to different payoffs. Once an alternative attains the aspiration level, it is deemed as satisfactory, and further search is terminated. The word ‘satisficing’ is a neologism that alludes to the fact that such decision procedures are satisfactory and they suffice. Empirical evidence is supportive of the theory (Caplin, Dean, & Martin 2011).

The G&O programme focuses on procedural rationality by employing heuristics that are fast and frugal (economize on time and information). The KT&O programme also focusses on fast and frugal heuristics and also uses procedural rationality, although in a different way from the G&O programme. For instance, searching through the available information (availability heuristic); being influenced by prior information (anchoring); being influenced by the law of small numbers in making decisions (representativeness heuristic); and being unduly influenced by one's own prior beliefs (hindsight-bias and confirmation-bias).

In the G&O programme, procedural rationality is introduced by using cues to make decisions; ranking cues according to their ecological validity; deciding on the basis of the first discriminatory cue, or some weighted average of the cue values. ¹⁷

The take-the-best cue (TTB) then proceeds in the following manner.

Search rule: The cues are examined in decreasing order of ecological validity.

Gigerenzer and Goldstein (1996) used data on 83 German cities with 9 ecological cues and find that the TTB heuristic outperforms other methods such as tallying (giving equal values to all cues and adding them up), weighted tallying (weighted values of cues are added), and logistic regression (which computes the probability based on regression estimates that a city is more populous, conditional on the cue values).

OPEN IN VIEWER

Table 2.

Determining which of two foods has more cholestrol

Cues/ecological validity	Time 1	Time 3
Saturated fat (80%)	cake ? pie	cake >pie
Calories (70%)	cake >pie	cake >pie
Protein (60%)	cake >pie	cake >pie
Choice	cake	cake
Confidence	70%	80%

In other contexts, Borges, Goldstein, Ortmann, and Gigerenzer (1999) applied the recognition heuristic to stock market choices, for example, among two candidate stocks, pick the one that is recognized. The authors find that it outperformed several other methods of stock market investment such as mutual funds, market indices, and chance investment (or dartboard portfolios). Czerlinski, Goldstein, and Gigerenzer (1999) also find support for the TTB heuristic against other alternatives such as regression analysis. Subjects had to guess high school dropout rates in 57 Chicago public high schools. The cues, many with high ecological validities, included percentage of low income students, and average SAT scores.

Example 4. (Hoffrage et al., 2000): Patricia decides, at Time 1, which of two foods, a cake and a pie, has more cholesterol. She does not know the answer, so she uses cues to make an inference. The experimenter directly provides her with cues (saturated fat, calories, protein) and their ecological validities, stated as a percentage, just after the cue names; see Table 2. It is critical to note that ecological validity, the percentage of pairwise cases in which the cue can distinguish between the choices in the sample, can only be defined with respect to a reference class of foods. Here, the reference class selected by the experimenters is a random sample of foods from a Chicago supermarket. If a cue favours choice A over B, we write A >B, and if the cue cannot discriminate between the two choices, we write A ? B. At Time 1, saturated fat cannot determine if a cake has more cholesterol relative to a pie (cake ? pie), but calories can discriminate (cake >pie).

The TTB heuristic at Time 1 is used. The cue with the highest ecological validity cannot discriminate (cake ? pie), so the next cue, calories, which successfully discriminates in favour of cake (cake > pie) predicts that a cake is chosen. Patricia's confidence in her judgement equals the ecological validity of the first discriminatory cue, so the confidence is 70 per cent.

Having made the Time 1 choice, Patricia makes the same choice at a future time, Time 3. For whatever reason, she might have imperfect recall of her choice and the choice procedure at Time 1. It is now assumed by the authors that ‘...the cue values are not verdically remembered but show systematic shift towards feedback.’ We are dubious about the suitability of this inference and the reasoning preceding this inference; it appears to assume hindsight-bias, the very phenomenon that the authors wish to explain. The authors use this assumption to construct the information in Column 3 of Table 1. Proceeding as above, cake is chosen with 80 per cent confidence. This, the authors claim, explains hindsight-bias.

