Behavioural mechanisms and public policy design: Preventing failures in behavioural public policy

Introduction

Since more than a decade, insights from behavioural economics, psychology, neuroscience and other behavioural sciences have been integrated in the design of policies across multiple areas and jurisdictions (Lourenco et al., 2016; Straßheim and Beck, forthcoming; United Nations, 2017). Especially the programme of research known as the ‘heuristics and biases approach’ is inspirational by showing that many errors in the judgement of citizens or even experts can be traced back to a specific set of cognitive mechanisms (Kahneman and Frederick, 2002; Kahneman and Tversky, 1982 [1974]). Under conditions of uncertainty and whenever situations are complex, individuals may use a conceptually, semantically or logically simpler judgement (a ‘heuristic’) as a substitute for the more complex assessments. People may answer difficult questions as if they were simple ones. This basic mechanism of ‘attribute substitution’, of shifting from a complex mode of deliberation (called ‘system 2’) to a fast and simple mode (called ‘system 1’), may lead to behavioural biases and failures (Kahneman, 2011; Kahneman and Frederick, 2002, 2005). This paper discusses the role of mechanisms in public policies inspired by behavioural sciences and the implications for preventing policy failures.

There are at least three reasons why these insights on cognitive mechanisms became so instructive in public policy:

First, behavioural approaches help to explain the failures of those approaches in policy design that are based on rational mechanisms of action (Datta and Mullainathan, 2014; Howlett and Mukherjee, 2014). More often than not, programmes, policies or instruments fail because people act against their own interests (at least as defined by policy-makers). Incentives are taken up less enthusiastically than expected; information is ignored; resources are used for the wrong purposes; habits prevail even when confronted with high risks or immediate danger (Bovens and t' Hart, 2016; Dunlop, 2017; Nair and Howlett, 2017). Rational choice explanations treat these phenomena either as the result of unrecognized factors such as hidden costs or informational asymmetries or simply as residual deviations from the standard model (Howlett, 2012). In contrast, behavioural economics argue that people systematically act against their own interests because of heuristics that work as cognitive shortcuts in situations of complexity. By identifying cognitive mechanisms leading to biases, behavioural approaches might help policy designers to realize that

the assumptions we make - sometimes without realizing - when we design programs do not match the way people actually make decisions. Our intuitions - and those in economic models - overlook many of the important things that make people tick. (Datta and Mullainathan, 2014: 11)

Third, and as a result of the increased focus on second-order mechanisms, policy actors and decision-makers themselves have become objects of behavioural studies (BIT, 2018; World Bank, 2015). As the most recent report on ‘Behavioural Government’ issued by the BIT (2018: 3) argues, ‘elected and unelected government officials are themselves influenced by the same heuristics and biases that they try to address in others’. This helps to understand failures in decision-making and raises awareness of biases in public policy. For example, the ‘confirmation bias’ is probably one of the most frequent biases influencing the judgement of policy professionals. It refers to the selective gathering (or the undue relevance) of information in order to support a previously held belief and to neglect other information. When complex policies are evaluated, this second-order mechanism distorts the interpretation of data or other observations, the analysis of situations or the appraisal of instruments independently of cognitive capabilities or seniority (World Bank, 2015: 183). Identifying the biases of policy professionals and decision-makers might be an especially important contribution to policy design.

