What could neuroeconomics add to economics?
A methodological discussion and insights from research on addictive
behaviour
Floor de Jager 21 August 2014 Master Thesis Economics
Floor de Jager 10456910
21 August 2014, 15 ECTS
MA-Thesis Economics. Specialization in Behavioural Economics and Game Theory University of Amsterdam, Faculty of Economics.
Supervisor: Prof. dr. F.A.A.M. van Winden word count: 12,273 excluding references
- Contents -
- 1. Introduction - ... - 4 -
- 2. Discussion on the potential of neuroeconomics - ... - 5 -
- 3. On choice prediction from neural data - ... - 9 -
3.1. Choice, valuation and neuroeconomics ... - 10 -
3.2. Complementing exogenous environmental factors with neuroeconomics ... - 11 -
3.3. Possible criticism on the Falk studies ... - 13 -
- 4. Addiction models and neuroeconomics - ... - 16 -
4.1. (Beyond) rational addiction ... - 16 -
4.2. Reinforcement and tolerance ... - 20 -
4.3. Instances of rational and non-rational behaviour ... - 21 -
4.4. Critique on and discussion of both models ... - 23 -
4.5. Choice, information and revealed preferences ... - 27 -
- 5. Conclusion - ... - 30 -
[ - 4 - ]
- 1. Introduction -
The last decade has seen the quick maturation of a new field of science born out of the interdiscipli-nary work of dedicated scientists. Anytime such a thing happens this is both fascinating and poten-tially very valuable. The notable progress that has been made in the understanding of the neurobiol-ogy of choice behaviour led to the emergence of the field of neuroeconomics (NE). A variety of dif-ferent disciplines are implicated, but the central sciences of this cross disciplinary area of study are economics, neuroscience and psychology. A high rate of new discoveries underpins the viability of the field as an independent and stand-alone field of investigation. A host of new and fascinating re-search is conducted and many interesting findings have resulted. It is the ultimate hope that these results lead to foundational contributions to economics, neuroscience and psychology. The purpose of this thesis is to evaluate the potential of neuroeconomics to contribute to economic theory.
For some it is apparent that any facet of economic decision making is also an aspect of eco-nomics. Through this lens the contribution of NE to economics is self-explanatory as it expands the set of empirical phenomena that economists can investigate. Upon scrutiny this viewpoint is not self explanatory however and firm but relevant criticism has been voiced. It is therefore that I will inves-tigate whether and how neuroeconomics can contribute to economic questions of interest.
To search for this potentiality, chapter 2 will start with the methodological debate which has taken place wherein both fervent proponents and adamant opponents of neuroeconomics’ relevancy to economics voice their opinion. As it turns out, it is no easy feat to accurately indicate the relevancy of the field for economics while not sticking to platitudes and overly hypothetical cases. Two specific ways in which neuroeconomics might be able to add to the science of economics are investigated. One way in which NE could help economists is by making neural (endogenous) data available for pre-dicting choice behaviour. This abstract goal is illustrated with research on smoking cessation in chap-ter 3. Neural data seems to be predictive of how successful smokers will be in quitting their behav-iour. Another way in which NE might prove valuable to economics is by helping with model building and selection. This is illustrated in chapter 4 by comparing two models which deal with addiction. The first one is a standard rational model which adheres to orthodox economics. The second is a model that already incorporates neuroeconomic results. Throughout these chapters I will discuss whether these findings are truly of any value to economics and how so.
Unlike the centipede that ended up in the ditch by thinking too much on which foot to put in front of which, I hope that thinking hard and meticulous not only keeps this thesis going, but will provide some important insight in the science of economics, that of neuroeconomics, and how the latter might be of value to the former.
