3.1. Financial behaviour

Policymakers and experts in the financial domain have been concerned that many people lack sufficient financial literacy to successfully plan lifetime consumption or run a small business. Different behavioural interventions have been proposed with the goal of improving financial behaviour.

Save More Tomorrow (SMT) is a behavioural intervention designed to increase people's retirement savings. Its design is based on viewing decisions about retirement savings as trade-offs between current and later (retirement-age) consumption and aims to influence these decisions by changing one of the options in the trade-off (Thaler and Benartzi Reference Thaler and Benartzi2004; Benartzi and Thaler Reference Benartzi and Thaler2013). That is, instead of asking people to choose a trade-off between consumption now or later, the policymakers ask people to choose a trade-off between consumption in the near future (say, a year from now) and later. The rationale behind this intervention is that people tend to systematically overvalue the immediate present over the future, thus reducing willingness to save for retirement now. The theory of hyperbolic discounting (Loewenstein and Prelec Reference Loewenstein and Prelec1992) captures this finding but also suggests that modifying the timing of options may be used to reduce this effect. The SMT program also takes into account people's inertia (increases in savings are automatic) to ease them into self-control restrictions (by projecting them into the future). Various observational studies show that this intervention is effective. Among people who said that they could not afford a cut in pay now, 78% joined the programme (see Thaler and Sunstein Reference Thaler and Sunstein2008).

An alternative policy targeting financial behaviour trains people in the use of simple rules of thumb (SRT) for making financial decisions (Drexler et al. Reference Drexler, Fischer and Schoar2014). For example, people were trained to keep their money in two separate drawers (or purses) for their business and household, respectively, and to transfer money from one drawer to the other only when accompanied by an explicit ‘I Owe You’ note. This simple rule of thumb for structuring revenues had significant and substantial effects on business practices, reporting errors and revenues, in particular in low-skilled individuals (Drexler et al. Reference Drexler, Fischer and Schoar2014). Notably, training a control group in standard accounting practices did not have comparable effects.

Both SMT and SRT have been shown to be effective interventions to facilitate far-sighted financial behaviour. And although they pursue closely related goals, they do so through different pathways. SMT intervenes on a feature of the environment – the temporal location of the saving choice. By shifting it into the future, SMT seeks to harness the relatively low discounting rates between two events that are located at a sufficient distance in the future, as implied by hyperbolic discounting. SMT expects that the change, under the hyperbolically discounting heuristic, will cause an increase in saving choice. It thus uses knowledge about a feature of people's heuristic repertoire to change the choice environment such that people are likely to adjust their behaviour accordingly. According to our distinction, SMT is a nudge.

SRT, in contrast, is a boost because it intervenes in people's heuristic repertoire by teaching them a new rule of thumb. The rule helps choosers to interpret the environment in new ways (according to the household/business account distinction), and will only be applied if matching information cues (i.e. separate sources of income and expenditure that match this distinction) can be found in the environment. To the extent that the training is successful, it expands the agents’ heuristic repertoire so that they exhibit better business discipline and are more likely to succeed in their businesses.

3.2. Gym attendance

Policymakers and experts in the health domain have been concerned that many people do too little exercise and therefore are more likely to suffer from cardiovascular and other diseases. Again, different behavioural interventions have been proposed with the goal of increasing gym attendance.

Providing gym membership by default is such a policy. Typically, gym memberships are ‘opt-in’, since an employee might get a subsidy for membership but must choose to become a member to use this subsidy. The default-setting (DS) policy changes this to an ‘opt-out’: you are automatically enrolled, but if you explicitly choose not to be a member, you might use the subsidy for other services. The effectiveness of such policies has been shown in domains like saving for retirement (Beshears et al. Reference Beshears, Choi, Laibson, Madrian, Brown, Liebman and Wise2009) but it has also been applied with the goal of increasing the exercise regime. For example, as part of a broad regime of health-improving measures, the Cleveland Clinic has implemented such an opt-out system and reports substantial weight loss and reduction in blood pressure amongst its employees (Klein Reference Klein2011).

