Endogenous preferences,environmental economics,and welfare

Abstract

This article targets the problems of endogenous preferences for welfare assessment in the context of large-scale interventions that, for instance, concern us in environmental economics. Prominent economists have argued that the problem of endogenous preferences is particularly pressing in these cases. However, I argue that while there are promising solutions to deal with this problem when it comes to small-scale interventions (e.g. nudging), those do not work for large-scale interventions. Based on this, I offer a framework that allows us to preserve the central role of preferences for welfare assessment also in cases of large-scale interventions.

Introduction

Endogenous preferences have been identified as a pressing problem for environmental economics by prominent economists such as Nicholas Stern and Joseph Stiglitz. They hold that the fact that environmental economics deals with long-time horizons and large-scale interventions undermines the assumption of exogenously given, context-independent preferences (Stern et al., 2022). More precisely, these authors argue that endogenous preferences foil the case for cost-benefit analysis in the context of climate policy and favor other decision rules (e.g. precautionary principles).

Nevertheless, there is also considerable optimism that the explicit recognition of preference endogeneity will allow for better climate policy assessments. To illustrate, mitigation costs could be considerably lower due to preference adaptation. Individuals may, for instance, not only change their diet in response to a policy, but also end up preferring lower levels of meat consumption, making it possible to achieve reductions in emissions at lower welfare costs.

Partially because of such considerations, there is now a small but growing literature on endogenous preferences in environmental economics. For example, Mattauch et al. (2022) model a consumption decision between low-carbon and high-carbon goods, where policy instruments (e.g. taxes or infrastructure programs) change preferences. Moreover, Konc et al. (2021) model consumers who change their preferences under the influence of their peers. However, this literature has not yet developed a coherent framework for thinking about welfare effects in the presence of endogenous preferences.

In light of this, the aim of this article is to use environmental economics and the large-scale interventions it tries to assess as a test case for our ability to rely on preference-based welfare measures for policy assessment. More specifically, I will argue that promising methods for dealing with endogenous or context-dependent preferences when assessing small-scale interventions like nudge policies do not readily translate to the large-scale interventions that concern us, for instance, in environmental economics. Against this background, I then outline a package of assumptions that can still justify relying on preferences for welfare analysis even in large-scale interventions.

Toward this goal, I first outline why giving up on the assumption of exogenously given preferences threatens to undermine preference-based welfare economics. In particular, I highlight the prominent argument that efforts to salvage preference-based welfare economics in the absence of this assumption are wedded to a psychologically implausible understanding of “mistakes” (Infante et al., 2016).

Subsequently, I argue that recent attempts to deal with this problem are not fit for large-scale interventions. For example, Beck (2023) argues that we can select between preferences exhibited in different contexts based on how well they cohere with the overall set of agents’ attitudes. Yet, it is highly plausible that large-scale interventions (like those we are concerned with in environmental economics) also imply substantial shifts in agents’ other attitudes. Moreover, authors like Fabian and Dold (2025) and DesRoches (2020) provide accounts based on the psychologically defensible notions of “agentic” preferences or “value-based” preferences. However, large-scale interventions will likely also affect these preferences. Hence, I argue that if we are dealing with such large-scale interventions, we cannot make use of these frameworks.

Motivated by this discussion, I then offer a new way of looking at the problem that can support the appeal to preferences to also assess large-scale interventions. My proposed framework relies on the assumption (a) that preferences in the contexts of interest are fundamentally fuzzy, i.e. they merely specify a range.

Moreover, my framework builds on (b) the so-called evidential account (Hausman and McPherson, 2009), i.e. the idea that preferences (under the right circumstances) are (strong) evidence for welfare. This account denies that preference satisfaction is identical to welfare. Yet, it reserves a prominent role for preferences in welfare judgments.

The final assumption (c) concerns additional sources of evidence about welfare (cf. Hausman, 2025). As the evidential account is compatible with combining such evidence with evidence from preferences, we can use it to go beyond fuzzy preference-based welfare judgments, thereby allowing us to arrive at more informative policy assessments.

The article is structured as follows. I first present the problem of endogenous preferences. Subsequently, I argue that there are some promising solutions to deal with this problem for small-scale interventions like nudging and healthcare interventions, but not for the large-scale interventions that concern us, for instance, in environmental economics. Finally, I outline my framework that preserves the central role of preferences for welfare analysis, also in cases of large-scale interventions.

Endogenous preferences and welfare economics

This article starts from the observation that the problem of endogenous preferences is particularly pressing in fields like environmental economics that deal with large-scale interventions. The reason is that large-scale interventions not only undermine the assumption of exogenously given or context-independent preferences, but also speak against potential solutions for addressing this issue that have been developed in the literature (see Stern et al., 2022). ¹

Of course, the fact that people change their behavior due to large-scale interventions could also be explained in terms of changing beliefs. Maybe people alter their behavior not because of changes in their preferences but because of changes in how they perceive the world. For instance, a new infrastructure policy may lead to people switching from cars to public transport—to a greater extent than what could be explained by a change in the relative costs of these options—not because of a change in how much they “like” the two options, but because the change in policy enabled them to learn more about the respective options. In the latter case, we could still represent their preferences by a stable utility-function, while updating the agents’ beliefs about the relevant options.

