2. Context-independent proxy

There are several proposals for how to identify which choices merit deference that aim to reconcile welfare economics with recent behavioural findings. One of the proposals that comes closest in spirit to traditional welfare economics assumes that only those choices, among all possibly context-dependent choices, that remain constant across context and time (context-independent choices) are reliable proxies of welfare.

This position, which I call the context-independent proxy of welfare, has been formalized by Bernheim and Rangel (Reference Bernheim and Rangel2009) (hereafter B&R). ^{Footnote 6} They consider a general framework for describing choice behaviour in which X denotes the set of all possible alternatives such as consumption bundles or any other state of affairs, as long as alternatives are complete and mutually exclusive descriptions of the world. To model context-dependent behaviour, B&R define a generalized choice situation, denoted $GCS = (A,d)$ , as the combination of a standard choice situation $A \subseteq X$ and an ancillary condition d. An ancillary condition can be the way in which information is presented, the labelling of a particular option as the status quo, or any other feature of the choice environment as long as it ‘may affect behaviour, but is not taken as relevant to a social planner’s evaluation’ (B&R: 55).

Let ${{\cal G}^*}$ denote the set of all GCSs contemplated by an observer, ^{Footnote 7} and assume that for all choice situations $A \in X$ there is some ancillary condition d such that $(A,d) \in {{\cal G}^*}$ . Individual behaviour is modelled through a choice correspondence $C:{{\cal G}^*} \Rightarrow X$ , that assigns a non-empty set of alternatives $C(A,d) \subseteq A$ to every generalized choice situation $(A,d) \in {{\cal G}^*}$ . An alternative $x \in C(A,d)$ is interpreted as an option that the individual selects and is willing to choose when facing $(A,d)$ . Welfare analysis can then be performed by defining a welfare binary relation, P, where xPy means that x is ‘better than’ y. B&R’s preferred welfare ranking of alternatives is based on what they call an ‘unambiguous choice relation’, denoted ${P^*}$ and defined as follows:

(1) $$x{P^*}y\;{\rm{if\,and\,only\,if\,for\,all}}\;(A,d) \in {{\cal G}^*}\;{\rm{such\,that}}\;x,y \in A,\;{\rm{we\,have}}\;y \,\notin C(A,\,d).$$

In other words, x is said to be better than y if and only if y is never selected when x is available. Since by assumption every subset of X, including $\{ x,y\}$ , is in the domain of ${{\cal G}^*}$ , this means that x is only said to be better than y when x is chosen at least once over y and y is never selected when x is available. Thus, only choices between two alternatives that remain stable for all observed choice situations and ancillary conditions determine the welfare ranking of different alternatives.

At this point, it is worth noting that B&R are agnostic about the process that gives rise to choices. In contrast to other authors (e.g. Rubinstein and Salant Reference Rubinstein and Salant2012), they do not assume the existence of an underlying context-independent stable preference that can be reconstructed by eliminating mistakes. The presumption is that ‘choices provide appropriate guidance because they are choices’ (B&R: 52), or in other words, respecting choices is required to satisfy individual ‘self-determination’ (Bernheim Reference Bernheim2009: 290–293). Prima facie, their proxy seems like a natural extension of the principle of individual sovereignty to settings in which context-dependent behaviour is prevalent (see however, section 5.3 below).

This approach is appealing because it relies exclusively on choice data. However, this same advantage will often lead to a welfare ranking that is not very discerning and that becomes less so as the number of choice observations increases (Rubinstein and Salant Reference Rubinstein and Salant2012). In such circumstances, many pairs of alternatives x and y are not comparable under ${P^*}$ (hereafter denoted $x{N^*}y$ ). In answer to this criticism, B&R deviate from their context-independent proxy and propose to ‘prune’ ${{\cal G}^*}$ by using non-choice data to delete ‘suspect’ GCSs. This refinement allows them to identify a welfare-relevant domain, ${\cal G} \subseteq {{\cal G}^*}$ , consisting of all GCSs that merit deference and from which the observer takes normative guidance. Note that when referring to the context-independent proxy, I mean the criterion presented above that does not take this refinement into account. In other words, the context-independent proxy does not prune ${{\cal G}^*}$ (i.e. ${\cal G} = {{\cal G}^*}$ ). On the other hand, the next two proxies I will discuss can be represented as criteria for pruning ${{\cal G}^*}$ , and I will illustrate their welfare inferences using this approach. ^{Footnote 8}

Using B&R’s framework and their context-independent proxy of welfare, the introductory example can be represented as in Table 1.