Alternatively, Temptation Bundling (TB) teaches people a strategy to overcome self-control problems by coupling ‘instantly gratifying “want” activities (e.g., watching the next episode of a television show, checking Facebook, eating an indulgent meal) with engagement in a “should” behaviour that provides long-term benefits but requires the exertion of willpower (e.g., working out, completing a paper review)’ (Milkman et al. Reference Milkman, Minson and Volpp2013: 1–2). Specifically, the experimenters recommended that the subjects allow themselves to enjoy a number of desirable audio novels only while exercising, thus making their gym attendance more tempting. In comparison to a control group that was just given a bookstore voucher of the same value as the audiobooks (0.94 gym visits/week), the subjects using TB showed a higher weekly gym attendance rate (1.42 visits/week).

Both TB and DS aim at increasing gym attendance rates. As in the previous case, the two interventions do so through distinct pathways. TB aims to enrich the agents’ heuristic repertoire by training them in the use of a new heuristic. This heuristic helps them interpret their environment in a novel way (‘Am I facing a self-control problem with respect to the gym or not?’). If the chooser detects a positive cue in this dimension, she reacts by applying the bundling rule. This in turn affects her behaviour, making her visit the gym more often. Thus, according to our model, TB is a boost.

DS in contrast intervenes in the environment by changing the default setting of a choice situation. Many experiments have shown that people tend to choose the default option more often, be this for reasons of inertia, loss aversion or recommendation effects (Thaler and Sunstein Reference Thaler and Sunstein2008; Beshears et al. Reference Beshears, Choi, Laibson, Madrian, Brown, Liebman and Wise2009). DS uses knowledge about these tendencies to change the environment in such a way that people are likely to choose the desired option. According to our model, DS is a nudge.

4.1. Trigger stability

The Heuristics & Biases programme has documented many persistent cognitive biases, mainly through laboratory experiments (Tversky and Kahneman Reference Tversky and Kahneman1974; Kahneman et al. Reference Kahneman, Slovic and Tversky1982; Haselton et al. Reference Haselton, Nettle, Andrews and Buss2005). However, even if the existence of a bias is established, it is an additional question whether it will be stable when a nudge is implemented.

Let's consider the Save More Tomorrow (SMT) intervention discussed in Section 3.1. This nudge works if people show stable preferences related to hyperbolic discounting, that is, overvaluation of benefits obtained in the present over similar or even higher benefits available in the future. However, if the agent were aware that what matters for the decision is her view in a year's time she might adjust her evaluation and apply a different heuristic. In such cases, changing the timing of the options will not result in the predicted behavioural reaction. Similarly, if the agent interpreted the intervention as an attempt to manipulate her, she might seek to avoid the manipulation by changing her reactions (for evidence of these kinds of reactance, cf. Brehm and Brehm Reference Brehm and Brehm1981; Campbell Reference Campbell2007). Similar complications might arise with respect to the example of Default Setting (DS). For example, ‘in contexts that involve less effort, and clearer antecedent preferences, default rules are less likely to stick, and disclosures might make them less sticky still’ (Sunstein Reference Sunstein2016: 118). Also, if existing defaults are greater attractors than other options, because they are more salient, carry a signal better, or because they are less offensive, reactance to changes in default might result (Grüne-Yanoff Reference Grüne-Yanoff2016). In other words, when switching defaults as a policy intervention, one is not guaranteed that the same number of people will stick with the new default as they did with the old.

The goal of a nudge is to change the environment in such a way that the application of a given heuristic, which is assumed to be stable across environments, leads to a defined behavioural outcome. Hence, what needs to be established is that the triggering relationship between the environment and a heuristic is stable. If the relationship is unstable, policymakers cannot hope to reliably change behaviour by altering the environment. In other words, nudge interventions require trigger stability between planned changes in the environment and known heuristics. By contrast, since in our model of boosts heuristics are selected by the agent rather than triggered, the stability between the environmental trigger and the behaviour is not required for a boost to work. As we will see, boosts require a different sort of stability, namely sufficient (environmental) resources so that the agent is able to select the right heuristic from the heuristic repertoire.