However, I will assume that not all responses to the kind of large-scale policy interventions we are interested in can be understood along these lines. On the one hand, it appears plausible that at least some policy interventions cannot be readily conceptualized as belief updating. For instance, consider an agent deciding whether to buy an energy-efficient durable good that, because of the uncertain potential for energy savings, appears to offer chances for losses and chances for gains relative to their current status quo (see Greene, 2011). Imagine we ask whether we should mandate that only these efficient goods can be sold. Now, assume that our agent has context-dependent preferences that can be represented by a prospect theory model, e.g. our agent values perceived losses with respect to a certain reference point (i.e. their status quo) at twice the rate of gains. ² Further, assume that even if the mandate is implemented, the situation will still be perceived as a risky bet by the agent. Under these assumptions, implementing the mandate can alter what the agent takes the relevant reference point to be, and thereby change their preferences. In a context in which the mandate is not implemented, the agent would value potential losses related to the good in question at twice the rate of the respective gains. Yet, in a context in which the mandate is implemented, taking the risky bet associated with the good becomes the new status quo. This will lead the agent to value losses and gains related to the good at the same rate. Such changes in the agent's valuation can effectively change their preferences for or against the energy-efficient durable good without any belief updating happening. To see this, consider that if this—admittedly somewhat stylized—scenario occurred, we could “switch” the agent's preferences back and forth by wiggling whether the mandate is in place or not. Hence, we cannot plausibly construe what happens here as instances of belief updating.

What is more, apart from the artificial, but clear-cut example above, environmental economists have modeled more complex situations like the impact of peers on choice behavior or the impact of infrastructure programs on consumption as cases in which behavior is not only explained by changes in costs or beliefs but involves also genuine preference change (e.g. Konc et al., 2021; Mattauch et al., 2022). Moreover, Bernheim et al. (2021) recently provided a framework for modeling endogenous preferences that can be used to build more specific models of applied phenomena. Considering this, I will proceed under the assumption that belief updating or changes in relative costs do not always suffice to explain changes in behavior when it comes to the large-scale interventions that motivate us here. ³

With this in the background, I can now explain why endogenous preferences are so problematic for welfare economics. For the types of welfare assessments economists usually perform, preferences need to satisfy certain characteristics. In particular, they need to (i) exhibit certain structural properties (e.g. transitivity, context-independence, stability), (ii) concern our own welfare (i.e. have the right content), and (iii) be sufficiently based on the relevant facts (cf. Hausman, 2016). Only for preferences that satisfy these conditions is it plausible to say that they exhibit a close relationship to welfare. ⁴

The main problem that concerns us here relates to the structural properties specified in (i). More specifically, it is about the property of context-independence or exogeneity. If this property is violated, agents can exhibit different preferences in different contexts. Hence, it becomes unclear in which context the agents exhibit the preferences that truly track their welfare. In principle, we could then end up with a case in which an agent's preferences change with the policy we implement. Yet, we would like preferences to be the standard by which we adjudicate which policy we should implement. If preferences change with the policy, they are unable to fulfill this role. In other words, if preferences are endogenously determined, the yardstick with which we want to measure welfare changes with the system we want to measure.

This problem is discussed extensively in the growing literature on behavioral welfare economics. So, a natural suggestion for dealing with endogenous or context-dependent preferences, also when it comes to the large-scale interventions, would be to look at how behavioral welfare economists deal with this problem. Hence, we will now have a look at this.

The search for true preferences

Foundations of behavioral welfare economics

Behavioral welfare economics has mainly been concerned with addressing problems arising from violations of condition (i), particularly context-dependence. Most efforts of this research program are aimed at developing methods for identifying a set of “true” or laundered preferences that can be used to make welfare assessments despite context-dependency or endogeneity.

This is motivated by the antipaternalistic stance of welfare economics, i.e. the idea that respecting (certain) attitudes of the agent is a requirement of welfare analysis. There are various reasons why this is seen as a requirement of welfare analysis. For instance, it is argued that people usually know better what is good for them (cf. Hausman and McPherson, 2009). An additional—and in my view even stronger—reason for the antipaternalistic stance is that welfare policy needs to respect preferences to be legitimate. This second argument is, for example, developed by Haybron and Tiberius (2015: 717), who argue that legitimate welfare policy “must not impose an external standard of wellbeing on people.” While they do not think that this means that we should trust people's preferences at face value—they think that not every preference people reveal in their choices is indicative of what is good for them—Haybron and Tiberius hold that we need to base welfare assessment on people's values. In making this claim, they have a very specific conception of values in mind. They take them to be “relatively robust pro-attitudes, or clusters of pro-attitudes, individuals take to generate reasons for action (Haybron and Tiberius, 2015: 723).” The topic of how to best conceptualize values is, of course, important for welfare measurement (see Jahn and Beck, 2023). However, here I want to simply point out that Haybron and Tiberius make quite a common move: We are required to show some form of respect for what people value or prefer. Yet, this requirement does not entail respecting all attitudes of an agent, e.g. their revealed preferences. Instead, what we have to respect is what people truly value (under some specification of what this means). In behavioral welfare economics, this type of attitude is sometimes referred to as people's “true” preferences. What is important for the view of Haybron and Tiberius, though, is that they do not presuppose that welfare is grounded in what people value or truly prefer. Yet, they argue that welfare policy has to proceed under the assumption that this is the case in order to be legitimate. They call this pragmatic subjectivism.

I hold that pragmatic subjectivism offers good reasons for giving a strong role to subjective attitudes of the affected people when it comes to assessing their welfare. However, one could argue that if revealed preferences are context-dependent and we need to go on a tedious hunt for “true” preferences, why not simply ask the agents what they think is good for them? Would that not be the most straightforward way to respect the antipaternalistic requirement? As Bernheim (2016: 21) puts it, “if choice is simply a constructed judgment, one could argue that other types of constructed judgments should be equally admissible for the purpose of evaluating welfare.”