After more than a decade of debating and implementing behavioural insights and interventions however, we are now starting to learn more about the limits and biases of behavioural public policy itself. While the number of studies on successful behavioural interventions continuously increases, scholars are reporting unintended side effects, implementation problems with interventions working in one context but not in another one and other forms of policy failures associated with behavioural public policy (House of Lords, 2011; John, 2018; Sunstein, 2017; Weaver, 2016). As John (2018: 96–97) has argued, the current focus on ‘what works’ might be misguided: ‘a mature research programme is not dependent on continual successes but needs to learn from failure as part of a genuine desire for reform’. Taking this call for a more critical analysis as a starting point, this paper aims at reassessing and rebuilding some of the conceptual and methodological foundations of the behavioural change agenda in order to understand and, if possible, prevent behavioural policy failures. From a mechanism-based perspective, scientific knowledge and its use for public policy design expands by adding new or formerly ignored mechanisms to the toolbox of already known mechanisms and by showing how they relate to each other (Hedström and Ylikoski, 2010: 61). In a similar vein, it is being argued that behavioural public policy could benefit from systematically linking a specific set of already well-known but often ignored mechanisms on different levels of policy-making that might help to understand the observed failures and side effects. In many ways, research on and practices of behavioural public policy need to overcome a tendency to focus on isolated mechanisms without taking into account coexisting mechanisms on the same or different levels that might interfere with them and therefore inhibit, impede or distort the expected outcome. This paper proposes to link mechanisms on three levels:

Linking first-order mechanisms related to system 1 and system 2: Taking into account the interdependencies of reflexive (system 2) and automatic (system 1) mechanisms on the cognitive level could help to understand problems with reactance and other behavioural failures. It could also support the design of ‘thought-provoking’ nudges that motivate people to act on their own behalf (John, 2018). Future behavioural public policy needs to overcome the sharp separation between the two.

To develop these three assumptions further, the paper presents a critical re-evaluation of the current literature on behavioural mechanisms based on a broad range of disciplines such as public policy research, political sociology, cognitive and behavioural science and science and technology studies. It is structured as follows: the next section gives a brief conceptual introduction to the key terms and assumptions in analysing behavioural public policy failures. ‘Linking system 1 and system 2 mechanisms’ section explores the cognitive mechanisms that lie at the heart of the dual process framework developed by the ‘heuristics and biases’ school and problematizes the sharp splitting between both systems. ‘Linking individual- and collective-level mechanisms’ section discusses selected findings on the interference between cognitive mechanisms and collective-level mechanisms in groups and cultural contexts and elaborates on the implications for policy design. ‘Linking policy and expertise’ section brings together insights on biases resulting from the interlinkages between decision-making and scientific expertise. Finally, ‘Behavioural mechanisms: Lessons for policy design’ section summarizes the findings and draws some conclusions for the design of behavioural public policies.

Understanding behavioural public policy failures

In many respects, the rich and complex literature on public policy failures has benefited from behavioural sciences (for an overview on this literature, see Bovens and ‘t Hart, 2016; Dunlop, 2017; Howlett, 2012; Nair and Howlett, 2017). Even in the face of pressing problems such as climate change, policies may not be able to generate an adequate response by citizens because of cognitive failures to perceive risks (Kemmerling and Makszin, 2018). In consumer policy, more information on the contents of food or the nutritional value might increase the probability of health-damaging behaviour (Reisch and Zhao, 2017). Policy-makers and professionals themselves tend to under- or overestimate risks, overreact to certain issues or problems while ignoring others or interpret evidence in line with their own beliefs (BIT, 2018; Bovens and ‘t Hart, 2016; Dunlop, 2017).

According to a widespread definition, a policy fails ‘if it does not fundamentally achieve the goals that proponents set out to achieve, and opposition is great and/or support is virtually non-existent’ (McConnel, 2011: 221). Behavioural sciences, however, also describe conditions under which a policy does not achieve the goals that proponents set out to achieve even if opposition is low and support is high. Two insights are key for this observation (Kahneman, 2011; Thaler and Sunstein, 2008):

First, when situations are complex or ambiguous the behaviour of individuals may be influenced by cognitive heuristics (Kahneman and Frederick, 2002, 2005 Kahneman and Tversky, 1982 [1974]). In contrast to standard models of rational choice, people sometimes use mental shortcuts and simple solutions even if this means acting against their interests as stated by themselves. They may ignore the fact that they do not know enough or fail to learn from new information. In the aftermath of an earthquake, more people are likely to purchase insurances against this unlikely event (Kahneman and Tversky, 1982 [1974]). Heuristics shape the ways risks are perceived. More information rarely solves the problem because it adds to the already existing amount of information, increases uncertainty and tends to trigger the use of simple heuristics (Thaler and Sunstein, 2008).