4.2. Agent motivation

By definition, boosts aim at changing the agent's heuristic repertoire and therefore require the agent's active participation at the intervention stage. If the agent is either not able or not motivated to participate, it is unlikely (though not impossible) that a boost changes behaviour (see also Grüne-Yanoff and Hertwig Reference Grüne-Yanoff and Hertwig2016). Because ability is something to which a boost can in principle be calibrated but motivation is not, we examine each condition separately. We focus on motivation in this subsection and turn to teachability in the next.

Even the most inquisitive student will be bored or distracted by disorganized teaching. Thus, motivation clearly has a component that can be addressed by good pedagogical design. Motivation to learn and practice is, of course, a large topic that cannot be comprehensively discussed here. However, we indicate some conditions that might affect an agent's motivation. First, an agent's motivation depends on having a goal that she wants to achieve such as better health, financial stability or improved self-control. If she does not conceive of the intervention as a relevant means for her goal, a boost is less likely to be effective because she will see no reason to learn or undergo training in the first place. Second, even if an agent sees an intervention as a means to achieve her goals, she might not see the need to learn new heuristics because she believes in her current strategy and does not see the added benefit of an alternative one. Third, even if people are motivated to adjust their current strategy, they may not be motivated to learn unless they are convinced that the boost provides them with effective means to achieve their goals (e.g. thousands of purported ‘dietary rules’ are peddled – often by agents with less than sincere motives – and the success of most of these rules is very uncertain at best). In other words, if the level of expected effort appears higher than the expected benefit from learning, people will not be motivated to learn.

Thus, in case agents do not see the added benefit of learning new strategies to achieve their goals, boosts tend to be less effective than nudges. Of course, motivation to learn may be increased by appropriate information about the pros, cons and effectiveness of the boost before training. Sometimes, lack of motivation may even be addressed with specific boosts (or nudges). Nevertheless, even the best training interventions and preparatory information might fail to motivate people, for example, because they suffer from inertia or keep procrastinating. In sum, in contexts where motivation is lacking, nudges tend to be more effective than boosts because a change in the environment does not require the agents’ active participation.

4.3. Teachability of heuristics

Motivated agents might fail to learn a heuristic because of pedagogical failures. Such failures might lie either in the design of the heuristic itself or in the way it is taught. For example, most people will be challenged to learn how to extract power roots without paper and pencil. Tools that require more memory and processing capacity than a person can muster are examples of such design failures. Less extreme examples include the difficulty of teaching low-skilled individuals standard accounting practices. As Drexler et al. (Reference Drexler, Fischer and Schoar2014) showed, training small-business owners in standard accounting practices had no effect in improving reporting practices and business success, while simple rules of thumb training had a significant effect.

The Fast & Frugal Heuristics programme has provided evidence that simple, easy-to-grasp rules can be used to make sufficiently accurate judgements compared to more complicated procedures (Gigerenzer et al. Reference Gigerenzer and Todd1999). The benefits of such decision-making strategies have been observed in a variety of settings, ranging from food consumption, to the stock market to online dating (van der Linden Reference van der Linden2011). Thus, boosts are most effective in domains where simple and effective rules of thumbs can be identified or simple ways exist to teach a more complicated strategy. Hence, when teachability is low, nudges tend to be the better choice.

The second issue concerns good pedagogy. Here again, the relevant literature is extensive. We want to mention only one contextual factor that strikes us as important. As we have defined them, boosts aim to teach the right application of simple heuristics, which must be trained in repeated applications in order to be learned. An important feature of training success is sufficient and prompt feedback. In most training situations (such as schools, universities or PhD supervision), feedback is provided by the teacher. Since boosts are intended for large populations, not for individual training sessions, the feedback should be provided automatically. In fact, many successful boosts include such feedbacks. For example, people trying out Temptation Bundling (TD) might see their exercise time go up, or at least find it easier to regularly visit the gym. Similarly, people who have taken a first lesson in simple rules of thumb for running small businesses might see that applying the rules positively affects their business behaviour.