Example 4 highlights some of the important features of the G&O approach. The method is fast and frugal (limited number of cues) using a well-specified procedure for choice (procedural rationality). However, the assumptions bear examination. The information on cues and their ecological validity is experimenter-provided. This is unrealistic. Subjects are likely to differ in their beliefs about the reference class of foods that might be relevant for comparison, hence, disagree on ecological validity and also the relevant cues or their number. Thus, if subjects had to search for the relevant reference classes and the cues, there will be a separate prediction for each subject which makes testing of the theory difficult. Furthermore, in an actual search problem, there is likely to be the well known infinite regress problem in search costs (Dhami, 2016, p. 1411). Despite the stated objective of the G&O programme supporting procedural rationality, limited progress has been made in this direction. Some human actions simply involve motor skills acquired during human evolution, for example, the gaze heuristic that is much cited in the G&O programme (Dhami, 2016, Example 19.6). But how are heuristics used to solve cognitive, social, economic, and cultural problems? The G&O programme is evaluated on these issues in the section ‘A Critique of the G&O Programme’.

Furthermore, we only observe Patricia's choices in Example 4, but not the mental process that she engages in. Any number of theories/mental processes can be consistent with her choice. Thus, despite the claim to the contrary, this class of theories are strictly ‘as if’ theories. Finally, using ecological validities to infer confidence in choices needs empirical confirmation.

A range of other heuristics such as LEX and LEXSEMI have been used to examine risky choices (Payne, Bettman, & Johnson, 1993). Unlike the TTB heuristic, the LEX heuristic does not order cues by ecological validity, but by some other measure of importance. The LEXSEMI heuristic imposes a slightly more stringent requirement for cues to discriminate between choices by requiring that the difference between cue values exceeds a certain threshold. The elimination by aspects combines the LEX method with Herbert Simon's satisficing strategy.

Heuristics have also been used to make choices among objects with several attributes (Payne, Bettman, & Johnson, 1993). Edwards and Tversky (1967) suggested the weighted adding strategy (WADD) in which the multiple attributes of objects, subjectively weighted, are added and the object with the highest such score is chosen. But the subjective weights also make WADD difficult to test. One may simply add all attributes (equal weighing strategy, EQW). In both cases, the choice of attributes matters; adding or deselecting unobserved mental attributes of people changes the rankings. Insofar as these attributes are subjective, and privately observed by the subjects, elicitation and predictions of such heuristics are vexed issues on which the G&O approach throws very little light.

A Critique of the G&O programme

In this section we critically evaluate the G&O approach.

Evaluation of the ‘Less Is More’ Effect

Gigerenzer (2008) writes: ‘That simple heuristics can be more accurate than complex procedures is one of the major discoveries of the last decades.’ This is also often described as the less is more effect. The source of this finding is that in prediction tasks, there is a tradeoff between unbiasedness and variance minimization, a well-known result in statistics (Gigerenzer, 2008; Gigerenzer & Brighton, 2011).

For a formal statement and derivation of the less is more effect result, see Dhami (2019, Chapter 4). In the work of G&O, the less is more effect is often stated informally as follows

where the noise term is a mean zero and constant variance noise term.

Consider the following simple example that illustrates the less is more effect. Suppose that we have a set of data points, and we wish to fit a polynomial through these points. The estimated polynomial is the relevant estimator of some true underlying function. The higher is the degree of the fitted polynomial, the better is the in-sample fit, that is, more unbiased is the estimator within the sample. However, the sample dispersion is partly determined by random noise. Fitting the in-sample noise-inclusive data too closely by a higher order polynomial, reduces the ability to fit out-of-sample data in a prediction task; such polynomials will exhibit higher variance in fitting data out-of-sample. In other words, while higher degree polynomials will have lower bias, they might have significantly higher variance; on net, the RHS of (9.1) may turn out to be higher for higher degree polynomials, beyond a certain degree. In this case, a lower degree polynomial that has higher bias in-sample but lower variance out-of-sample could make lower prediction errors; this is the essence of the less is more effect. The less is more effect lies at the heart of the less is more critique of G&O. The bias-variance tradeoff is statistically sound, but the relevant issue is whether it is useful to study human cognition. Gigerenzer and Brighton (2011, p. 12) write: ‘Our cognitive systems are confronted with the bias-variance dilemma whenever they attempt to make inferences about the world.’ However, we are not aware of any direct incontrovertible evidence in support of this assertion. ²²

On Mathematical Optimization and ‘As If’ Theories

G&O have criticized economic theories, particularly prospect theory, on the grounds that theories based on optimization are ‘as if’ theories that ignore procedural rationality (Gigerenzer, 2008). Gigerenzer (2016, p. 38) write:

Although behavioural economists started out with the promise of greater psychological realism, most have surrendered to the as–if methodology. Cumulative prospect theory, inequity-aversion theory, hyperbolic discounting are all as–if theories. They retain the expected utility framework and merely add free parameters with psychological labels... which is like adding more Ptolemaic epicycles in astronomy. The resulting theories are more unrealistic than the expected utility theories they are intended to improve on. Behavioral economics has largely become a repair programme for expected utility maximization.