The answer welfare economists give here is that talk is cheap. One of the main reasons to focus on preferences and choices in welfare economics is the fact that people's stated preferences (i.e. verbal judgments) often differ from the choices they actually make. As long as the agent does not have to live with the consequences of their judgment, how can we make sure that the purpose of making the judgment and the aim of the analysis are the same, i.e. improving welfare (Bernheim, 2016)? Hence, one strong reason to favor preferences over verbal judgments is that when we observe people choosing, we can usually be confident that they are in the business of making a choice that promotes their well-being. ⁵

So, how then are we supposed to get at people's true preferences? Various theoretical frameworks for operationalizing preference laundering (i.e. the method by which we identify true preferences) have been proposed. To illustrate, Bleichrodt et al. (2001) propose a methodology that is based on the idea that cumulative prospect theory (CPT) is descriptively adequate, i.e. it captures how agents actually behave. CPT models behavior in terms of a basic utility function, probability weights, and a loss aversion parameter. After estimating all of the elements via observing people's choice behavior, the core idea of Bleichrodt et al. (2001) is to take the basic utility function laundered from the other elements and use it within an expected utility theory (EUT) framework.

More generally, approaches to preference laundering often conform to the idea of supplementing the standard models of choice in economics (i.e. EUT) with additional elements representing the “cognitive biases” that allegedly prevent preferences from determining choices. For this to succeed, we need to assume that there is a utility function that captures choices. I will call this U(x, f ); f denotes the context of choice. Yet, on top of this, we also need to assume that there is a context-independent normative objective function, V(x), according to which we can evaluate welfare. In other words, we need to assume that there is a function that somehow represents an agent's “true” preferences. Under these assumptions, the difference, b(x, f ) = U(x, f ) − V(x), can be understood as reflecting an agent's biases. ⁶ The thought is then that if enough is known about the process mapping “true” preferences to choices (i.e. about the bias function b(x, f )), one can recover true, context-independent preferences based on observations about people's choice behavior.

One main application of these methods is the justification of so-called nudges. Thaler and Sunstein (2008: 6) define a nudge as “any aspect of the choice architecture that alters people's behavior in a predictable way without forbidding any options or significantly changing their economic incentives.” The classic example is a cafeteria manager who decides to place a healthy snack in a more prominent location than an unhealthy snack while keeping both options available. As a result of this, agents will be more likely to choose healthy snacks instead of unhealthy snacks. Yet, as both options are still available and there are no noteworthy costs of going for the unhealthy snack, the agent's option set is argued to remain the same. However, keeping the option set the same is only one of the requirements that proponents of nudges want to respect. The other requirement is the antipaternalistic stance, i.e. respecting people's preferences. But how are we supposed to do this if people reveal different preferences in different contexts, e.g. if their preferences are determined by which of the snacks is placed more prominently? The answer is that we can appeal to the agent's context-independent true preferences. So, preference laundering methods are frequently employed in the justification of nudges. Yet, other prominent examples can be found in health economics (see Lang et al., 2024).

So, why not just apply these methods to the problem of endogenous preferences also in the case of large-scale interventions? What I wish to point out here is that some promising approaches from this literature may be well-suited for cases like nudging and health economics, where we deal with small-scale interventions, but not for the large-scale policies that concern us, for instance, in environmental economics. To see this, I will first have to say more about the criticism that preference laundering methods have received.

Critics have argued that preference laundering rests on psychologically implausible assumptions (see Infante et al., 2016; Sugden, 2018). In particular, it has been argued that there is no psychological mechanism that can be plausibly seen as grounding true, context-independent preferences (i.e. V(x)). The critics have called this purported mechanism the inner rational agent. They argue that this inner rational agent is assumed to be clearly separable from other mechanisms (b(x, f )), that distort the preferences generated by the inner rational agent in such a way that we end up with the mistaken/actual preferences that we can observe (U(x, f )). Under this reconstruction, preference laundering is seen essentially as an attempt to uncover the preferences generated by the inner rational agent from the distortion of other bias-introducing mechanisms. Sugden (2018) and other critics persuasively argue that this two-stage model of human psychology, on which they claim preference laundering relies, is false. In other words, the only way to uphold the idea of preference laundering is to commit to a psychologically implausible account of mistakes.

Coherence-based approaches

Yet, several attempts have been made to show that preference laundering is not committed to this model and can be viewed as a feasible enterprise. Some authors (i.e. Beck, 2023; Hausman, 2016; Jahn and Beck, 2023) have proposed what I will here call coherence-based approaches. The core idea is that we can select between preferences exhibited in different contexts based on how well they cohere with the overall set of agents’ attitudes. Apart from coherence-based approaches, we also find accounts based on the psychologically defensible notions of “agentic” preferences or “value-based” preferences, which are intended to be more stable than our ordinary context-dependent preferences (e.g. DesRoches, 2020; Fabian and Dold, 2025). It is argued that these special types of preferences can be used to deal with context-dependence in our ordinary preferences. In the following, I will argue that while both sets of approaches have initial plausibility when it comes to small-scale interventions, they face severe obstacles when it comes to large-scale interventions. ⁷

Let us start with coherence-based approaches. According to these approaches, the question is not how the agent's preferences would look like if they were not distorted by certain psychological mechanisms. In fact, we do not need to assume that cognitive processes can be neatly separated into the categories of rational and biased. Instead, what we are concerned with is how a preference relates to other attitudes of the agents. To illustrate, imagine Simon, who always chooses a piece of cake instead of fruit when he visits the cafeteria at his workplace. Yet, in a different context, for example, when training for a marathon, Simon always deeply regrets this choice and expresses a preference for fruit. Which preference in which context tracks Simon's welfare better? Coherence-based approaches would suggest looking at other attitudes of Simon, for example, his feelings of regret and his intention to perform well in the marathon, and ask which of the two preferences aligns better with these attitudes. We then base our welfare judgments on the better cohering preference, which in this case appears to be the preference for fruit.