Second, the informational environment against which people make decisions may trigger some of these heuristics and inhibit others (Thaler and Sunstein, 2008). The way books in a bookstore are being presented or the order of the food on the menu in a restaurant makes certain decisions easier than others. Everyday action is embedded into informational infrastructures, simplifying the presentation of options, evoking certain associations or making certain options more visible than others. Thaler and Sunstein (2008) speak of ‘choice architectures’. According to the authors, choice architectures are ubiquitous. Since choice architectures are inevitable, policy-makers are advised to actively engage in designing arrangements that support desirable policy goals and reduce behaviour seen as suboptimal.

Behavioural public policy is based on such insights from behavioural economics, behavioural sciences, psychology or neurosciences and includes all means and modes of public policy aiming at influencing individual or collective behaviour (John, 2018; Oliver, 2015; Straßheim and Beck, 2019). The spectrum of policy instruments is large, including efforts of political and administrative simplification to reduce the cognitive burden on citizens, education programmes for decisions under the conditions of risk and insecurity, techniques of social norms marketing and behaviourally informed regulation to prevent industry from manipulating consumers (Oliver, 2015; World Bank, 2015). Nudges are arguably the most prominent subtype of behavioural public policy (Thaler and Sunstein, 2008).

Over the past decade, the debate on behavioural public policy has highlighted both advantages and challenges of behavioural approaches (for an overview, see Straßheim and Beck, forthcoming). While arguing that behavioural insights and interventions can improve the design and implementation of development policies, more and more authors also point to failures of behavioural public policy: multiple cognitive, socio-cultural, institutional and policy-related factors have to be taken into account to make behavioural public policy work (House of Lords, 2011; John, 2018; Sunstein, 2017; Weaver, 2016). The uncertainties and unintended side effects related to these factors are not fully understood. For example, a Swiss–US study on a behavioural energy conservation campaign found that successfully reducing water consumption by giving people feedback on their individual water use coincided with an overall increase in electricity. An explanation for this unforeseen outcome is what researchers call the ‘moral licensing effect’ – people who saved water simply felt entitled to be wasteful in another area (Tiefenbeck et al., 2013). Moral licensing is well known in many areas of behavioural governance such as obesity policies or fitness campaigns (Merritt et al., 2010). It is just one type of many confounding factors that intervene with the mechanisms triggered by behavioural interventions.

There are many more reasons why behavioural public policies could fail (for an overview on these and other examples, see Sunstein, 2017; Weaver, 2016). The compliance costs might be too high: A study in the United Kingdom found that a majority of people opted out of a savings plan with an unusually high default contribution rate. Essential infrastructures might not exist: Conditional cash transfer programmes that are supposed to change people's behaviour in public health or education work in many places. They might, however, fail where no schools or clinics exist or where people in local communities strongly oppose vaccination (Deaton and Cartwright, 2016). Especially the more recent research on compliance regimes and administrative mechanisms has started to systematically identify the barriers and constraints for behavioural change instruments (Lodge and Wegrich, 2016; Weaver, 2016). This research points to a number of problems, including insufficient or contradictory incentives, intervening cognitive effects such as myopia, hostility to government programmes and other value-based objections, herd effects and lack of autonomy. All in all, much more research is needed to develop a better understanding of the cognitive, socio-cultural and policy-related mechanisms influencing how behavioural instruments work. The complexity of these factors also explains the insistence on simplification as a core principle of behavioural public policy (Halpern and Mason, 2015; Sunstein, 2017): While nudges may work in carefully orchestrated and isolated settings, it is difficult to know how they behave in the real world (see Deaton and Cartwright, 2016 for a methodological critique).

Taking these findings together, many behavioural public policy failures seem to be the result of a deficit in understanding the links between multiple mechanisms on multiple levels (see Capano and Howlett in this Special Issue). From a mechanisms perspective, behavioural public policy failures occur whenever interfering mechanisms are poorly understood or even ignored. To get a better understanding of failures and to search for ways of preventing them, it might be helpful to distinguish between different types and levels of mechanisms.