Optimization is simply a tool used both by neoclassical economics and behavioural economics. But behavioural economics and its approach, methodology, results, and conclusions are very different from the canonical approach taken in neoclassical economics. ²³

It is entirely appropriate to criticize as–if optimization based models if their predictions are inconsistent with the evidence. Behavioural economics has shown, conclusively, we believe, that theories that are central to optimization-based models in economics, such as expected utility, exponential discounted utility, self-regarding preferences and several refinements of Nash equilibrium, are inconsistent with the evidence (Dhami, 2016).

Behavioural economics has never been a rescue project for expected utility; it has generated conclusive evidence to show that expected utility is untenable. Prospect theory is an ‘as if’ theory but overall, one that does best in situations of risk/uncertainty/ambiguity, as understood in economics (Dhami, 2016, Part 1). It is a leading example of a tenable ‘as if’ theory in behavioural economics. There are models of heuristics, such as the priority heuristic, that can explain some of the puzzles that prospect theory explains (Brandstätter, Gigerenzer, & Hertwig 2006). However, its predictions do not hold outside the dataset used in the paper and questions have been raised about the statistical techniques used (Birnbaum, 2008). The range of phenomena explained by prospect theory is too vast to be accounted by any other theory, heuristics or otherwise. ²⁴

Recent theoretical work also suggests that there is a link between heuristics based approaches in time discounting and optimization models (Dhami, 2016, Section 10.4, p. 627). The relevant heuristics in this case are non-compensatory and lexicographic. ²⁵

G&O have also expressed strong reservation to optimization based theories on the grounds that they have free parameters (see quote above). But there are statistical tests that enable us to determine whether these parameters add enough in the way of explanatory power that justifies their use or not (e.g., Akaike information criteria and Bayesian information criteria). G&O require that the parameters so estimated in these models should be held fixed in every context and frame in order to make predictions. ²⁶

On Distinct Domains of Choices in KT&O and G&O

As noted above, economics makes no predictions under true uncertainty, but individuals do make choices under true uncertainty. ²⁷

1. Overconfidence: Even under true uncertainty, overconfident individuals may believe, incorrectly, that they can see through all the possibilities and assign subjective probabilities to all these possibilities. One can then use any of the standard decision making models such as prospect theory (where special cases include expected utility and rank dependent utility), or some heuristics based model based on the insights of, say, the priority heuristic.

2. Conservation of cognitive effort: Under true uncertainty, individuals may not believe that they can see through all the possible outcomes and probabilities. However, exerting cognitive and search costs to explore possible unknowns and their probabilities, may be too costly. Individuals may then simply assign a vanishingly low probability to the ‘unknown unknowns’ and focus on the ‘known unknowns’ only. In this case, they can employ any of the standard decision making models. For an analogue of such models under risk, where people simply ignore events of very low probabilities and optimize over the rest, see Dhami (2016, Section 2.11).

3. Norms: A tantalizing possibility is that individual behaviour under true uncertainty might be determined by social norms or personal norms. Consider, for instance, marriage, arguably a problem in the domain of true uncertainty. One may decide to have a trial live-in relationship, and make the marriage decision based on an evaluation of the live-in period. This is a norm in some societies but not others.

4. Heuristics: Individuals may realize that they just do not have enough data to make any optimization based decisions at all, so they may resort to simple heuristics to make the choice.

There is no presumption that our four speculative categories are non-intersecting. For instance, heuristics may themselves be determined by social norms, which might in turn have been optimized to conserve cognitive effort in order, partly, to prevent overconfident individuals from making choices that are suboptimal. When we speak of true uncertainty, below, we only consider the very last potential explanation of behaviour, that is, a heuristic-driven account of true uncertainty, as if it were a stand-alone explanation.