This is, of course, only a rough sketch of this proposal, and there are various issues that need to be resolved. Depending on how we specify the coherence requirements on our attitudes, this proposal will deliver different results. Moreover, what about cases in which none of the agent's context-dependent preferences cohere well with their other attitudes? Does this also give us reason to base our welfare judgments on a purely hypothetical preference that would cohere better? I hold that these are pressing questions for coherence-based approaches, but I take it that they are answerable (for an attempt to address some of these challenges, see Beck, 2023).

However, I hold that there is a further issue that is far more problematic for coherence-based approaches. In order for these accounts to get off the ground, we need to assume that the overall profile of the attitudes of the relevant agents remains stable across contexts. If we want to rely on coherence with these attitudes to determine on which preferences we should base our welfare analysis, but these attitudes also change with the context, coherence-based approaches are in trouble. Different attitudes in different contexts may cohere better or worse with different (context-dependent) preferences. If we face such a situation, there is no clear answer to the question of which preferences cohere better with the agent's attitudes.

Yet, it is highly plausible that large-scale interventions also imply substantial shifts in agents’ other attitudes. For example, in 2007, Ljubljana unveiled “Vision 2025,” a comprehensive plan aimed at making the city greener, cleaner, and more sustainable. Central to Ljubljana's vision for an eco-friendly city was the conversion of the city center into a car-free zone. As a result, noise pollution and emissions in the area were significantly reduced. Nevertheless, there were protests when the plan was announced, and many people voiced their opposition. However, once implemented, the car-free zone was met with high levels of approval (for further details, see EU Urban Mobility Observatory, 2020).

What exactly did happen here? One explanation would, of course, be that people realized how great the car-free zone was. They simply got new information that they previously lacked, but their preferences stayed the same. In this case, it is clear that we should go with the better-informed preferences to assess welfare. For example, people might have been concerned with the possibility that many shops in the inner city will close as a result of the car-free zone. Once they realized that this would actually not happen, they may have changed how they viewed the option that is on the table. While they may prefer shops staying open to shops closing, they may also prefer shops staying open and having the car-free zone to shops staying open and not having the car-free zone. As it turns out, the choice was actually between the latter two options. While I cannot fully dismiss the possibility that the initial opposition was mainly based on false information, I would like to explore an alternative explanation.

Another possibility is that people simply exhibit some form of status quo bias. That is, among the two options (car-free or not car-free), they always prefer the option that is currently being implemented. Like in my previous example, we could then simply switch their preferences by exchanging the status quo policy. While this is, of course, somewhat simplified, it is a clear-cut example that preference laundering methods should be able to address. So, how would coherence-based approaches deal with this case?

They would try to determine the preference that coheres better with the agent's overall attitudes. However, I take it to be quite plausible that in our example of the car-free zone in Ljubljana, these are also in flux. While it is plausible that our other attitudes are not affected by the choice of where a cafeteria manager places the cakes (we may not drop our goal to run a marathon because of it), implementing a car-free zone is different. It has the potential to affect many of your other attitudes, e.g. how much you like being outside, how much you value efficiency in traveling compared to taking a bit longer but having time to mentally prepare for your workday, and so on. Of course, as a matter of empirical fact, we might get lucky, and it can turn out that most of the agents’ attitudes remain stable. However, if it turns out that they do not, coherence-based approaches are in trouble as there is now no fixed set of attitudes that we can use to adjudicate between people's context-dependent preferences.

What the example of the car-free zone illustrates nicely is that the larger the scale of our intervention is, the more likely it is that coherence-based approaches run into problems because, on top of our preferences, also the attitudes based on which we are meant to select between preferences exhibited in different contexts become more likely to change. Hence, while coherence-based approaches appear promising when it comes to preference laundering concerned with small-scale interventions—e.g. nudges and healthcare interventions—they are less convincing when it comes to the large-scale interventions.

“Value-based” or “agentic” preferences

Let us, therefore, briefly turn to other approaches that want to defend preference laundering against Sugden's critique. Some authors, i.e. DesRoches (2020) and Fabian and Dold (2025), propose that there is a special type of “value-based” or “agentic” preferences that are grounded in something like robust proattitudes that generate reasons for acting/choosing. Those preferences are argued to be psychologically plausible. Fabian and Dold (2025), for instance, draw on self-discrepancy theory and self-determination theory to sketch the process through which individuals generate “agentic” preferences. ⁸ While the details differ between these two proposals, I take it that they have much in common. The core idea is that agents can have “agentic” or “value-based” preferences that are closely linked to their well-being, and are relatively stable, compared to our ordinary context-dependent preferences.

To illustrate how this can help with nudging, consider the following example: Agents may have a “value-based” or “agentic” preference for walking more. Yet, as it stands, they never take the stairs at their workplace. If it turns out that we can change their behavior and make them take the stairs by repainting the stairs so that they look like a piano, the approaches under consideration would suggest that we should do so because taking the stairs is truly in line with their “value-based” or “agentic” preference (see Fabian and Dold, 2025).