Table 1. The snack choice: Context-independent proxy

The apple and the snickers are not comparable according to the context-independent proxy because the chosen alternative in the advance choice condition ( ${d_1}$ ) is contrary to the one chosen in the immediate choice condition ( ${d_2}$ ). In other words, the context-independent proxy is agnostic about which snack is welfare superior because there are conflicting choice patterns.

3. Reason-based proxy

Even when confronted with the evidence that observed behaviour is often ‘inconsistent’, economists usually take the satisfaction of a given stable and context-independent preference as the benchmark for welfare analysis (e.g. Koszegi and Rabin Reference Koszegi and Rabin2007; Rubinstein and Salant Reference Rubinstein and Salant2012; Apesteguia and Ballester Reference Apesteguia and Ballester2015). One approach among these, which underlies many recent economic models, is to assume that only choices that are made after rational deliberation (reason-based choices) are reliable proxies of welfare. ^{Footnote 9}

A prominent example of this reason-based proxy of welfare is given by ‘dual-system’ models recently popularized by Kahneman (Reference Kahneman2011) (see Alós-Ferrer and Strack Reference Alós-Ferrer and Strack2014 for a short review of different models). According to this view, human psychology can be divided into two systems or modes of thought: one fast, effortless and automatic (System 1), and another slow, effortful and controlled (System 2). In economics these models have been used, among other things, to represent the intrapersonal conflict between present and future preferences (e.g. Thaler and Shefrin Reference Thaler and Shefrin1981; Bernheim and Rangel Reference Bernheim and Rangel2004; Fudenberg and Levine Reference Fudenberg and Levine2006, Reference Fudenberg and Levine2012). For instance, Bernheim and Rangel (Reference Bernheim and Rangel2004) build a model to study addictive behaviour in which an agent alternates between a ‘hot mode’ and a ‘cold mode’. Whenever ‘cued’ towards the hot mode the agent always takes an addictive behaviour ‘irrespective of underlying preferences’, while in the cold mode she ‘considers all alternatives and contemplates all consequences’ and selects her most preferred alternative (p. 1559). They assume that agents maximize a context-independent and stable preference relation on their cold mode and that choices taken under the hot mode are ‘mistakes’. Choices made after ‘cold’ deliberation – reason-based choices – are assumed to be reliable (and consistent) proxies of welfare.

A criterion based on rational deliberation has also found support in a recent influential book in which Hausman (Reference Hausman2012) aims to describe how the concepts of preference, value, choice, and welfare are and ought to be used in economics. In the book, Hausman argues that a preference is and ought to be an evaluation in the sense that it is the result of a rational deliberation about what one has most reason to do. This means that, according to Hausman’s view, a person’s choice that is not based on rational deliberation about what she has most reason to do does not reveal a preference. It follows that such a choice cannot be used as a proxy of welfare. Only reason-based choices are potential proxies of welfare. ^{Footnote 10}

B&R’s framework can be used to illustrate the welfare inferences of the reason-based proxy of welfare. Using this approach, data on internal deliberation prior to choosing can be used to delete certain GCSs. For example, if internal deliberation prior to choosing is ‘too fast’, then the corresponding GCS is excluded. In practice, it is often difficult to determine whether or not internal deliberation is rational in specific choices made by specific individuals, so observers rely on indirect data gathered from studies in psychology, neuroscience and other fields (e.g. eye-tracking or neuroimaging studies).

With these premises in mind, we can revisit the snack choice example. Evidence suggests that, on the one hand, immediate food consumption generally triggers impulsive behaviour (or System 1), while, on the other hand, being far from consumption generally encourages reason-based deliberation (or System 2) (e.g. Read and van Leeuwen Reference Read and van Leeuwen1998). According to the reason-based proxy, this information suggests that the immediate choice should be excluded from the welfare-relevant domain ${\cal G}$ and that the apple is welfare superior to the snickers as judged by Norah. This is represented in Table 2.