[ - 5 - ]
- 2. Discussion on the potential of neuroeconomics -
The field of neuroeconomics has sparked much interest and its rise let to a flood of articles with sur-prising and new findings on the border of economics, neuroscience and behaviour. Results such as the computation of utility in the brain (e.g. Glimcher 2003, Glimcher, Dorris & Bayer 2005 and Knutson, Taylor, Kaufman, Peterson & Glover 2005) are very exciting. Since the concept of utility is at the core of economic theory since the very beginning of the economic profession, finding that this concept is hard-wired into the brain naturally seems important and revolutionary. Such findings then channelled a lot of time and money towards this endeavour. Publications in Nature and Science seemed to validate the relevancy and importance of the field and some went as far as to claim that NE’s findings would revolutionize economics (Camerer, Loewenstein & Prelec, 2005). Claims like these however need to be carefully scrutinized. The economic profession has decades of experience in relating environmental variables to choice behaviour, in modelling and prediction. Even a wave of new findings that complement or even contradict accepted concepts is not yet a reason to discard old models straight away.
With the advance of neuroeconomics, criticism of the field and of its methods arose. Refer-ring to the images of neural activity in fMRI scans: ‘[c]olourful diagrams, which mean nothing to economists, are presented as clear evidence’ (Rubinstein 2008, p. 486). The statistical methods em-ployed to tease out brain activation are complex and vulnerable to critique too: ‘there are so many ways to go wrong’ (Savoy 2005, p. 362). This (methodological) discussion was held on conferences and in special issues in journals such as that in Economics and Philosophy (2008) or in the issue titled
Neuroeconomics: hype or hope In the Journal of Economic Methodology (2010). Acknowledging the
possibility of rhetorical excess, the early protagonist Camerer ceded that: [t]hese early neu-roeconomics papers should be read as if they are speculative grant proposals’ (Camerer 2008a, p. 44). To make the step towards a constructive use of neuroeconomics, it is worthwhile to take a look at a specific point of criticism towards it and how that came about.
A landmark article in the development of the field of neuroeconomics and one of the most outspoken and hopeful articles was Neuroeconomics: How neuroscience can inform economics by Camerer et al. (2005). By providing ‘an entirely new set of constructs to underlie economic decision making’, the authors speculate on the potential of NE to replace the as-if assumption1 of mainstream economics with a theory founded upon empirical findings on the working of the brain (p. 9-10). By referring to experiments that seem to unveil a literal computation of utility in the brain neu-roeconomists expected economists to embrace the field.
1
[ - 6 - ]
First, it substantiates the most fundamental assumption in economics. Because we now know that utility functions are real physiologic entities, the utility calculations that people were as-sumed to do really happen in brain. (Park & Zak 2007, p.50)
or
Given how interested these early economists were in direct measurement – when they could not readily do so – one might think that their intellectual descendants would be at least as in-terested when tools like those they fantasized about are actually available. (Camerer 2008b, p.370)
It is worthwhile to take a slight detour here and look at this concept of utility. To appreciate the criti-cism it is important to understand what different conceptions of utility there are as the most outspo-ken criticism towards NE relies on a theory where utility is a technical construct derived from obser-vation, instead of the physically existing process or entity Camerer, Park and Zak and others believe it to be. The direct measurement Camerer is referring to was speculated upon by Edgeworth in the second half of the 19th century. Different approaches by Fisher and Edgeworth are nicely presented in the analysis of the history of economic thought by David Colander (2007). Both Fisher and Edge-worth were concerned with the measurement of utility. Foreshadowing revealed preference theory, Fisher’s proposal was to do so indirectly by deriving utility from the observation of individual choice. In a nutshell this is what revealed preference theory is about, choices reveal preferences. Edgeworth however speculated about the development of a ‘hedonimeter’ which would in principal be capable of directly measuring utility. Edgeworth thus saw utility as an existing mental state whereby utility is substantive, consisting of an actual state of being. This interest in a substantive hedonistic (i.e. con-cerning the mental state) account of utility was pretty much abandoned throughout the 1930’s and replaced with something much closer to what Fisher advocated. It all but completely vanished with Samuelson’s work on revealed preference theory after which utility was largely (merely) defined through observed choices. It is now an ordinal concept which is able to rank different utility levels but there is no sense in thinking of utility in an absolute sense as one would like when it were to rep-resent welfare for example. This meant that utility is by now a technical concept without any inher-ent meaning, without substance beyond its technical role in ranking these preferences.