First-order mechanisms such as cognitive heuristics and biases are triggered by nudges or other behavioural interventions to induce certain behaviour or to prevent the manipulation of consumers by third parties such as the industry. In behavioural public policy, mechanisms can be activated at the individual level by providing a heuristic that can be used as a substitute for more complex modes of judgement under uncertainty, for example by making some choices more salient or by setting a default. At the social level, interventions such as social norms marketing can highlight certain options and reorient herd behaviour. These different levels of behavioural change mechanisms interact with each other, as for example when nudges in energy policy change individual energy consumption, which can affect collective behaviour via herd effects and lead to an overall increase in sustainable behaviour (BIT, 2011; United Nations, 2017).

Second-order mechanisms are those mechanisms that are involved in feedback processes and other reflexive dynamics informing the use of behavioural interventions by observation of the reactions of individuals, groups and policy systems behaviour. Second-order mechanisms are promoting reflexive governance and include individual and collective activities such as policy learning, diffusion and transfer. They might, however, also include self-undermining mechanisms related to confirmation bias, group reinforcements or illusions of control that become negative second-order mechanisms whenever they are entangled in policy feedbacks and reflections. These self-undermining mechanisms may be triggered in complex, uncertain or highly politicized situations when policy-makers develop a bias while evaluating or re-regulating the use of first-order mechanisms (BIT, 2018; Dunlop, 2017; Weaver, 2010). While first-order mechanisms have direct effects on individual and collective behaviour, second-order mechanisms influence the way policies are designed, appraised, diffused or changed.

The following argumentation is based on the assumption that failures associated with behavioural public policy are the result of a deficit in linking mechanisms on at least three different levels: First, linking first-order mechanisms related to system 1 and system 2 might help to understand problems caused by value-based objections, reactance and other behavioural failures. Second, linking first-order mechanisms on the individual and collective level could lead to a better understanding of how individual behaviours are interlinked with collective behaviour in groups and policy subsystems. Third, linking the behaviour of government professionals and scientific experts could help to identify those second-order mechanisms that explain why both actor groups reinforce each other not only in terms of learning and knowledge production but also in terms of collective myopia, ignorance and other pathologies.

Linking system 1 and system 2 mechanisms

Some behavioural public policy failures could be related to a lack of understanding the mechanisms linking both the automatic and intuitive processes of system 1 and the controlled and reflective processes of system 2. In many instances, nudges seem to intervene in complex dynamics of iterated shifting between system 1 and system 2 without doing justice to the complexities of individual judgements.

Kahneman and Frederick (2002) have argued that ‘attribute substitution’ is the basic cognitive mechanism underlying most heuristics. Using a dual process model perspective, they assume that cognitive mechanisms either belong to system 1 or system 2. Especially under conditions of complexity, the attributes of a specific object of judgement are substituted by properties of an attribute that comes easier to mind. In these cases, system 1 takes over. While system 2 is based on controlled and serial cognitive operations, system 1 automatically produces quick answers.

Over the past 30 years nearly 30 variants of dual process models have been developed and hundreds of other heuristics have been described (Fiedler and Von Sydow, 2015; Kahneman, 2011). While Kahneman and Tversky (1982 [1974]) have continuously emphasized the importance of system 1 and its potential for adaptability in complex contexts, the main focus of this research was on the errors and biases system 1 produces (Kahneman and Frederick, 2002: 52). Nudging and related policy instruments of behavioural change are designed for targeting system 1 mechanisms either by inhibiting mechanisms and avoiding errors made by system 1 (first-degree nudges) or by actively triggering these heuristic mechanisms so as to change behaviour in a targeted direction (second-degree nudges) (see Baldwin, 2014). The heuristics and biases approach with its dual process model is the core concept on which nudging and other behavioural change strategies are based (for an overview, see Taranu and Verbeeck, 2016).