Table 2. The snack choice: Reason-based proxy

4. Confirmed proxy

I have distinguished two prominent behavioural proxies of welfare that equate some (as opposed to all) choices with welfare. In this section, I propose a new proxy of subjective welfare: the confirmed proxy of welfare.

Consider a discrete time horizon ${\cal T} = \{ 1, \ldots ,T\}$ and assume, without loss of generality, that all generalized choice situations $(A,d) \in {{\cal G}^*}$ are ordered in time from 1 to $T - 1$ . Assume that the observer makes his/her welfare or policy evaluation at period T. Assume as well that the observer wants the welfare or policy evaluation to be reliable at period T. ^{Footnote 11} Then, for any period $t \in {\cal T}$ :

In other words, a choice is said to be confirmed whenever an individual would want herself to repeat that choice at the time of the welfare or policy evaluation if faced with the same menu and the same ancillary conditions. In practice, this proxy demands that, alongside choice data, an observer records self-reports at T about what an individual would want herself to do at T. I call such self-reports stated meta-choices. This can be elicited using questions at period T, such as: ‘From where you stand now, would you want yourself to choose the same alternative again?’ Stated meta-choices should not be confused with stated choices, which correspond to what an individual thinks she would choose. This difference is relevant, as many people, in many circumstances, would like to behave differently from how they think they would behave (e.g. in cases of limited self-control; see Benjamin et al. Reference Benjamin, Heffetz, Kimball and Rees-Jones2012 for empirical evidence).

We can use B&R’s framework to represent some welfare inferences of this proxy. In this case, ‘who’ prunes ${{\cal G}^*}$ is no longer the observer but the individual herself. Take again the snack choice example. Recall that at period 3 (after the two choices have been made) Norah would want herself to repeat her choice of the snickers over the apple but not her choice of the apple over the snickers. Table 3 represents the snack choice example according to the confirmed proxy of welfare.

Table 3. The snack choice: Confirmed proxy

In this example, the confirmed proxy provides an unambiguous welfare inference in favour of snickers over apple for period T. ^{Footnote 12} However, this is not always the case. Sometimes, the confirmed proxy provides a prudent inference. To see this, consider the case of addictive behaviour represented in the payoff table taken from Dalton and Ghosal (Reference Dalton and Ghosal2012: 594) (Table 4).

Table 4. The smoking choice

In Table 4, ${a_2}$ corresponds to smoking and ${a_1}$ corresponds to not smoking, and ${h_i}$ represents Norah’s health states ( ${h_1}$ being healthier than ${h_2}$ ). Assume that Norah chooses repeatedly from this set of feasible actions, ${a_1}$ and ${a_2}$ , in a ‘long-run’ horizon from periods 1 up to $T - 1$ (see Dalton and Ghosal Reference Dalton and Ghosal2012: 588–593). Suppose as well that she believes, wrongly, that her health state is stable over time. It follows that she always prefers to smoke at the beginning of the time horizon ( ${a_2}$ is the dominant action for each h). However, in the long run Norah’s health state deteriorates to ${h_2}$ and the unique long-run outcome is ( ${a_2}$ , ${h_2}$ ) with a payoff of 0. ^{Footnote 13} Now suppose that in period T Norah states that she would like to quit smoking. ^{Footnote 14} The confirmed proxy uses this piece of information to make more reliable comparisons of subjective welfare at period T. This self-report reveals that Norah’s observed behaviour, even though it is context-independent, is not aligned with how she would want herself to behave at T. According to the confirmed proxy, given that there is a conflict between choices and stated meta-choices at T, an observer should take an agnostic position at T as to whether smoking is better for Norah than not smoking as actually judged by herself.

It is worth noting, however, that the confirmed proxy can inform policy interventions even when it provides a prudent inference. For example, assume that a social planner is considering policies on smoking. The planner would like to set a policy framework that takes into account people’s choices and the inference that many smokers, like Norah, would like to quit smoking. The planner should not prohibit smoking, as this would be against smokers’ revealed preferences. At the same time, the planner should not only not prohibit smoking but also provide opportunities to ‘unwilling’ smokers to follow the behaviour that they would want themselves to follow (i.e. not smoking). Providing free consultations with specialized doctors is a policy in that direction. Since the two policies – non-prohibition and free consultations – are not mutually exclusive, the planner would respect choices and stated meta-choices and potentially improve the welfare of smokers who would like to quit smoking.