One might think that these approaches are not too dissimilar from the coherence-based approaches. However, there are some key differences. Coherence-based approaches give no special place to the agents’ other attitudes. All they say is that sometimes we can select between different preferences exhibited in different contexts by asking which of them coheres better with the agents’ other attitudes. No attitude is privileged in this process. In principle, the most straightforward way to make an agent more coherent could be to drop or change one of their other non-preference attitudes. For example, consider an agent who cannot stop talking about their goal of walking more. Yet, they never take the stairs, no matter how they are painted, and show no other attempts to walk more. The most straightforward way to make this agent more coherent could be to drop their goal to walk more. We should then simply conclude that walking more is not good for them.

Maybe proponents of “value-based” or “agentic” preference would like to say that the agent I just described has no “value-based” or “agentic” preference for walking more, and they would agree with coherence-based approaches. Be that as it may. I do not want to overstate or settle the differences between the views. What is important is that the two sets of accounts suffer from many of the same problems when it comes to large-scale interventions. What prevents an agent's “value-based” or “agentic” preference, e.g. a preference for walking more or eating healthier, from changing in the same way as their other attitudes in the case of large-scale interventions? DesRoches (2020) already states that in the long run, our value commitments can also change and that his approach is, therefore, best suited for what he calls “Joe-in-the-cafeteria cases.” Hence, independently of whether these approaches are really relevantly different from coherence-based approaches, they seem to face the same problems once we leave the domain of nudges and other small-scale interventions behind.

A new framework

In the previous section, I have argued that the approaches aimed to defend preference laundering against Sugden's critique seem to be well-suited to handle small-scale interventions, e.g. cafeteria cases, but are ill-equipped to deal with large-scale interventions. Against this backdrop, I now want to suggest a new way of looking at the problem. As I shall argue, this reframing of the problems will allow us to justify reliance on preferences also when it comes to the large-scale interventions that often concern us in environmental economics. More specifically, it justifies relying on a certain interpretation of the Bernheim and Rangel (2009) framework in those cases.

In a nutshell, my own framework

assumes that preferences in the relevant contexts are fundamentally fuzzy (instead of changing with context),

commits to an evidential account of the relationship between preferences and welfare, and

appeals to preference-independent evidence for welfare.

Fundamentally fuzzy preferences

The first and most crucial assumption is that preferences merely specify a range. In other words, they can be fundamentally fuzzy.

My motivation for this assumption comes from decision-theoretic work on imprecise credences. The concern here is that precise probabilities often fail to adequately represent severe uncertainties. In some cases, we can assign a precise probability to an event. For example, when rolling a die, each possible outcome can be given a precise probability. Decision-making involving such precise probabilities is referred to as decision-making under risk. Traditionally, this is contrasted with decision-making under uncertainty, where the probability of an event is unknown.

However, we are often in an intermediate state, where we lack sufficient information to assign a precise probability, yet we are not in a state of complete ignorance. The epistemic attitude we should adopt in these cases is more nuanced, reflecting uncertainty in a “fuzzier” form. Increasingly, decision theorists recognize the limitations of modeling belief states with precise probabilities (e.g. Joyce, 2010; Seidenfeld, 2004).

Consider, for instance, a person on the academic job market asking their advisor about their chances of securing a position (event E). A precise reply such as “your chances are 0.035” might raise suspicion. A more plausible response would be that the chances lie somewhere between 0.03 and 0.1. If the advisor's guess is the only information available, the applicant's epistemic attitude toward their job prospects might be better represented by a closed, convex set of probabilities rather than a single, precise probability.

Of course, in real-life conversations, advisors are unlikely to express their assessment in terms of a closed, convex set of probabilities. Nevertheless, the key point remains: regardless of how the advisor chooses to communicate their judgment, the applicant's resulting attitude is more appropriately modeled by a set of probabilities (e.g. [ $\underset{\_}{\mathrm{P}}(E)=0.03$ , $\overline{\mathrm{P}}(E)=0.1$ ]) than by a single figure (for a particularly clear outline of this topic, see Wollesen forthcoming).

Hence, it appears plausible that some of our epistemic attitudes can and should be fuzzy. Yet, if we already accept the idea of imprecise credences, the idea of fuzzy evaluative attitudes is not too far off.

Let us now return to welfare economics. The Bernheim and Rangel framework is an approach toward welfare economics that allows for ambiguous welfare judgments. While it does not by itself require that we should understand evaluative attitudes as fuzzy, in the following, I will argue that this idea sits particularly well with this approach. But before doing so, I first have to outline the key concepts of the Bernheim and Rangel (2009) framework.

At the core of the framework lies the distinction between direct judgments, which pertain to outcomes we care about for their own sake, and indirect judgments, which pertain to alternatives that lead to those outcomes (see Bernheim, 2021, 2025). For example, a direct judgment may pertain to my mental states (e.g. how much I like to taste apples), while an indirect judgment may pertain to goods that get me into those states. Based on this distinction, the approach defends the following two premises.

Premise A: With respect to matters involving either direct judgment or correctly informed indirect judgment, each of us is the best arbiter of our wellbeing.

Premise B: When we choose, we seek to benefit ourselves by selecting the alternative that, in our judgment, is most conducive to our wellbeing.

Premise B is not remarkable in the context of welfare economics. If we drop it, choices would simply not be a good guide to welfare, regardless of whether we are dealing with context-dependent preferences. Hence, I will not further discuss B here.

What about premise A? Here, the core idea is that there are some judgments about which you are the ultimate authority. For instance, no one can tell you how much you like the taste of apples. Yet, this does not mean that we will have to defer to all of your judgments. Imagine that I place two boxes in front of you. Both are opaque. In one, I place an apple; in the other, I place a kiwi. Now, imagine you can observe what I am doing. You then get a choice between the boxes and pick the one that contains the apple. In this case, even though your choice would be what Bernheim calls an indirect judgment, we still seem to have good reason to rely on it when assessing your welfare. You appear to be correctly informed.