Based on these observations, the heuristics and biases approach and the dual process model have also been met with scepticism (Bago and Neys, 2017; Fiedler and Von Sydow, 2015; Gigerenzer and Gaissmaier, 2011; Lizardo et al., 2016). The debate sheds some light on potential problems that could arise from too sharp a separation between system 1 and system 2:

First, the theoretical and empirical status of heuristics is highly unclear. Even the most prominent heuristics as described above seem to lack the conceptual clarity and the empirical evidence that could directly connect them to a specific cognitive system (Gigerenzer and Goldstein, 1996: 650–669). Recent research shows that the distinction between system 1 and system 2 is more fluid, causing a mutual interference and iteration between habitual behaviour and reflective, deliberate action (Lizardo et al., 2016). In this sense, many nudges seem to trigger not only a shift from system 2 to system 1, but also back to system 2 that is (sometimes briefly) invoked in reflections on the consequences of system 1 actions (John, 2018: 127–134; Mols et al., 2015). When designing nudges, these effects need to be taken into account to avoid reactance or unintended consequences in the long run (Lepenies and Malecka, 2015).

Second, a related argument concerns the underlying assumptions about rationality that characterize the dual process model put forward by the heuristics and biases programme (Gigerenzer and Gaissmaier, 2011). Proponents of the ‘fast and frugal’ school criticise that system 1 processes have been routinely and nearly automatically associated with biases and limits in rationality. The general message underlying the heuristics and biases programme, they argue, is that people are better off if they do not rely too much on what comes easily to mind. Over the past two decades however, Gigerenzer (2002) and Gigerenzer and Goldstein (1996) have constantly pointed out that this interpretation provides only a limited understanding of system 1. Under specific conditions heuristics do not lead to judgemental errors. Instead, they seem to provide an ‘adaptive toolbox’ that helps decision-makers to cope in uncertain environments (Gigerenzer, 2002; see also Kelman, 2011 on the two schools and Sent, 2004). ‘Debiasing’ people (Kahneman and Tversky, 1982) might lead to a loss of such adaptive capacities.

Taken together, these findings from behavioural economics and cognition science provide a more nuanced picture of the dual process model. They suggest that more research should be invested in analysing the links between the mechanisms of the two systems. This is very much in line with social theories on behaviour, action and agency (Giddens, 1984; Schütz, 1959). These theories distinguish behaviour and action. Behaviour is conceptualized as an ongoing, more or less continuous flow of conduct based on certainties taken for granted until further notice. In contrast, action is understood as the competence of actors to anticipate a future time when certain conducts have been accomplished and the resulting act will already have been materialized (Schütz, 1959: 85). In the co-presence of others, this capability to reflect upon one's own action by anticipating future and monitoring present actions interrupts the continuous flow of behaviour and forms the foundation for knowledgeability and social interaction (Giddens, 1984: 190). In other words, system 1 and system 2 are related to each other by mechanisms of self-monitoring and reflection.

This shift from isolated behavioural mechanisms to iterative mechanisms of behaviour and action, from a dual system to an interactive system perspective is highly plausible from both a cognition science as well as a social theory perspective (see Kahneman, 2011). It has practical consequences for the design of behavioural public policy: John (2018: 127–134) has recently argued that even in the most impulsive moment of behaviour people tend to activate other parts of their cognitive system as well. Many nudges that are supposed to change the choice environment to activate system 1 heuristics also always activate a component of system 2. Following social norms to change health behaviour ‘often require that the people in the trial have gone through a thought process about how the desired actions will affect them’ (John, 2018: 128). To induce long-term behaviour changes and to avoid unintended side effects, nudges need to be ‘thought provoking’ (John, 2018: 129). They should stimulate peoples’ interest and engagement with specific problems by making visible alternative ways of action and by appealing to citizen's identity. Helping young people to slow down and think about future consequences of action or boosting patient's perception of treatment statistics requires some understanding of the linkage between system 1 and system 2 (Grüne-Yanoff and Hertwig, 2016; John, 2018: 130–131; Mols et al., 2015). This, of course, would mean that designing public policy is less about behaviour change and more about mechanisms of motivating certain action.