5. Why confirmed choices

In my opening remarks, I mentioned three justifications that have been traditionally used to support the link between choice and welfare. In this section, I revisit these justifications to support the link between confirmed choices and welfare. The overall argument can be summarized as follows:

1. (1) If A is a more reliable proxy of preference satisfaction than B at t, if A is a more reliable proxy of SWB than B at t, and if A is more respectful of individual sovereignty than B at t, then A is a more reliable proxy of subjective welfare than B at t as far as the most common theories of subjective welfare in economics are concerned.

2. (2) Confirmed choices are a more reliable proxy of preference satisfaction than context-independent and reason-based choices at T.

3. (3) Confirmed choices are a more reliable proxy of SWB than context-independent and reason-based choices at T.

4. (4) Confirmed choices are more respectful of individual sovereignty than context-independent and reason-based choices at T.

5. ––––––––––––––––––––––––––––––

6. (5) Therefore, confirmed choices are a more reliable proxy of subjective welfare than context-independent and reason-based choices at T as far as the most common theories of subjective welfare in economics are concerned.

In the following, I provide support for premises [2], [3] and [4] (sections 5.1, 5.2 and 5.3 respectively). While doing so, I link my arguments to common phenomena of interest to economics, such as updating beliefs, changing preferences, ex post regret, taste for variety, habituation and limited self-control.

6. Discussion

The previous arguments suggest that confirmed choices have important advantages over context-independent and reason-based choices as proxies of welfare. These advantages are particularly prominent when assessing what is good for a person as judged by herself at the time of the welfare or policy evaluation (period T). The focus on a person’s welfare as judged by herself accords with the tradition in economics of treating individual subjective attitudes as decisive in assessing the relative welfare associated with different alternatives. I depart from neoclassical economics, however, by requiring synchronicity between the observer’s and the observed person’s judgements. This requirement seems sensible as soon as we acknowledge, as often done in psychology, philosophy and behavioural economics, that individuals’ subjective attitudes change over time. Synchronicity is thus a necessary condition for respecting individual attitudes at the time of the welfare/policy evaluation, which I have argued to be an essential quality of a proxy of subjective welfare (section 5.3). Requiring synchronicity is also important for the proxy of welfare to be robust to common behavioural phenomena such as updating beliefs, changing preferences and ex post regret. While the timing of an observer’s evaluation is often overlooked in the literature, some authors have articulated views that point towards synchronicity between the observer’s and the observed person’s judgements. For example, Gul and Pesendorfer (Reference Gul and Pesendorfer2005: 433) argue that to determine whether a policy ‘improves the welfare of the agent it suffices to determine whether the agent would vote for the policy in the period in which it is introduced’.

This raises the question, however, of whether confirmed choices at T are a reliable proxy of welfare for periods other than T. ^{Footnote 18} To see this, consider the following example. Suppose that last Wednesday, at $T - 1$ , Norah went to a theatre matinee and loved it. This Wednesday, at T, Maria (a friend of Norah who is deciding what they will do that day) asks Norah if she would want herself to repeat the choice that day; Norah says no. However, the reason behind Norah’s answer is that on the previous night she went binge-drinking with other friends and that today she does not feel like going to the theatre. Without knowing this last piece of information, Maria takes Norah’s answer into consideration and decides not to invite her to the theatre matinee that day. This seems to be a correct inference by Maria, the ‘planner’, about what is best for Norah as judged by herself for that day (period T). However, Maria’s inference may be incorrect for other periods. Next Wednesday, at $T + 1$ , Maria may decide not to invite Norah to the theatre when actually Norah would love to go.

This example illustrates the problem that may arise in instances of intertemporal substitutability or complementarity of a good with itself or other goods. Although this problem is not exclusive to the confirmed proxy, it is useful to understand how it impacts the reliability of its welfare inferences. On the one hand, as the example illustrates, this issue is not problematic for the confirmed proxy’s welfare inferences if a policy is introduced at T and its consequences are restricted to T. On the other hand, this issue can be problematic if intertemporal substitutability/complementarity impacts stated meta-choice at T in ways that it does not affect other periods, as this makes the welfare inference especially contingent on T.