However, let us now assume that I distract you and exchange the content of the boxes. You then choose the box in which I have placed the kiwi. Should we trust that this choice is best for you in terms of your well-being? The answer appears to be no. According to Bernheim (2021), your judgment here is merely an indirect one, and because of my distraction, it is likely to be an incorrectly informed one. Such incorrectly informed indirect judgments should not be used to guide our welfare assessment. Note that for Bernheim, direct judgments seem to be the type of judgments that cannot be incorrectly informed. That is, you cannot be incorrectly informed when it comes to judging how much you like the taste of an apple. Hence, we can simplify Bernheim's claim and say that we should not base our welfare analysis on incorrectly informed judgments, that is, those that are based on false beliefs about the options.

There are further issues that complicate the matter. For example, whether we classify a judgment as incorrectly informed will always depend on assumptions about the relevant domain of objectives. Sometimes this is easy to identify. For instance, investors want to maximize their return on investment. In our example, we tacitly assumed that you want to choose the type of fruit you like better (instead of making a choice between two opaque boxes). Yet, if we are wrong about the relevant domain of objectives, our attribution of an incorrectly informed judgment is likely to be wrong.

I will not try to resolve this issue here. Instead, I want to emphasize that while there are still many unanswered questions about when we are licensed to classify a judgment as incorrectly informed, it is clear that there are some incorrectly informed (indirect) judgments and that there are also cases like my kiwi example in which we can easily and uncontroversially identify them. ⁹

With its basic premises clarified, we can now move on to the central concept of the Bernheim and Rangel framework, i.e. what is called a Generalized Choice Situation (GCS) (Sugden, 2018). These do not only consist of a set of options from which the individual must choose, but also of a set of “ancillary conditions.” Ancillary conditions are defined as properties of the choice environment that may affect people's preferences or choices, but not the options from which they are choosing.

Equipped with this concept, the framework then proceeds in the following way. We first exclude all GCSs whose ancillary conditions are suspect from the welfare-relevant domain. What do we mean by suspect ancillary conditions here? Suspect ancillary conditions are those likely to lead to incorrectly informed (indirect) judgments. Just think of the kiwi example, in which the agent got distracted. Judgments made in these conditions are unlikely to track your welfare well. Hence, we should exclude those conditions from the welfare-relevant domain, that is, not base our welfare analysis on preferences displayed in those conditions.

Once we have excluded all suspect GCSs from the welfare-relevant domain, we look at the preferences that individuals display in the remaining GCS. If those are not affected by changes in the remaining ancillary conditions, we may end up with a standard utility function that tracks people's welfare. However, there is no assumption here that preferences are context-independent. If we have excluded all suspect GCS and preferences still change with ancillary conditions, Bernheim (2021: 391) advises to “live with whatever ambiguity remains.” What does that mean in practice?

Suppose Simon has two standing-room tickets to a football game and is wondering if he should use them or sell them (see Bernheim, 2021). His willingness-to-accept (WTA) differs across “ancillary conditions.” Let us assume that there are only three of them:

In A, career success becomes salient, and Simon's WTA is 20€.

In B, enjoyment of the game becomes salient, and his WTA is 30€.

In C, Simon thinks he has seating tickets, and his WTA is 100€.

In this example, the Bernheim–Rangel framework implies that we should exclude C, as this is clearly a suspect GCS. This means that we have to base our judgment about how good having and using the ticket is for Simon only on A and B. Of course, we get different answers from Simon's choices in A and B. Nevertheless, if those are the only two conditions, we can conclude that having and using the tickets improves Simon's welfare by no less than 20€ and by no more than 30€. Now, we have to ask what exactly we mean by this claim.

There are two options that we have available here:

Option 1: We just do not know where exactly on the relevant range (e.g. between 20€ and 30€) Simon's “true” preferences lie.

Under this option, we face the critique of the inner rational agent again. We have to say that Simon's choices are the result of his true valuation of the ticket plus some bias. As I have argued before, this is highly unattractive. It would basically associate the Bernheim and Rangel framework with those approaches to preference laundering that suffer from psychological implausibility. What is more, under this option, it remains unclear why we assume that the true preferences need to lie in the range delineated by the choices in the different ancillary conditions. If we assume that all choices are biased in the same direction but to a different degree, the true preference of the agent could also lie outside of the range delineated by the choices. All in all, I do not take this option to be an attractive interpretation of the Bernheim and Rangel framework. Luckily, there is a second option that is also more in line with Bernheim's (2021) more recent work.

Option 2: Simon's preferences merely specify a range, i.e. they are fundamentally fuzzy.

This option is much more attractive. First, it does not share the downsides of Option 1. What is more, it also aligns well with proposals in decision theory that conceptualize our attitudes as fuzzy. Option 2 is, thus, the more appealing interpretation of the Bernheim and Rangel framework. Hence, the framework and the idea of fuzzy preferences appear to be a natural fit. ¹⁰

This is already good news. The resulting view avoids the worry that we measure with different yardsticks in the presence of endogenous or context-dependent preferences. It does so by introducing fuzzy preferences, which hold across different contexts.

Importantly, in cases like the example of Simon, we can still reach (some) decisive verdicts by just looking at the relevant fuzzy preferences. For instance, if we could acquire a football ticket for Simon for merely 10€, based on his fuzzy preferences, we can conclude that welfarist considerations favor doing so.