Linking individual- and collective-level mechanisms

Related to the critical debate on the linkage between system 1 and system 2 mechanisms, there are recent findings showing that dual process models are in need of a more systematic understanding of social and cultural factors that might interfere with the cognitive processes (Fiedler and Von Sydow, 2015; Le Mens and Denrell, 2011). The general argument is that the characteristics of the two systems described by the heuristics and biases approach are not as highly correlated, as one would expect. On the contrary, these studies show that even judgements clearly following the deliberate and reflective system 2 processes exhibit strong biases very similar to classical heuristics (e.g. representativeness or anchoring). This happens when collective mechanisms are influencing individual cognitive mechanisms: in-group bias, for example, can be the outcome of a highly rational information gathering and reflecting strategy; in such cases, the bias can be the result of a pre-structuring of information by a peer group (e.g. white freshmen in colleges), leading to stereotypical appraisals of other people even if additional information is given (Le Mens and Denrell, 2011).

Other findings point to a second set of factors that might influence how behavioural governance works (Wang et al., 2017; World Bank, 2015: 67–97). Anthropologists and ethnographers have made clear that cultural contexts have an effect on the formation and expression of heuristics and biases. For example, loss aversion is highly relevant for the design of nudges in pension policy (Thaler and Bernatzi, 2004). The degree and impact of loss aversion, however, varies depending on cultural effects. Even when controlling for economic demographic factors, substantial cross-cultural differences exist (Wang et al., 2017: 278–279). When combining similar countries into clusters, between-cluster variation is large. In Eastern Europe, loss aversion seems to be larger than in Anglo-American countries. Gender differences and the dominance of masculinity in cultures are also significantly influencing risk preferences. Moreover, in countries with high asymmetries in terms of power, wealth and social status, average individuals feel more helpless and could be more pessimistic when confronted with losses (Wang et al., 2017).

These cultural mechanisms are relevant for the way nudges and other behavioural public policies need to be designed to be transferable to different communities and jurisdictions without unintended side effects. Of course, these results are also highly relevant for public policy evaluation: As already mentioned above, Deaton and Cartwright (2016) have criticised the blindness of randomized control trials (RCTs) for cultural, institutional and other social contexts and recommended a broader range of empirical strategies. Some of the observed failures and side effects of nudging and other behavioural change strategies, for example objections based on group values, are the direct outcome of the methodological and conceptual shortcomings of an undersocialized perspective on behavioural change mechanisms. This means that behavioural interventions targeting individual and collective mechanisms need to be combined.

Group pressure and ‘herd mentality’ can be tackled by social norms marketing or techniques of ‘peer education’ targeting system 2 contexts (Mols et al., 2015). Interventions to change ‘mental models’ or establish new ones focus on recalibrating group-related information selection (World Bank, 2015). ‘Think’ strategies could establish institutional spaces for citizen engagement as ‘safe havens’ for consultation, deliberation and re-examination of existing institutions (John et al., 2009). In this sense, the design of behavioural change strategies is always confronted with both challenges, the integration of the cognitive and social context and the integration of system 1 and system 2 mechanisms. There is, however, a third challenge that concerns policy-makers and behavioural experts themselves.

Linking policy and expertise

The research carried out by Kahneman and Tversky was originally motivated by the question as to how the intuition of experts and professionals works (Kahneman and Tversky, 2011). This question, however, seems to have faded away in the debate about nudging. In their reports on the role of behavioural insights in development policy, both the World Bank (2015) and the United Nations (2016) have shifted the attention back to this question: What are the cognitive limitations and biases of experts and decision-makers? How do we nudge the nudgers? In a recent article, Lodge and Wegrich (2016) call this the ‘rationality paradox of nudging’: ‘[A]t the heart of nudge is a basic paradox: it assumes bounded rationality, but offers a comprehensive vision of rationality to address problems caused by bounded rationality.’ (253). In their analysis, Lodge and Wegrich identify the mechanisms that shape the bounded rationality contexts of behavioural public policies. While nudge and other behavioural public policies are aware of the cognitive limitations of citizens, proponents seem to be too optimistic on the decision-making capacities of nudgers themselves. As a result of multiple constraints in governments and administrations in terms of institutional structures and group pressure, biases in policy judgment such as ‘overcommitment’ or ‘oversimplification’ may lead to the failure of nudging as a policy tool: ‘nudge is not sufficiently reflective of its own limitations’ (Lodge and Wegrich, 2016: 263).