Our main claim in this section is that the domain of decisions in the KT&O and the G&O accounts of heuristics is, in many cases, non-intersecting. However, in some cases, the domains intersect, as in the use of the priority heuristic to deal with risk, and the efficacy of the frequency format relative to the probability format in reducing biases relative to the BRA. We have already dealt with these issues elsewhere in this essay. For this reason, here, we focus exclusively on the difference in domains between the two approaches.

The KT&O demonstration of heuristics and biases mainly considers the domains of risk and uncertainty alone. In several of their most important demonstration of biases, they explicitly provide data on outcomes and probabilities to subjects. In other cases, data on probabilities and outcomes is not needed. This includes their demonstration of the law of small numbers (which includes the gambler's fallacy and hot hands fallacy); base rate neglect and violation of Bayes’ law; conservatism or underweighting of the likelihood of a sample; hindsight-bias; confirmation-bias; regression to the mean; false-consensus effect; conjunction fallacy; and confusion between necessary and sufficient conditions.

In other cases used in KT&O, the distinction between risk/uncertainty and true uncertainty is less clear; this includes the affect heuristic and the anchoring heuristic. But even here, empirical evidence shows that these heuristics continue to hold when the environment is unquestionably in the domain of risk and uncertainty, even when the subjects are experts in the area (Dhami, al-Nowaihi, & Sunstein, 2018).

One can, therefore, safely conclude that the relevant domain that KT&O used to demonstrate their biases was either certainty, risk,or subjective uncertainty but not true uncertainty. We argue below that the domain of problems that G&O deal with typically lies within the domain of true uncertainty, which they describe as large worlds. This observation on the G&O programme is important not only because we really need to understand human decision making in such environments, but also to note that many of the contributions of G&O and KT&O lie in non-intersecting domains. In light of this, the characterization of their positions as adversarial, as has been the case in the great rationality debate, can be misguided and misleading.

The research agenda in G&O speaks directly to problems of true uncertainty. They distinguish between small worlds and large worlds, drawing on terminology originally introduced by Savage (1954). Small worlds corresponds to our use of the terms ‘risk’ and ‘subjective uncertainty’ (that KT&O were mainly interested in) and large worlds corresponds to our use of the term ‘true uncertainty.’ In introducing their research agenda, and summarizing their approach Gigerenzer et al. (2011, p. xviii) write:

How should we make decisions in the large world–that is, when Bayesian theory or similar optimization methods are out of reach?...In sum, the accuracy-effort trade-off indeed exists in a small world. In a large world, however, where there is no general accuracy-effort tradeoff, heuristics can be more accurate than methods that use more information and computation, including optimization methods.

In this quote the greater accuracy of heuristics relative to other methods that rely on more computation/information is predicated on the less is more effect that we have already analyzed above.

A good example of this approach for economists is in financial market investment. It is probably impossible to predict or perhaps even know the set of all possible outcomes and probabilities in stock markets, so it is, arguably, a problem in true uncertainty. The standard assumption in financial economics is to treat this as a problem in risk and uncertainty, yet even in this case there may be no unique prediction (see Example 6). Candidate optimization benchmarks that are typically used in the G&O programme, to evaluate their heuristics against, such as logistic regression, require information that is simply not available under true uncertainty. But people do invest in the stock market. We draw on two examples that are frequently cited in the G&O programme as examples of successful heuristics (Gigerenzer et al., 2011).

Example 5. Borges et al. (1999) study the performance of the recognition heuristic in choosing among stocks (i.e., among two stocks, choose the one that can be recognized) against the performance of other mutual funds. In six out of eight tests, the performance of the recognition heuristic was better than the performance of selected mutual funds. Economists would need more data and tests to be persuaded by the results; the data was over a period of 1 year in the midst of a strong bull market that might have favoured stocks that are more easily recognized. More tests would need to be done to disentangle the predictions of the recognition heuristic in this case from other predictions of models in finance.