Note that this is basically a reframing of the problem. But why is this reframing justified? If we have good reasons to disregard some of the context-dependent preferences of an agent, e.g. on coherence grounds or based on incorrectly informed indirect judgments, we may still end up with a model of the agent that conforms to the standard assumptions of welfare economics. Yet, how should we view the agent in the absence of such reasons? We can, of course, just say that the agent has irreducible endogenous or context-dependent preferences and that we have no reason to disregard any of these preferences. Yet, this would mean that we are still stuck with the problem. In contrast, reframing what we observe as an instance of fuzzy preferences allows us to arrive at a model (or interpretation) of the agent that can still be usefully employed in welfare analysis. If both of these interpretations are equally permissible, pragmatic consideration related to our goal of welfare analysis seems to suggest adopting the latter.

So far, everything looks quite nice. However, in the next section, I will outline a problem for the resulting view that motivates the other two assumptions, namely the evidential account and other evidence about welfare.

The evidential account and further evidence

Here is the bad news: If people's (welfare-relevant) preferences specify merely a range, then preferences sometimes allow us to make only highly uninformative welfare judgments.

To see this, let us return to the example of a decision about an energy-efficient durable good. Because of the uncertain potential for energy savings, it appears to offer a 50–50 chance of a gain of 1500€ or a loss of 1000€ (see Greene, 2011). We consider two “ancillary conditions”:

In A, we “value” losses at twice the rate of gains.

In B, we “value” losses at the same rate as gains.

Should we mandate that only these efficient goods can be sold? Our framework delivers the verdict that the welfare effect of the policy is no less than −250€ and not more than 250€. Hence, it does not give us any meaningful indication of whether the policy is good or bad in terms of its impact on the agent's welfare.

Here is where the evidential account comes in (Hausman and McPherson, 2009). It holds that preferences (under the assumption that they satisfy conditions (i), (ii), and (iii) specified above) are merely (strong) evidence for welfare. This account denies that preference satisfaction is identical to welfare. Yet, it still reserves a prominent role for preferences in welfare judgments.

Why adopt this account? The most straightforward reason is provided by Richard Kraut (2007: 118) when he asks, “is it the case that whenever S wants (rationally, with proper information, and reflectively) P to occur, and P does occur, that is good for S simply because it is something he wants in that way?” Proponents of the evidential account hold that the answer is “no.” What makes P good for S is whatever the good-making features of P are, but never merely the fact that S wanted those good-making features. In other words, if we could perform a miracle intervention that changes S's preferences for the good-making features of P but leaves all other details about the world untouched, P would still be good for S. In what follows, I assume that this argument is correct. Yet, under this assumption, we require an argument for why so many people hold that preference satisfaction constitutes welfare.

The answer given by the evidential account is that the strong evidential role preferences can play in welfare analysis often gets confused with a constitutive role. If we assume that people are good judges of their own welfare, then looking at preferences that concern their own welfare and are based on the relevant facts seems like a good idea when it comes to gathering evidence about what is really good for people. Of course, as we have seen, sometimes people have incorrect beliefs about the available options. In other cases, we may be skeptical of a preference because it coheres badly with other goals and attitudes of the agent, as coherence-based approaches suggest. Yet, these are special cases. If we have an agent who is concerned with judging their own welfare, we usually have strong reasons to consider their preferences to be good evidence for welfare analysis. The conflation of this strong evidential role of preferences with a constitutive role can plausibly explain why some hold that preference satisfaction constitutes welfare. So, the evidential account is not itself a theory of welfare, but compatible with various theories of welfare as long as they grant a strong evidential role to preferences.

Why does it matter to clearly distinguish between the evidential and the constitutive role of preferences in welfare analysis? The reason is that the evidential account allows us to make the following suggestion: If we treat preferences as mere evidence for welfare, we can go beyond preferences and include other information about welfare to reach a more definitive verdict than the uninformative verdict we get in the durable good example.

More specifically, I propose to treat people's preferences as upper and lower bounds on permissible welfare judgments. My reasons for this are grounded in the anti-paternalistic stance of welfare economics and the claim of Haybron's and Tiberius’ (2015) pragmatic subjectivism, which holds that legitimate welfare policy needs to be based on agents’ values. In the case of fuzzy preferences, this means our chosen policies need to be compatible with the range specified by the agent's preferences. Hence, even if there is strong preference-independent evidence about how good a policy is for an agent, but we cannot find any ancillary conditions in which the agent would choose in line with what our other evidence suggests, i.e. the other evidence points us outside of the range of the agent's fuzzy preferences, we have non-welfarist reasons for not imposing this policy on the agent. Staying within the range specified by the agent's fuzzy preferences can then be seen as a guardrail that prevents policymakers from becoming overly paternalistic and doing welfare policy that is illegitimate by the standards of Haybron's and Tiberius’ pragmatic subjectivism. ¹¹

Let me illustrate the resulting view. In the case of the energy-efficient durable good, given that both imposing and not imposing the policy is compatible with the agent's fuzzy preferences, we may be able to rely on other evidence, e.g. evidence from a randomized control trial or an instrumental variable estimation that shows a difference in welfare (according to a preference-independent metric) that can be attributed to the good in question. Alternatively, we could rely on generalizations about wellbeing for further guidance (see Hausman, 2025). No matter the details of how we generate such additional evidence, the core idea is that preferences, as mere evidence for welfare, can be combined with other information about welfare to reach more definite verdicts. Importantly, any such information does not concern whether the agent's preferences in either ancillary condition A or B are their “true” preferences. Instead, what is relevant here is just further evidence about what is good for people.