This snippet of utilities’ history and different approaches is interesting in a few respects. First of all it helps to discard what I will not be talking about, namely the substantive hedonistic account of utility as a measure of happiness or welfare as discussed in for example Dolan & Kahneman 2008, Frey 2010, Bruni & Porta 2005, Easterlin 2006 and Layard 2006. Bernheim (2009, p. 29 - 33) provides
[ - 7 - ]
a comprehensive list of arguments as to why neuroeconomics will not offer such a measurement anytime soon. As Camerer illustrates by referring to Edgeworth in the quote above, it is easy to mix the two approaches. Most relevant for now however is that NE’s claim of revolutionizing economics based on the NE findings on utility and the claim of the replacement of the as-if assumption, can be an easy focal point of criticism on the relevancy of NE in general.
The most explicit objection to a physiological grounding of economics, or a synthesis of eco-nomics and neuroscience for that matter, came from Frank Gul and Wolfgang Pesendorfer in The
Case for Mindless Economics (2008). In their provocative yet thorough paper, the authors discard the
possibility of any role of neuroeconomics for the economic profession except for that as an inspira-tion like any other science, phenomenon or event. Since ‘economic models make no predicinspira-tions about physiological processes that underlie decision making’ (p. 21) they maintain that many argu-ments articulated by neuroeconomists are beside the point. It is the account of utility as discussed above that they discard.
Their argument against the relevance of neuroeconomic findings for economics in general is very basic. They argue that for economics the focus should be on the decisions made and not on the process of decision making leading up to the ultimate choice. That would imply that only choice be-haviour, observable bebe-haviour, can be used for testing economic models and theories and that eco-nomics has no interest in the black box that produces the resulting choices. Since neuroscience is inherently about the process in the brain, any of their findings have no relevancy whatsoever to eco-nomics. This seems to put defenders of NE’s value for economics in a difficult position. Vromen (2010) however meticulously articulates why even orthodox economics2 does implicitly make some assumptions on the decision making process. Constructing utility and preferences requires a specific type of consistency in choice behaviour. As behavioural economics has demonstrated over and over again however, such consistency might not always exist in the real world. An irrelevant alternative might for example shift choices, or preferences can reverse over time. Economics therefore restricts the set of possible decision making processes it allows. By excluding some decision making processes, economics does make implicit assumptions on the decision making process. The strict separation of observed choices and the process leading up to that choice is an illusion. This leads Vromen to con-clude that ‘the gap between economics and neuroscience is not as wide as Gul and Pesendorfer want us to believe’ (p. 172).
Gul and Pesendorfer reject NE’s relevancy while others deem it to be self-explanatory. Doug-las Bernheim takes the middle road in On the Potential of Neuroeconomics: A Critical (but Hopeful)
Appraisal (2009). In this paper he examines the multiple ways in which NE might be able to
2 Please note that I will use rational choice theory, revealed preference theory and orthodox economics
[ - 8 - ]
ute to economic theory. Two of the uses of neuroeconomics he provides will be central to this thesis. The first possibility is that ‘neuroeconomics might redraw the boundary between the set of variables that economists treat as observable (x), and those they treat as unobservable (ω)’ (p. 9). The next chapter will discuss research on smoking addiction and stopping behavior. There I will illustrate how previously unobservable variables can add predictive power to a model that does not include en-dogenous neural data. The second possible use of NE that this thesis engages with is that of model selection and guidance. Neuroeconomics ‘could conceivably generate suggestive findings that infor-mally guide the search for an appropriate empirical model in useful directions, leading to more rapid and effective identification of the best predictive relationship’ (p. 25). This contribution of NE to eco-nomics is illustrated in the fourth chapter.