We do not believe that feedback is a necessary condition for the effectiveness of boosts in the strict sense, however. In the short run, Simple Rules of Thumb (SRT) may be effective even without the user being aware of its function. But without positive feedback, people have no reason to continue using a heuristic in the longer run. For example, it is plausible to assume that people would eventually discontinue using separate drawers for family and business accounts, unless they came to understand that this helps them achieve better financial discipline and ultimately better business outcomes. Also, feedback is not a sufficient condition for learning. For example, outcome feedback does not make people's tendency to underestimate rare events go away (Hertwig et al. Reference Hertwig, Barron, Weber and Erev2004). However, there is evidence that feedback helps people learn to use their heuristics adaptively (Rieskamp and Otto Reference Rieskamp and Otto2006). In general, feedback will help to create confidence in the effectiveness of a boost, which in turn will increase the likelihood that people persist and maintain motivation to learn and apply it.

The previous three conditions concern requirements for effective boosts and nudges at the intervention stage. When the corresponding conditions are satisfied, nudges will trigger heuristics systematically and reliably and boosts will successfully enrich people's heuristic repertoire. However, this does not ensure that either intervention will change behaviour in the intended direction. For this to happen the heuristic triggered or the newly learned heuristic must reliably affect the behaviour of at least the majority of the targeted population in a particular direction. That is, there should be some positive net effect: for example, that more people show the behaviour than before, or – when there is a quantifiable behavioural outcome, like a retirement contribution – that the average outcome has changed in the desired direction. This leads to the discussion of the implementation stage of Figure 1, where heuristics or behavioural tendencies produce actual behaviour.

4.4. Homogeneity of heuristic repertoire across individuals

We start by analysing the conditions for the effectiveness of nudges at the implementation stage. Nudges typically affect the choice environment for a whole population based on the so-called equal incompetence assumption, which ‘treats all . . . actors in all situations as if they were equally predisposed to commit errors of judgment and choice’ (Mitchell Reference Mitchell2002: 2).

Thus, the behavioural triggers that nudges exploit do not, and typically cannot, discriminate between individuals who may have different heuristics at their disposal or simply differ in their levels of competency. Let us illustrate the problem of heterogeneity with an example from the Default Setting (DS) intervention. Beshears et al. (Reference Beshears, Choi, Laibson, Madrian, Brown, Liebman and Wise2009) describe the effects of resetting the default of the retirement saving rate. Employees hired before a set date were offered a choice with a default rate of 3%, while employees hired after that were offered the same choice with a default rate of 6%. The intended effect was an increase of those who chose the 6% contribution (from 24% to 49%) and a decrease of those selecting the 3% contribution (from 28% to 4%). It is tempting to consider this as evidence of the effectiveness of the default resetting in this particular context. However, such an interpretation neglects the effects the default resetting had at the individual level on other, non-default contribution rates. In particular, an analysis at the individual level showed that it reduced the frequency of higher contribution rates (the proportion of those selecting contribution rates over 7% fell from 41% to 30%) and increased the frequency of low contribution rates (the number of those selecting less than 3% increased from 5% to 15%). These effects are substantial: a revisit of the data shows that the mean contribution rate is actually lowered by the default resetting (from 6.88% to 6.31%). Consequently, the resetting did not just shift some people from a 3% to a 6% contribution rate. It also affected people with other contribution rates, so that (i) about the same number of people chose the new default and (ii) more people chose a lower contribution, yielding a mean reduction in contributions. Contrary to first impressions, the nudge produced a worse result because of the heterogeneity in the population. One can easily imagine similarly unwanted effects of the Save More Tomorrow (SMT) nudge if the heterogeneity of the underlying discounted evaluation is too large. In such a context, SMT would make some people save less, while other people would save more, without obtaining an optimum for anyone.