Importantly, this approach will also work for large-scale interventions where those approaches tailored only to “Joe-in-the-cafeteria cases” can offer no further guidance. ¹² To illustrate this, consider the model developed by Mattauch et al. (2022). They aim to examine the impact of large-scale climate policy-induced changes in consumer preferences. They are interested in policies like implementing a carbon tax, changes to transport infrastructures, and policies that can lead to societal changes in dietary habits. Toward this goal, they develop a highly stylized theoretical model that abstracts away from detailed cognitive and social mechanisms. They represent the influence of policies on preferences by adding a parameter α to the agent's utility function. In their model, an agent has to choose between a clean good (C) and a dirty good (D). A higher α means that more weight is placed on C. In this way, the agent's utility is determined by C, D, and α (i.e. U(C,D,α)). They then proceed by exploring various implications of a tax on dirty consumption that affects not only the price of D but also α.

In the context of this project, they also comment on welfare analysis. They note that we could still compare the different preferences that result from changes in α by appealing to so-called meta-preferences. The idea here is that the agent may, for example, “like herself more when intrinsically motivated to protect the environment,” that is, when she exhibits a higher α (Mattauch et al., 2022: 12). This is effectively saying that the welfare-relevant preferences are those that she exhibits at the highest possible level of α. Hence, we are dealing with an attempt at identifying the agent's true and welfare-relevant preferences. In this case, the agent also seems to be aware of what her true preference would be. After all, she has a meta-preference for having these preferences. Nevertheless, the agent fails to act on these preferences. While this picture may fit some agents in certain idiosyncratic circumstances, arguing that people usually have such meta-preferences comes dangerously close to endorsing the two-stage model of human psychology on which Sugden (2018) bases his critique of the inner rational agent. Therefore, I do not think that meta-preferences are an attractive option for doing welfare analysis in the context of the policy decision that Mattauch et al. try to illuminate with their model.

Another option that the authors consider is that there is a “unit by which one can compare different preferences and their corresponding utility functions (Mattauch et al., 2022: 12).” In this context, they note that a “consumer may experience greater utility from higher α from a positive self-image (Mattauch et al., 2022: 6).” The problem with the proposal for the present context is that it departs from a preference-based way to measure welfare and instead seems to work with an experiential notion of wellbeing. This illustrates the need to think more carefully about welfare analysis when dealing with the type of large-scale interventions the authors try to model.

So, what does my framework suggest for how we should do welfare analysis in those cases? The answer is straightforward. Different policies lead to different levels of α and, thereby, to different utility functions. Unless we have a good reason for excluding some of the circumstances that lead to a specific α from the welfare-relevant domain, this set of utility functions should be used to determine the fuzzy preferences of the agent i.e. the range of utilities that represents how good the outcome is for the agent in question. This highlights that my framework is not only applicable to relatively small-scale interventions like in the example of the energy-efficient durable good but can, with the help of modeling tools such as the ones developed by Mattauch et al. (2022), also be employed in assessing large-scale interventions. In such a case, the fuzzy preferences represented by a set of utility functions can then either help us to directly inform policy or provide the first step in a process that needs to be supplemented with preference-independent evidence for welfare.

What about the quality of preference-independent evidence for welfare? My stance here is that the answer will depend heavily on context. Alexandrova (2017) persuasively argues that we have better prospects for gaining the understanding required for developing welfare measures by building theories geared to specific groups of people in specific contexts (i.e. so-called mid-level theories). If this is correct, also our prospects for good preference-independent measures may differ between contexts. While we may have a theory for guiding the selection of the relevant components of welfare and their respective weights in some contexts, we may lack such a theory for other contexts. Nevertheless, even in cases where we do not have good mid-level theories, we may still be able to make use of what Hausman and McPherson (2009) would call platitudes, or rely on what Sunstein (2020) calls an incompletely theorized agreement about welfare. In line with this view, I here do not argue for any specific proposal for developing such welfare measures.

Instead, one way to operationalize my framework is as follows: Suppose we have a non-preference-based measure for the context at hand that leads to a welfare evaluation of x for a specific outcome. Further suppose that the fuzzy preferences of an agent for that outcome are given by [ $y$ , $z$ ] (e.g. [−250, 250]). My proposal is then to calculate the value we should use for welfare assessment by min (z, max (x, y)).

In other words, once we have $x$ we can either use it to select a point on the range of ambiguity given by the agent's fuzzy preferences, or if x lies outside this range, we use it to justify a choice of one of the endpoints of the interval given by the agents’ fuzzy preferences, i.e. y if x is smaller than y and $z$ , or z if x is larger than y and $z$ . ¹³ For example, assume that the fuzzy preferences of an agent for the outcome in question are given by [−250€, 250€]. Further, assume that our non-preference-based measure yields a welfare evaluation of 50€. In this case, my proposal would suggest using 50€ for our welfare assessment. Yet, if our nonpreference-based measure leads to an evaluation of 300€, we should use 250€.

Of course, my framework does not guarantee that we can always generate additional evidence that would allow us to determine x. As a consequence, we can remain stuck with an uninformative evaluation of a policy in terms of welfare. So, what should we do? The silver lining is that even then, preferences can still play an important role as guardrails.

Conclusion

I have tried to suggest a way to rethink the problem of endogenous preferences, which is especially useful for dealing with large-scale interventions such as the ones investigated by environmental economics. While it may still be possible in the assessment of small-scale interventions to deal with context-dependency by disregarding all but one preference, for example, on coherence grounds, for large-scale interventions, we might be better advised to reinterpret agents’ preferences as fuzzy. I have pointed out how we can still reach informative verdicts about welfare policy in line with this reframing.