Example 6. Consider the following simple investment strategy. In a portfolio of N assets, invest a share 1 / N in each asset. This is sometimes known as the 1 / N heuristic. Treating the problem of financial investment as one of risk and uncertainty (but not true uncertainty), DeMiguel, Garlappi, & Uppal (2009) compare the performance of the 1 N heuristic with 14 other optimal investment theories often recommended in finance. ²⁸

We offer three main comments on this work. (a) No evidence is given that people (as distinct from investment firms) do indeed use any of the optimal investment strategies described in the paper. (b) The 1 / N heuristic does not belong to the class of TTB heuristics in the G&O approach (e.g., no cues are used at all). (c) The problem of investing in financial markets is arguably a problem in true uncertainty. For these reasons, DeMiguel et al. (2009) address an entirely different class of issues: Namely, that the 1 / N heuristic performs better than the optimization benchmarks that were developed under risk and subjective uncertainty.

As our final point in this section, it is a tantalizing possibility that people may employ the heuristics in the KT&O programme when dealing with true uncertainty, although these heuristics were not designed for this purpose. Suppose one faces a problem in true uncertainty. Then one may actually be influenced by an externally provided anchor or suggestion (anchoring heuristic); or by the availability of apparently similar information in the past (availability heuristic); or by the emotional tags association with the given information (affect heuristic). It might also be the case that Herbert Simon's aspiration adaptation theory may be used in such cases. We believe that it is worth exploring this possibility in future research. This would provide a different account of decision making under true uncertainty, as compared to the G&O programme.

Nudges and Nudging

In neoclassical economics, observed choices reflect the underlying preferences of individuals. In this standard view, people's choices are assumed to promote their welfare. The reason is that such choices represent sufficiently careful and considered judgements about the welfare effects of actions; the preferences underlying these choices are sometimes termed as normative preferences. At the same time, individuals in neoclassical economics are assumed to be self-regarding, which gives rise to a role for public policy to internalize the externalities caused by self-regarding actions. Examples of relevant policy interventions include Pigouvian taxes to regulate the emission of pollution by firms, subsidies to consumers to buy electric cars, and tax-exempt charitable contributions. Regulation (such as fuel economy standards) might also be used to internalize externalities, though it is a crude response.

Do individuals really know what is in their best (including the long-term) interests all the time? Are they able to make choices that are in their long term interests? Consider the following empirically documented examples, F1–F6, of human behaviour that by no means exhaust the scope of our argument.

F1. Individuals may overconsume some goods (sugar, saturated fat, alcohol, tobacco) and suffer ill health, a shorter lifespan, and a lower quality of life.

F2. Individuals may procrastinate too much, fail to complete their projects in time, fail to consolidate their finances in a timely manner that might be more conducive to their long term interests, not enroll in suitable pension plans, and take up annual gym memberships when a pay-as-you go membership could have saved them money.

F3. Individuals might under-save for their retirement; the data indicates a sharp drop in consumption at retirement (Bernheim, Skinner., Skinner, & Weinberg, 1997). Individuals may also retire too soon.

F4. Individuals might make marriage and divorce decisions too hastily. They might also make purchase decisions on impulse, in a hot state, and regret it afterwards.

F5. Individuals might be misled by manipulative or deceptive, but not untruthful, advertising by firms that cleverly frames the options (Akerlof & Shiller, 2015).

F6. Individuals might not purchase fuel-efficient vehicles or energy-efficient appliances. Focussing on the short-term, they save money in the present but lose money over a period of years, and with any reasonable discount rate, they are likely to be making an unjustifiable choice.

Under normative preferences, we should not observe the phenomena in F1-F6. For example, in the BRA framework individuals rationally choose to consume excessively unhealthy foods; never procrastinate; perfectly smooth consumption over their lifetime; choose to marry, divorce, and make their purchases rationally; and make frame-invariant choices, provided the frames are informationally equivalent.

We do not deny that in cases that appear to show mistakes (as in F1–F6), it is possible that some people, or most people, are doing exactly what they should do. We agree that third parties, including public officials, should hesitate before concluding that people have erred. But in many cases, the decisions of some or many people are not in their interests. In actual practice, public policy and self-imposed control mechanisms might be used to remedy the problems described in F1–F6, above. For example, such remedies may take the form of: (a) Corrective taxes to respond to ‘internalities’ (public policy), individual use of private organizations (such as Alcoholic's Anonymous), or voluntary decisions to check into rehab at a substantial financial cost (self-control mechanisms). (b) Internally or externally imposed deadlines to undertake the appropriate action such as to fill one's tax forms, and submit an assignment by a due date. (c) SMarT (save more tomorrow) savings plan that commits individuals to save a fraction of future increases in their incomes (Benartzi & Thaler, 2004). (d) Cooling-off periods for divorce and a 21 day return policy for impulse purchases. (e) Regulation requiring financial firms to state the annual percentage interest rates (APR) on their financial products, and clear, colour-coded declaration of the key nutritional facts of food products, such as calories, saturated fats and salt content. (f) Clear labels, so that people can see, and appreciate, the potential economic and social benefits of fuel-efficient automobiles and energy-efficient appliances.