[ - 9 - ]
- 3. On choice prediction from neural data -
As Bernheim recalls, one of the main promises (and to him, objectives) of neuroeconomics is to add to the available variables for modelling. These new variables can be used to (partially) explain or pos-sibly predict choice behaviour. Whereas the unobservable process inside the head is, in economics, often treated as mere noise in the environment which the error term would capture, NE provides the opportunity to reduce this noise and create an observable variable for otherwise standard analysis. That is, when endogenous neural data would predict choices above and beyond the readily observ-able exogenous environmental variobserv-ables. It is thus of interest to dig into this ‘prediction-from-the-brain’. This chapter will focus on how neural data from fMRI scans seems predictive of whether cur-rent smokers will be successful in quit doing so. Before this however, it is instructive to quickly gloss over the most important tool in current neuroeconomics and look at other studies that relate previ-ously unobservable data to choice behaviour.
When a subject is asked to perform a task, any task, the brain exerts effort which requires amongst others oxygen. It is delivered by blood flows containing this oxygen. Since different areas of the brain are engaged depending on the task at hand, different blood flows with different levels of oxygen run within the brain at any moment. These blood flows and their differing oxygen levels can be differentiated between in magnetic resonance imaging (MRI) scanners due to the differing mag-netic susceptibility of oxygenated blood vis-à-vis blood with less oxygen. With a spatial resolution of about 1 millimetre and upwards, a 3D picture of the brain can be created consisting of tiny cubes called voxels. Changes in oxygen levels occur on a time scale of seconds so the combination of the temporal and spatial resolution gives a near real time picture of brain areas involved in the processes and tasks a subject carries out. This MRI scanning combined with brain activity and changing oxygen levels is called functional magnetic resonance imaging (fMRI). The constructed intensity of activity of the respective brain region is called the blood oxygen level dependent (BOLD) signal. It is useful to keep in mind that the spatial and temporal resolution has a limit however. Very short timeframes (about less than 1 or 2 seconds) or very small involved areas (less than 1 mm) cannot be differenti-ated between. Future research might run into these limitations as finer grained pictures are needed to get to an ever increasing understanding of the exact mechanisms that occur in the brain. For now we can ignore these limitations however. Functional MRI has become the dominant technique in current neuroscience practice as it has resulted in many interesting findings as we will soon see. Be-fore we turn to NE and smoking cessation, let us look at some related and supporting findings that involve choices, valuation and brain scanning. This research lays a foundation to interpret and appre-ciate the findings later on.
[ - 10 - ]
3.1. Choice, valuation and neuroeconomics
The first thing I would like to elaborate on is a potential criticism that could be applied to much neu-roeconomic research. Due to various (practical) limitations, neuneu-roeconomic experiments often offer hypothetical choices instead of real choices. Typical examples would be choices over gambles with high stakes, or choices involving consequences that fall into a distant future such as for savings. High rewards are usually not feasible in academic experiments with limited resources and for studies over longer time frames the subjects and their choices are much harder to control (e.g. get subjects to return to the laboratory). One might object that hypothetical choices differ in a fundamental way from real choices as they do not seem to carry the same weight. A hypothetical choice is more like an intention whereas a real choice has real, unavoidable consequences. It seems common sense that the former is likely to be less thoroughly thought through and probably more whimsical. Since no real costs or benefits are associated with making the hypothetical choice, there is no way to really check if this is a ‘true’ preference (if such a thing exists). The discussion on preferences will be further ex-plored in section 4.5. For now it is reassuring and important to see that Kang, Rangel, Camus and Camerer (2011) find that although there is stronger activation of the common valuation areas in the brain during real choices, the same brain areas are involved for both hypothetical and real choices. This indicates that the same mechanism is at play during both types of choices and this makes ex-trapolating from observed hypothetical choices to real world choices much more credible. This find-ing allows me to not dwell on this potential objection in the rest of this thesis. Hypothetical choices are an acceptable instrument and experiments employing them do not need to be automatically discarded.