A similar issue may arise with intertemporal choices, i.e. ‘decisions in which the timing of costs and benefits are spread out over time’ (Loewenstein and Thaler Reference Loewenstein and Thaler1989: 181). Examples include how much schooling to obtain, how much to save for retirement, buying a house or choosing an insurance policy. While the confirmed proxy accounts for intertemporal implications before T that the other proxies fail to account for, the confirmed proxy may not account for intertemporal consequences after T. This will be the case if the stated meta-choices at T disregard these consequences or incorrectly account for future tastes (for theory and evidence on the incorrect account of future tastes, or ‘projection bias’, see Loewenstein et al. Reference Loewenstein, O’Donoghue and Rabin2003 and Frederick et al. Reference Frederick, Loewenstein and O’Donoghue2002: 373).

Another potential objection to the confirmed proxy is that it gives a prominent (even if auxiliary) role to self-reports. Economists have traditionally been suspicious of self-reports. The usual criticism is that talk is cheap (e.g. Grüne-Yanoff Reference Grüne-Yanoff2012: 643). I agree. However, self-reports seem to reveal information about people’s goals and values that is not captured in choice behaviour and that it is important for an observer’s welfare or policy evaluation (see also Hirschman Reference Hirschman1984; Lewis Reference Lewis1989; Beshears et al. Reference Beshears, Choi, Laibson and Madrian2008). For example, stated meta-choices that are contrary to choice behaviour seem to decrease our confidence that revealed preferences are welfare-enhancing; if a consumer writes a bad review of a product, we should be less confident that the product the consumer bought is good for her as judged by herself. In the confirmed proxy, self-reports help the observer to make an educated guess about individuals’ attitudes towards their behaviour at the time of the welfare or policy evaluation (see Manzini and Mariotti Reference Manzini and Mariotti2014: 344 for an argument in favour of using non-choice data for a similar purpose).

Even so, self-reports are more reliable in some contexts than in others. In some typical economic settings, such as consumption and labour market behaviour, people are usually familiar with the context and they have few reasons to be dishonest or strategic in their answers about what they would want themselves to do. However, this is not necessarily the case in other contexts. For instance, self-reports are likely to be unreliable when concerning criminal behaviour, tax-avoidance, or other forms of antisocial behaviour, since deception and/or self-deception are likely. In other contexts, such as difficult or unfamiliar decisions (e.g. some healthcare choices), self-reports, like choices, are likely to be unreliable due to false or incomplete information. In still further contexts, where self-reports may have a political impact, individuals may respond strategically. Perhaps even more generally, self-reports, like choices, may be susceptible to changes in viewpoint or framing.

The concerns highlighted in this section suggest that in some contexts it may be important to create favourable conditions for revealing well-considered self-reports, i.e. self-reports that are honest, informed, reflected, and robust to trivial changes in viewpoint or context. Many procedures can help to create these ‘favourable conditions’. For example, one can provide general information and/or other aids for decision-making (e.g. impartial advice from an expert); introduce truth-telling-commitment devices that help the elicitation of honest self-reports (e.g. asking people to sign a solemn truth-telling oath before giving their answers, as in Jacquemet et al. Reference Jacquemet, Luchini, Rosaz and Shogren2019, Reference Jacquemet, Luchini, Malezieux and Shogren2020); present several frames of the same issue to counter framing effects (e.g. in Druckman Reference Druckman2001, Reference Druckman2004; Benjamin et al. Reference Benjamin, Fontana and Kimball2020); point out choices/self-reports that are inconsistent with compelling postulates of decision-making (e.g. point out non-transitive answers, as in Tversky Reference Tversky1969; see Benjamin et al. Reference Benjamin, Fontana and Kimball2020: section V for a review); highlight intertemporal consequences if present (e.g. reminding people to consider implications for $T + 1$ and onwards); and/or inform individuals about the timing of the policy introduction if the policy is not introduced at T (e.g. Maria, in the example above, could have asked at T if Norah would want herself to go to the theatre at $T + 1$ ).