F1–F6 are easily explained within a behavioural economics framework. These and several other problems can arise from a range of judgement heuristics and biases (Dhami, 2016; Thaler & Sunstein, 2009). Many individuals report that they would prefer to save more, or prefer the median portfolio to their own, and others pay to kick a bad habit, indicating that they recognize their choices to be suboptimal (Sunstein & Thaler, 2003; p. 1169). Many individuals exhibit self-control problems as a result of present-biased preferences; this typically leads to temporally suboptimal outcomes such as inadequate savings for retirement, obesity, procrastination, and drug use (Dhami, 2016, Part 3). US data shows that many individuals do not utilize the full limits of their 401(k) pension plans. People also diversify their portfolios inadequately, and often use the 1 N heuristic in forming portfolio. This skews their pension portfolios (e.g., towards their own company equity). Coupled with overconfidence in their own companies, they might hold too much company equity, with possibly adverse consequences (e.g., Enron employees). Despite the steep interest rates on payday loans, these are used excessively.

Individuals may suffer from limited attention (the BRA assumes unlimited attention); thus, they react more to salient taxes, and in the used car market there are discontinuous drops in the prices of used cars at 10K thresholds (Dhami, 2016, Section 19.17). Emotions play an important role in individual economic decisions, such as buying consumer durables. Sometimes these decisions are taken in a hot state and regretted in a cold state (Dhami, 2016, Part 6); admittedly, the normative issues are especially complicated in some of these cases. Individuals might also be heavily influenced by the framing of the same problem in informationally equivalent ways. Indeed, framing plays an important role in the advertising industry, which spends more than 500 billion per year. The frame and context dependence of preferences has been actively studied in behavioural economics (Dhami, 2016; Kahneman & Tversky, 2000).

In light of these departures from the assumptions of neoclassical economics, what welfare or normative significance should one attach to individual choices, and how must we modify classical welfare analysis? ²⁹

Issues of paternalism in behavioural economics are often captured under the umbrella term of soft paternalism. This term includes libertarian paternalism (Sunstein & Thaler, 2003; Thaler & Sunstein, 2003, 2009), asymmetric paternalism (Camerer, Issacharoff, Loewenstein, O'Donoghue, & Rabin 2003), and light paternalism (Loewenstein & Haisley, 2008). Such approaches tend to share two important features: (a) they allow people to go their own way, that is, to reject the direction suggested by the paternalist; and (b) they reflect a form of means paternalism, rather than ends paternalism, in the sense that they are respectful of the chooser's view of the ultimate destination. It is for this reason that a GPS device is a defining example of libertarian paternalism.

In the relevant cases, the policy is introduced by a planner or a choice architect (who may work in the private or public sector). The objective is to enlist policies that do not distort (or distort minimally) the choices of fully rational individuals, but at the same time nudge others in a direction that is in their considered best interests (as judged by themselves). By considered best interests is meant the decisions that would be made by individuals if they had complete information, unlimited cognitive abilities, and no lack of self-control. In other words, individuals are enabled to make choices that make them better off, again as judged by themselves (Thaler & Sunstein, 2009, Chapter 5).

Admittedly, it can be challenging to know whether this criterion is met (Sunstein, 2017). If so, it may be best to insist on a particular form of choice architecture: active choosing (Sunstein, 2017). But in some contexts, it is difficult or perhaps impossible to avoid some kind of steering of individual choices. Any website has to have some kind of layout. Any form has to have some questions first and some questions last. Individuals are given choices and information by firms, other individuals, and the government. These are necessarily framed in a particular manner and already embed a certain choice architecture. Many examples can be given (Thaler & Sunstein, 2009) from doctors who advise patients on the choice of a treatment to architects who advise clients about house designs, to interior decorators who advise people on alternative interior designs.