It is interesting to see that valuation of the different components in a subject’s environment seems to be an automatic and ongoing process irrespective of whether the decision maker is focus-ing on the aspect which is under evaluation (Lebreton, Jorge, Michel, Thirion & Pessiglione, 2009). Lebreton et al. find that even when subjects are distracted, valuation in terms of pleasantness is an ongoing process which can be used to predict preferences. When subjects look at faces, houses and paintings and are asked to evaluate their age, BOLD signals constructed during this evaluation can be used to predict preferences in terms of pleasantness of pairs of these items. So even though their conscious is focused on evaluating the age, the pleasantness of these pairs is still encoded for in the brain. Closely related is the research of Knutson, Rick, Wirnmer, Prelec & Loewenstein (2007), which shows that the neural data obtained during the valuation process which precedes an actual choice, can reliably predict these choices even after controlling for self-reported preferences.
[ - 11 - ]
Another potential criticism to overcome is that the subjects in fMRI studies are placed in an artificial choice environment while they are lying on their back in a noisy, narrow tunnel. This makes the gap (both conceptually and time wise) between the valuation and the actual choice, consider-able. It is therefore interesting to find that neural data during non-choice environments, where sub-ject are merely faced with different goods without the a priori awareness of these choices being pre-sented to them later on, is predictive of actual choices made outside the fMRI scanner (Smith, Bern-heim, Camerer & Rangel, 2013).
Still, as interesting as this research is, these findings would be countered by the objection of Gul and Pesendorfer (2009) that economists are (and should be) concerned with the relation of envi-ronmental variables and choice behaviour. A detailed understanding of the neural mechanisms be-hind choice behaviour is interesting in its own respect and undoubtedly has other applications too but it ads little to the traditional economic endeavour. Since we have easy access to the observed choices, little is gained by an understanding of the exact process inside the skull, so they would say.
3.2. Complementing exogenous environmental factors with neuroeconomics
This strict position held by Gul and Pesendorfer becomes harder to maintain when data obtained through neuroeconomic techniques can complement the available exogenous environmental factors. This previously unavailable data on neural processes can be used to better predict choice behaviour. Observed choices can be related to the available exogenous environmental factors that are available, but including neural data and parameters into a model could then for example be used to diminish the size of the remaining error term in regression analysis.
One such application would be a case where agents hold private information they would not readily share for strategic reasons. Wang, Spezio and Camerer (2010) conduct some fascinating re-search on this topic. They apply eye tracking and pupil dilation information in a strategic sender-receiver game where it is advantageous for the sender to inflate her earning prospects. She earns more if she is able to nudge the receiver into choosing an action that is not in the receivers’ best in-terest given the objectively given state of the game. The researchers find that they can elicit some of the private information senders hold through measuring the size of the pupil during the game and by tracking the area of focus of the senders’ eyes. When sending deceptive information the pupil dilates (and it dilates more the bigger the deception is) and the senders look disproportionally at the payoffs of the true state (and not that of the state of which she is convincing the receiver they are in). Incor-porating these variables in decision making can greatly benefit the receivers in their ability to predict the actual information senders hold. In a similar vein of research, Krajbich, Camerer, Ledyard and
[ - 12 - ]
Rangel (2009) use fMRI data to obtain subjects’ private valuation of a public good within a public goods game. These games often encounter the free-rider problem which can be understood as fol-lows: for any player, revealing her true valuation of the public good to the other players of the game is strategically disadvantageous. Players understate this valuation when they realize that their share of the costs increase with their publicly communicated value. The private valuation is not readily observable by any other means and so the other players need to rely on the stated valuations in-stead. Obtaining the true valuation of each player is thus of great value to other players. Krajbich et al. obtain information on this private valuation through using fMRI scans of players during their game. Of course there is still a long and bumpy road ahead, but when these methods and results could be carried over to the real world this would greatly aid policy makers in designing optimal pub-lic goods provision as they would now have access to previously unobtainable information which is critical for optimal decision making. The class of research I would like to study in detail in the re-mainder of this chapter likewise uses neural data to predict choice behaviour. They add to the pre-dictive strength of models which used to be limited to standard observable environmental factors.