These favourable conditions can be interpreted as a particular ancillary condition. The previous discussion suggests a refined confirmed proxy that uses this particular ancillary condition to elicit well-considered confirmed choices:

In other words, a choice is said to be considerately confirmed whenever an individual would want herself to repeat that choice at the time of the welfare or policy evaluation if faced with the same menu under favourable conditions for revealing well-considered self-reports. While recording confirmed choices may be enough in a variety of contexts, recording well-considered confirmed choices seems desirable (or even necessary) in contexts that face the challenges mentioned in this discussion.

At this point, a reader may wonder whether the well-considered confirmed proxy is not similar to the reason-based proxy of welfare. To see that this is not the case, consider how data are used in each proxy. On the one hand, the reason-based proxy usually relies on indirect data about rational deliberation (e.g. eye-tracking studies that assess how attentive people are in different contexts) in order to determine which choices merit deference. In addition, this proxy usually relies on the behaviour of some individuals to determine when other individuals follow rational deliberation. On the other hand, the well-considered confirmed proxy demands self-reports to be well-considered when individuals are themselves asked to determine which of their choices merit deference. Therefore, these two proxies can lead to very different welfare inferences. According to the arguments in the previous section, the welfare inferences from the confirmed proxy are more reliable than those that arise from the reason-based proxy. The discussion in this section suggests that the well-considered confirmed proxy can correct for potential mistakes (as judged by individuals themselves) linked to ‘fast’ thinking, that could in principle grant an advantage to the reason-based proxy in contexts where ‘slow’ reasoning is important.

Concerning the context-independent proxy, its main advantage is that it is less demanding than the confirmed proxy in terms of data. Even though the confirmed proxy only requires gathering extra data at a single point in time, this can be a relevant advantage in some contexts. However, gathering the extra data for the confirmed proxy is important whenever one expects phenomena such as updating beliefs, changing preferences, ex post regret, habituation or other reasons for changes in subjective attitudes to be prevalent. It is also important when one expects conflicts between what people do and what they would like themselves to do (e.g. in cases of limited self-control). In addition, a well-considered confirmed proxy corrects for potential mistakes (as judged by individuals themselves) that can make the context-independent proxy unreliable, due, for example, to a lack of information or projection bias. The union of these phenomena seems non-empty for many (if not most) contexts, suggesting that the extra data required for the confirmed proxy is often necessary if the observer wants to have a more reliable proxy of welfare.

In sum, the arguments in section 5 suggested that confirmed choices are a more reliable proxy of welfare than context-independent and reason-based choices. This holds especially for the period of the welfare or policy evaluation (period T). In this section, I have argued that this inference holds for other periods and many (if not most) contexts, particularly if – whenever deemed necessary – one adopts favourable conditions for revealing well-considered self-reports.

7. Nudges and boosts

The arguments presented so far have implications for two influential behavioural policy programmes that are non-incentivizing (they do not provide monetary incentives) and non-coercive (they do not forbid or impose options), nudges and boosts. I discuss each in turn.

Nudges generally consist of interventions that change the choice architecture (the background environment against which people make decisions) to alter people’s behaviour in a predictable way without eliminating freedom of choice or significantly changing their economic incentives (Thaler and Sunstein Reference Thaler and Sunstein2008). ^{Footnote 19} These interventions are usually intended to promote the welfare of targeted individuals as judged by themselves (Thaler and Sunstein Reference Thaler and Sunstein2008: Ch. I). ^{Footnote 20} Paradigmatic examples include displaying information about the socially acceptable behaviour of others, setting the default option to opt-out in pension schemes (as opposed to opt-in), and presenting healthier food ‘at eye level’ in cafeterias, in order to direct individuals into more social, prudent and healthier behaviours respectively.

Stated meta-choices address at least two challenges facing ‘nudgers’ who are committed to improving the welfare of ‘nudgees’ as actually judged by themselves. First, stated meta-choices provide a way for a nudger to infer whether a nudgee is making a bad choice as actually judged by the nudgee. This information is usually not known, and it is necessary (though not sufficient) for a nudge to respect the individual sovereignty of targeted individuals (see Sugden Reference Sugden2009: 371). Second, stated meta-choices provide a way for a nudger to distinguish between people with different goals and infer the distribution of goals in the population. This information is essential for steering individuals with different goals towards different ‘optimal’ options and for avoiding nudges that are arbitrary or implement special interests (see Grüne-Yanoff and Hertwig Reference Grüne-Yanoff and Hertwig2016: 166). This means that stated meta-choices provide a criterion for choosing the ‘direction’ to which nudges shall steer behaviour that all the while respects individual sovereignty and does not rely on nudgers’ external judgement about what is best for others.