Similarly, the most recent report by the Behavioural Insights Team on ‘behavioural government’ argues that elected and unelected officials are themselves influenced by the same heuristics and biases they try to address (BIT, 2018: 7–13). Focusing on core activities of governments, the report identifies a multiplicity of decision-making biases (Bellé et al., 2018; Bovens and ‘t Hart, 2016; Sheffer et al., 2018):

Politicians and civil servants choose risky policies depending on how problems are presented (framing effects);

The BIT (2018: 11–12) also suggests strategies to overcome these biases such as transparency about the evidence base used, building networks to access expert advice and insight, assembling of teams that are cognitively diverse or integrating experimental trials into policy execution ‘wherever possible’.

The call for an even more systematic integration of experimental evidence, however, seems to be also somewhat overoptimistic as ‘gold standards’ RCTs have become the universal currency for evaluating policies. The global wave of experimental trials managed by organizations such as the Behavioural Insights Team (BIT) in the UK or the Abdul Latif Jameel Poverty Action Lab in the US has already raised some criticism (Pearce and Raman, 2014). RCTs create a controlled environment in which the social and cultural embeddedness of dual systems processes is suppressed in favour of internal validity. Or, in the pointed judgement made by Deaton and Cartwright (2016):

RCTs are not capable of telling us what would happen if these policies were implemented to scale, of capturing unintended consequences that typically cannot be included in the protocols, or of modelling what will happen if schemes are implemented by governments […]. (61)

In complex policy-making environments, these second-order mechanisms can lead to a de-politicized and de-contextualized view on policy-making that emerges as the result of combined biases in science–policy interactions. Identifying the mechanisms behind this specific form of science–policy ‘groupthink mechanism’ (Wellstead et al., 2018) seems to be vital for policy learning. Transdisciplinary training courses designed to inform policy-makers and experts alike about potential pathologies, questioning the separation between values and facts and allowing the collective exploration of behavioural public policy failures could be a way to overcome these biases (Dunlop and Radaelli, 2015).

Behavioural mechanisms: Lessons for policy design

It may sometimes be tempting to think that both small and simple behavioural interventions can create large effects (Halpern and Mason, 2015). Yet, this paper has argued that the multiple linkages between mechanisms of public policy-making require attention. These mechanisms may have been overlooked because they vary in different contexts and across different levels and because some instruments in policy evaluation are not sensitive enough. Getting a better understanding of them, however, may help to explain serious failures in behavioural public policy. It may also provide some lessons for public policy design. First, it could be important to shift the focus from behaviour to action. Even in the most spontaneous moments, people tend to reflect on what they are doing. To induce long-term changes, behavioural interventions need to stimulate people's reflection. Designing ‘thought-provoking’ nudges (John, 2018) requires a more interdisciplinary understanding of human action. Second, many problems of compliance with behavioural interventions are related to mechanisms on the level of groups and cultural contexts. Behavioural public policy needs to break from the microfocus proposed by behavioural economics and pay more attention to collective and cultural mechanisms. Some of the failures and side effects of nudging and other behavioural change strategies are a direct outcome of the methodological and conceptual shortcomings of an undersocialized perspective. Third, both experts and policy-makers are prone to cognitive biases. Even more problematic, however, are those mechanisms leading to collective pathologies in the relationship between science and policy. RCTs and the quest for ever more robust and value-free evidence are one expression of these pathological mechanisms in the science–policy nexus. Finding ways to overcome such biases and to develop ‘regulatory humility’ (Dunlop and Radaelli, 2015) seems to be more important than ever. It would indeed be ironic if in the quest for more rational behaviour both experts and policy-makers failed because a collective illusion of control has finally taken over.