The studies below combine fMRI scans with the complex issue of smoking cessation. They use the BOLD signal in response to anti-smoking messages to partially predict the consumption of cigarettes in the weeks and months following the scans. Smoking cessation is a complex issue where many physical and mental processes are at play but this stopping behaviour is predicted with greater accuracy by adding the BOLD signal to the model, than with just the historically accessible exogenous variables or even intrapersonal reflection by subjects themselves. Let me explain.
The central case in these studies is from Falk, Berkman, Whalen and Lieberman (2011) in which 28 subjects who were determined to quit smoking and had entered the American Lung
Asso-ciation’s Freedom From Smoking program were studied. All participants were heavy smokers and had
very high quitting intentions as measured by scoring more than 9 out of 10 on the Contemplation Ladder. This is an often used scale to assess the readiness of smokers to quit (Biener and Abrams 1991). The subjects were exposed to professional advertisements encouraging smoking cessation whilst being scanned in an fMRI scanner. What the researchers found was that the resulting BOLD signal within a region of the ventromedial prefrontal cortex (vmPFC) during these ads was negatively related to the smoking rate 1 month afterwards while controlling for the baseline level of smoking and quitting intentions. That is, subjects who displayed higher BOLD signals within this region during these advertisements were smoking significantly less after a month. Self-report variables used were: intentions to quit, self-efficacy (a psychological concept to capture the agent’s belief in fulfilling the task) and the ability to relate to the advertisements. Neural data on top of these self-report variables explained 20% more of the behaviour change (R2 = .351) than the variables on their own (R2 = .146). Adding the neural data thus more than doubles the original predictive power of the model. Such a
[ - 13 - ]
significant increase in predictive ability with data not otherwise accessible (not even to subjects themselves) seems a very interesting case to scrutinize. A critical inspection of the study’s validity and the usefulness of the results is called for.
3.3. Possible criticism on the Falk studies
A first thing to establish is whether the region of interest (ROI3) of Falk et al. (2011) is truly involved in the observed behaviour change, or whether this is a case of spurious causation. If one were to scan the entire brain with an fMRI during some tasks, certain areas will most likely be indicated to signifi-cantly relate to this task. Due to the sheer volume of the data and the used significance thresholds, this is however not informative but rather the result of probability gone wrong. Fortunately support-ing findsupport-ings in related studies where this ROI serves a prominent role such as Falk, Berkman, Mann, Harrison and Lieberman (2010), Chua et al. (2011) and Falk, Berkman and Lieberman (2012), indicate that there is a strong link between this ROI and subsequent choices or behaviour change. Since these studies interlink with Falk et al. (2011) I will dub these findings (including Chua et al.) the Falk studies for the remainder.
Further support for this ROI comes from Hare, Camerer and Rangel (2009) and Hare, Mal-maud and Rangel (2011) who find that the vmPFC is involved in issues of self-control. Bechara, Tranel and Damasio (2000) and Bechara et al. (2001) link diminished vmPFC activation and bilateral lesions in the vmPFC to decision-making deficits, notably towards future consequences and they link a dys-functional vmPFC to stimulant abuse. These papers further stress the causal role of the vmPFC in decision making. It would be premature however to equate the vmPFC with exercising self-control as this is a complicated process that includes amongst other the modulation of the vmPFC by the dorso-lateral prefrontal cortex. The finding of vmPFC activity’s predictive value on smoking cessation is certainly valuable and very interesting but the exact causal mechanism is yet to be discovered. That the vmPFC activity has a critical role in the behaviour and choices is quite clear however.