Some proponents of nudge policies do consider the possibility of such heterogeneity. Asymmetric paternalism (Camerer et al. Reference Camerer, Issacharoff, Loewenstein, O'Donoghue and Rabin2003), for example, assumes that some members of a population may be fully rational, and hence not in need of a nudge that others require. Asymmetric paternalism refers to an intervention that ‘creates large benefits to those who make errors, while imposing little or no harm on those who are fully rational’ (Camerer et al. Reference Camerer, Issacharoff, Loewenstein, O'Donoghue and Rabin2003: 1212). Consequently, it seeks to devise policies that affect only those ‘in need’ of a nudge. Asymmetric nudges are designed so that those whose behaviour is considered optimal or more than optimal remain unaffected and continue to implement their behaviour, even if deviant from the suggested option. Although a considerable improvement over the equal incompetence assumption, which runs into problems such as those described above, even asymmetric paternalism assumes that those who are subject to the bias are uniformly affected so that the same nudge can steer them toward more optimal choices. Thus, the conclusion remains that to be effective, nudge interventions require sufficient homogeneity in the population's heuristic repertoire.

4.5. Temporal effects

Nudges are designed to trigger the deployment of an existing heuristic to generate a certain behaviour, without trying to convince and sometimes without explicitly informing the agent about the goal of the intervention. That, of course, is the point of nudging: rather than convincing the agent that it would be in her best interest to save more or exercise more now, the nudge circumvents this difficult task by exploiting known tendencies or heuristics that reliably predict behaviours given certain environmental stimuli. The problem with this approach is that because the agent is not actively participating, she might engage in additional behaviour that counteracts or compensates the desired behavioural change.

Imagine, for example, a company policy that aims at making people exercise more, and thus subscribes all employees for 10 monthly gym visits by default. Assume that the policy makes a difference: once it is implemented, people spend more time in the company gym than before. However, spending more time in the gym is not the ultimate goal of the policy: making people exercise more is. It is possible that some people go more often to the company gym, but keep track of their overall exercise time and cut back proportionally on their previous exercise time outside the company gym. Whereas adaptation to boosts implies that a new and effective heuristic is learned and can be applied successfully, adaptations to a nudge are less predictable because the nudged are not consciously aware of the intervention and might be pursuing a different goal than the one intended by the policymaker. While the repeated and successful application of a heuristic acquired via a boost is likely to increase the agent's competence, the effect of a nudge might not persist over time because the repeated implementation adversely affects the stability of the triggering relationship between environmental stimuli and the agent's heuristics. Thus, before designing and implementing a nudge, potential temporal effects, that is, reactions to a nudge over time, must be carefully analysed.

Of course, over time the wearing-off effect might be counteracted by learning effects. For example, a study of energy conservation nudges shows that the effects of a one-time intervention wear off quickly. However, repeated exposure may ‘gradually change [people's] capital stock of habits or physical technologies’ so that ‘after two years, . . . effects decay at only 10 to 20 percent per year’ (Allcott and Rogers Reference Allcott and Rogers2014: 3034). In this case, we suspect that a repeated nudge may come close to a boost in that it ends up slowly changing people's repertoire of strategies and/or resources for action.

Similar evidence concerning the malleability of people's heuristic strategies comes from research into default setting, which shows that after receiving a good default, people are more likely to accept defaults in the future, even if doing so results in worse outcomes than going against the default (de Haan and Linde Reference de Haan and Linde2016). The reverse effect is also possible, however. Let's consider the gym-exercise example again. One likely effect of increased gym attendance is that with time, employees will build preferences. That is, they will discover whether they like going to the gym and whether they feel the perceived benefit is worth their while. For people with clear-cut preferences, ‘default rules are less likely to stick’ (Sunstein Reference Sunstein2016: 118). In that case, an initial increase in gym attendance due to the default might be followed by a drop-off, as only those who actually like going to the gym will stick with it.