Nudges directed by stated meta-choices (hereafter confirmed nudges) can be targeted to a non-arbitrary sub-population. In the case of smoking, for example, a confirmed nudge that aims to steer people towards giving up smoking can target the sub-population of smokers who would want themselves to stop smoking. Importantly, this nudge would not interfere with smokers who have no desire to quit smoking. In practice, this could be implemented by creating a database of potential nudgees who would like to quit smoking. ^{Footnote 21}

Confirmed nudges have at least one ethical, one behavioural and one welfare advantage over traditional nudges that are not targeted to a sub-population. The ethical advantage lies in the fact that confirmed nudges are more likely than traditional nudges to not interfere with people that do not want to be nudged. It follows, I conjecture, that confirmed nudges are more likely to be accepted by people than nudges that are not targeted to a sub-population (see Hedlin and Sunstein Reference Hedlin and Sunstein2016, Reisch and Sunstein Reference Reisch and Sunstein2016 and Arad and Rubinstein Reference Arad and Rubinstein2018 for surveys on the acceptability of nudges). The behavioural advantage lies in the reasonable assumption that an individual is more likely, ceteris paribus, to change her behaviour if she does not confirm it (e.g. a smoker is more likely to quit if she would want herself to quit than if she would not want herself to quit, all else being equal). It follows that confirmed nudges are more likely to be effective in steering the behaviour of targeted individuals than traditional nudges. Finally, confirmed nudges are more likely to have positive welfare implications than traditional nudges (both for most targeted individuals and on average). This is particularly the case when confirmed nudges are directed to a sub-population with similar goals. When this is the case, confirmed nudges exclude most people who would be made worse off from nudging and include people who, according to their stated meta-choices, could be made better off from nudging.

It is worth noting that Thaler and Sunstein (Reference Thaler and Sunstein2008: 80, 116–117) sometimes appeal to a New Year’s resolution test to support nudges. They ask, rhetorically, ‘how many people vow to smoke more cigarettes … in the morning next year?’ Very few, we are prompt to agree. However, what this question omits is that many smokers do not vow to smoke fewer cigarettes in the morning next year either. Confirmed nudges are desirable because they have the potential to benefit targeted individuals (e.g. smokers who vow to smoke fewer cigarettes) while imposing no costs on individuals who confirm their behaviour (e.g. smokers who do not vow to smoke fewer cigarettes). Using stated meta-choices to target nudges, as advocated here, is therefore different from using the self-reports of a limited set of people to justify nudging other people as sometimes endorsed by Thaler and Sunstein (Reference Thaler and Sunstein2008).

I now turn to the second behavioural policy programme that focuses on interventions that usually target individuals’ skills, knowledge and set of decision-making tools (their ‘heuristic repertoire’) to help them to apply their existing or new set of competencies more effectively (see Grüne-Yanoff and Hertwig Reference Grüne-Yanoff and Hertwig2016; Hertwig and Grüne-Yanoff Reference Hertwig and Grüne-Yanoff2017; Grüne-Yanoff Reference Grüne-Yanoff2018). The goal of such interventions is to ‘boost’ the decision makers’ set of competencies such that they identify and reach their objectives (Grüne-Yanoff and Hertwig Reference Grüne-Yanoff and Hertwig2016). Examples include improving the representation of statistical information in health brochures to improve patients’ understanding of different treatments (Gigerenzer et al. Reference Gigerenzer, Gaissmaier, Kurz-Milcke, Schwartz and Woloshin2007), and providing physicians with ‘fast-and-frugal’ decision trees for screening their patients in order to improve their performance in doing so (Jenny et al. Reference Jenny, Pachur, Williams, Becker and Margraf2013).