Example 7. (Sunstein & Thaler, 2003): Consider the arrangement of food items in a cafeteria in a certain order. On plausible assumptions, boundedly rational consumers, who pay special attention to items placed at eye level, could benefit from the placement of healthy options at eye level, relative to unhealthy snacks; this is an example of a nudge. Consumers with normative preferences are not influenced by the arrangement of items, so little cost is imposed on them. The intervention of the choice architect is a form of paternalism. But ultimately it is the consumer who exercises choice; in this sense, the proposal respects consumer sovereignty.

Example 8. (Dhami, 2016, p. 1593): Here are some everyday examples of nudges. Cash machines dispense cash only when the relevant bank card has been taken out first; nozzles for dispensing different kinds of car fuel, for example diesel and petrol, are often of a different size or colour to prevent incorrect use of fuel; many cars produce a beeping sound if the driver is not wearing a seat belt; automatic electric switches in cars and offices are often used to conserve energy, or prevent batteries from running out; default options in savings and pension plans enable people to invest better; and government warnings on cigarette packs.

Of great interest to behavioural economics are cases, traced above, that arise from self-control problems. These problems are often a product of present-biased preferences and imperfect awareness of such preferences in conjunction with a model of multiple selves; this is shown rigorously in Dhami (2016, Part 3). Since this appears to be such a common and repeated source of misunderstanding of the rationale for nudges, we summarize a basic result; for the full details, see Dhami (2016, Parts 3, 8).

Remark 1. (Dhami, 2016): People who have present-biased preferences will exhibit the behaviour outlined in F1–F4 even if they satisfy all other features of the BRA. The degree to which they exhibit such behaviour, and whether the behaviour is to procrastinate or preproperate, will depend on their degree of self-awareness of their own future self-control problems.

Critics have objected to several features of libertarian paternalism. Sugden (2008, 2013) has used the contractarian approach in which appeals are directly made to individuals, stressing the personal advantage to each individual from taking a particular action. The interested reader can get acquainted with some of the features of alternative proposals in Dhami (2016, Chapter 22)

Conclusions

The heuristics and biases approach, begun by Daniel Kahneman and Amos Tversky, is one of the most important research programmes in the social sciences and certainly one of its stellar achievements; we call it the KT&O programme. It showed, incontrovertibly, that the empirical evidence is not consistent with the BRA in economics. There has been increased formalization of the KT&O programme. It has been intensely scrutinized and criticized, which we believe has benefitted it. We have given several examples of a formalized version of the various heuristics in the KT&O programme, evaluated the criticisms of the approach, and described how these criticisms can be answered, and existing understanding improved by a fuller consideration of the evidence and the emerging models.

The BRA approach deals with situations of certainty, risk, subjective uncertainty, and ambiguity. In contrast to these situations, many interesting and important real life problems belong to the domain of true uncertainty. Furthermore, many real world problems are NP complete, that is, the time required to solve them grows too fast as the problem becomes more complicated. In fact, problems like this, such as the travelling salesman problem, cannot be solved in polynomial time. How do people solve such problems?

Decision making under true uncertainty is a vexed problem. Despite progress, we believe that the G&O programme has not yet been able to provide a persuasive account of decision making under true uncertainty. It is not clear what benchmark to use to assess the performance of the heuristics in the G&O programme; in contrast the benchmark is clear in the KT&O programme. There remain severe problems in stringently testing the heuristics proposed in the G&O programme because they rely on unobserved mental processes; we give several examples.

Finally, we consider an application to an approach in behavioural welfare economics, libertarian paternalism or nudging. This has recently been criticized in the G&O programme. We show that libertarian paternalism does not depend on controversial psychological claims and the criticisms do not always take account of the role of self-control problems and imperfect awareness of such problems.

We offer one final note. The neglect of true uncertainty in economics is truly astonishing. The other social sciences do not provide a coherent framework in this regard either. More research is needed on how people make decisions in this case. We also need to have a basic agreement on what constitutes a benchmark in these situations against which proposed alternatives can be evaluated. Should, for instance, this benchmark take an ex-ante perspective (prior to the resolution of true uncertainty) or an ex-post perspective (after the resolution of true uncertainty)? Heuristics in the KT&O programme that were designed for evaluating the BRA may, in our view, be also potentially interesting candidates to think about decision making under true uncertainty. Or it could be that mental models and social norms might have evolved to deal with such cases. It is staggering that the black box of true uncertainty remains so impenetrably black.