Both types of effects – immediate compensations for nudges and changes in habits or preferences over time – might impact the effectiveness of nudge policies, as they modify the desired behaviour in hard-to-predict ways. More specifically, the very intervention that caused the behavioural change in the first place may cause its increase or degradation due to adaptations over time. Although such effects cannot be completely precluded for boosts, they are less likely to occur because the application of a newly learned cognitive heuristic is goal-directed by definition.Footnote ⁴ If, for instance, people apply Temptation Bundling (TB), they do so because they aim to increase their exercise regime. Consequently, they have little reason to compensate for more gym visits afterwards, unless they deliberately decide to do so. Nor should one expect the effect to wear off, because more use of the heuristic means more training in its application, and, as with most training, a more effective application as a consequence.

4.6. Selection criteria

To be effective, boosts require not only that agents learn a new heuristic, but also that they are able to recognize when to apply it under at least some range of variation in environmental conditions. More precisely, in order for a boost to reliably change behaviour in the long run, two conditions must be satisfied. The first is that the agents should be able to recognize the conditions for the effective selection of the new heuristic. The second is that the agents should continue to select the newly learned heuristic rather than fall back on previous heuristics. We focus on the former condition here and turn to the latter in the next subsection.

Boosts, such as Simple Rules of Thumb (SRT) and Temptation Bundling (TB), increase people's competence to flexibly apply heuristics in specific environments. As such, a heuristic ‘may appear simplistic, but their underlying intelligence lies in selecting the right rule of thumb for the right situation’ (Gigerenzer Reference Gigerenzer2007: 49). To solve this so-called strategy selection problem, people must be sensitive to their applicability criteria, which are hidden in the structure of cues in the choice environment. Take, for instance, the recognition heuristic. If you recognize the name of one city, football club or stock on a list but not the others, then this option is likely to rank highest in value and choosing it will likely be successful. However, the recognition heuristic breaks down if you recognize more than one option or if the reason why you recognized a particular option is no longer valid (e.g. the population size of Detroit decreased after the crisis of the automobile industry; the competitiveness of a football club decreased because its best players left, etc.).

Similarly, a simple lexicographic decision rule examines cues in a ranked order one at a time and makes a binary decision based on the first cue that discriminates between two options. Such a simple rule has been shown to perform at least as well as a linear decision algorithm that integrates all informational cues, but only if the environment is such that the predictive power of the highest-ranked cue indeed dominates all others and cannot be outweighed by the remaining cues (Şimşek Reference Şimşek2013). Which cue structure is required for the effective application of a specific rule is an empirical question. The Fast and Frugal literature has shown that there are environments for which simple rules are a good match, but not all environments have a cue structure that can be matched successfully by some simple rule (e.g. Gigerenzer et al. Reference Gigerenzer and Todd1999; Şimşek Reference Şimşek2013). There might be environments in which decisions can be satisfactorily made only with computationally more intense strategies. In sum, when the effectiveness of a boost depends on the cue structure of the environment, people should be taught not only the heuristic but also the right selection rule.

4.7. Individual/environmental resources

What ensures that the agents will continue to select the newly learned heuristic rather than fall back on previous heuristics? Even when the agents are capable of recognizing the conditions for the application of the new heuristic, this in itself does not ensure that they will continue to select it. Other features of the decision environment, such as time, money, will power or value incongruence, might get in the way, especially in the long run. Consider for example the Temptation Bundling (TB) intervention. If personally relevant temptations cannot be effectively matched in terms of timing (e.g. I must go to the gym and want to go bungee jumping), financial means (e.g. I have money only to go to the gym or to go bungee jumping), or values (e.g. working out is healthy and eating cake is pleasant), it might be hard to stick with the new behaviour in the long run. In general, if resources such as time or money become scarce, it is relatively easy to fall back on pre-existing habits and behaviours. Similar problems occur if the new behaviours are not congruent with people's value systems. Even if in the beginning the tension is resolved in favour of the new behaviour, there is no guarantee that this will be maintained in the long run. Thus, whereas nudges change environments to match people's heuristics by definition, when designing and implementing boostable heuristics, environmental conditions such as the particular value system of the targeted population and the (temporal and financial) resources available to them must be well analysed before a boost can be reliably implemented.