The arguments in this paper suggest that it may be relevant to boost individuals’ ability and opportunities for reflecting upon their past behaviour. The underlying assumption is that by being able to better appraise their behaviour, individuals will be better prepared to reach their objectives.

Several interventions can be justified on these grounds. A prominent example is (non-coercive) ex post ‘cooling-off periods’ that aim to encourage individuals to critically reconsider their own past decisions. Ex post cooling-off periods have been used, for instance, on door-to-door sales in the USA, by imposing that these sales need to be accompanied by a written statement informing the buyer of her right to rescind the purchase within three days of the transaction (Thaler and Sunstein Reference Thaler and Sunstein2008: 250). They allow people to change their minds after choosing, which can be relevant for people to make better decisions as judged by themselves.

It is worth noting that Thaler and Sunstein (Reference Thaler and Sunstein2008) and other soft paternalists also support ex post cooling-off periods. However, their justification lies in the ability of these interventions to countervail limited self-control. Therefore, Thaler and Sunstein (Reference Thaler and Sunstein2008: 250) support ex ante cooling-off periods as well, such as mandatory waiting periods before a couple can get divorced. Importantly, while ex ante cooling-off periods shrink the opportunity set of an individual (Grüne-Yanoff Reference Grüne-Yanoff2012: 638–639, 644), non-coercive ex post cooling-off periods enlarge the opportunity set with the option to cancel one’s decision. The arguments in this paper provide support for ex post cooling-off periods, and not for (mandatory) ex ante cooling-off periods.

Another relevant example is the design of methods to elicit subjective attitudes (e.g. in the health and environmental domains). For instance, the use of interactive designs in which individuals are asked to reflect upon their choices has the potential to encourage individuals to form considered subjective attitudes that are less liable to choice reversals (see e.g. Slovic Reference Slovic1995: 369–370; Bleichrodt et al. Reference Bleichrodt, Pinto and Wakker2001: 1499; Gilboa Reference Gilboa2010). In line with the boost programme, one could create heuristics for boosting individual competencies that would supplement these opportunities for ex post reflection. For example, Benjamin et al. (Reference Benjamin, Fontana and Kimball2020) implement a procedure for the elicitation of risk preferences that not only provides the opportunity for individuals to revise their choices, but also breaks down the independence axiom of expected utility theory into ‘baby steps’ that are easy to understand.

8. Concluding remarks

When data on choice behaviour are available and conflicting choice patterns are present, some observers are faced with the task of discriminating between choices for welfare or policy evaluation. I proposed confirmed choices as a reliable (though fallible) proxy of subjective welfare for welfare or policy evaluation at a given point in time. This proxy uses choices as the main ingredient combined with auxiliary data – stated meta-choices at the time of the welfare or policy evaluation – that is often available and sometimes easy to collect. According to the arguments presented in this paper, this proxy has decisive advantages over two influential proxies of welfare based on context-independent and reason-based choices. Finally, I have also argued that stated meta-choices can usefully inform behavioural policy programmes that use nudges and boosts.

It is worth emphasizing that the comparison of the three behavioural proxies of welfare made in this paper is sound under the proviso that it applies to the most common theories of subjective welfare in economics. Therefore, my analysis excludes subjective notions of welfare that are used less frequently in economics, such as those based on experiences of pleasure and pain or on the activation patterns of specific areas in the brain. Future comparisons of different proxies of welfare would also benefit from empirical studies (e.g. experiments) that aim to compare these proxies directly.

Finally, it is worth noting that, given the potential contingency of choices, preferences and other subjective attitudes on changes in viewpoint or context, false or incomplete information, and adaptation, among other phenomena that may impair welfare inference, it can sometimes be difficult to identify credible rankings based on subjective information. Objective information, such as that concerning people’s adaptation to their conditions, may be relevant in some contexts for welfare or policy judgements. Yet, we may not want to forget about subjective information altogether, but try instead to find richer and more reliable data sets that include information on both individual choices and self-reported attitudes. As argued above, this sometimes demands creating favourable conditions for revealing self-reports that are honest, informed, reflected and robust to trivial changes in viewpoint or context. Hence, without the presumption of being able to recover stable and context-independent latent preferences in every situation, we may still be able to record choices and self-reported attitudes that are meaningful for normative analysis for a given context and time.