Besides questioning the involvement of the vmPFC, another valid objection to Falk et al. (2011) might be the plain fact that there could be other exogenous variables that would predict the variance in smoking cessation. The aforementioned studies did not control for a host of factors that might be causally involved in the regression analysis (e.g. age, number of years smoked, socioeco-nomic background, etc.). Most notably however, they fail to incorporate the average number of ciga-rettes consumed into the regression. All participants were heavy smokers, meaning that they smoked
3 The ROI is the vmPFC in this case. In neuroeconomics the ROI is that part of the brain that is a priori indicated
[ - 14 - ]
over ten cigarettes a day, seven days a week. However, this still permits for a fat tail where consider-able heterogeneity in the amount of cigarettes consumed can exist. We will later elaborate upon discount rates, smoking and impulsivity, but for now it suffices to note that Laporte, Karimova and Ferguson (2010) find that the strength of what they call the forward looking effect does seem posi-tively related to the number of cigarettes consumed. Subjects with higher discount rates smoked a larger number of cigarettes. Discounting the future more heavily makes the damaging effects of smoking less costly. From this it can be expected that subjects with higher discount rates are less likely to stop smoking. This indicates that the predictive strength of the vmPFC bold signal might be weakened by observing individual discount rates. Before economists are to be convinced of the added value of neural data in prediction, such easily observable variables should be ruled out. Since it is not clear whether discount rates could play the same role as fMRI scanning in the observed vari-ability in the Falk studies, I will take the results at face value. Sufficiently interesting point can be discussed while doing so.
The Falk studies seem to create some wiggle room for the role of NE within the economic profession in predicting choice behaviour, but still it could be asked what the relevance of this neural data is. Why would we need to predict behaviour from neural data when we could also just observe actual choice and use that data for subsequent studies? Surely the latter is cheaper than costly neuroimaging.
If the causal mechanism of addiction could be mapped (assuming that there would be a sin-gle constant mechanism underlying all addictive behaviour) this would help with guiding effective treatment and policy. An actual insight into the black box would surely be better than to rely on trial and error. It would allow for a faster feedback loop when developing treatment or policy. Let’s say you would need to wait a year to conclude whether someone successfully quit smoking after some intervention or not. When the neural mechanism at the outset of this intervention would be strongly predictive of the end result, you would no longer have to wait to observe the actual choice made. This faster feedback could make it much easier to device effective treatment then. For society as a whole this would be valuable too. Falk et al. (2012) for example use fMRI data to predict which ad-vertisements are most effective in aiding smoking cessation. Such a focus group approach using neu-ral data as a predictor gives much faster feedback on the effectiveness of policy than waiting for ac-tual behaviour to play out.
Another answer would be that variants of fMRI scanning would become cheap and ubiqui-tous. A cheap scan could then guide health care workers in what treatment is most effective for a certain individual. If the neural data shows that someone might benefit from extra help, faster and more effective treatment could be employed (quite possibly saving money over time). This beneficial effect is even greater for stronger/ more damaging drugs than cigarettes. Getting a finer grained
[ - 15 - ]
picture of the heterogeneity within a society could help save on health care expenses by foregoing ineffective treatment or a one size fits all approach. This answer is speculative but that is no reason to ignore it altogether. For cancer research for example, genetic information is more and more used to decide between treatment options. Specific chemotherapy is used depending on the genetic makeup of the patient. In a likewise manner specific neural data might point towards specific treat-ment options.
These answers show that there quite well could be a role of NE in prediction and the devel-opment of policy. At the same time they also show that this role is not self-explanatory. The next chapter will explore one of the other potential contributions of neuroeconomics that Bernheim high-lighted. It is closer to a real economic application and will illustrate how neuroscientific findings can help with model selection and guidance.
[ - 16 - ]
- 4. Addiction models and neuroeconomics -
At first sight (substance) addiction seems to pose a challenge to a theory of a rational, utility maxi-mizing agent. As we will see however, many features of addiction can be, and have been, modelled with fairly standard economic assumptions. Nonetheless, findings of research on addiction might challenge the standard economic paradigm of revealed preferences theory for which observed choices reveal true underlying preferences of a decision maker. By looking at two papers that try to model addiction, we will get a better understanding of how models adhering to economic rationality can account for addiction, how neuroeconomics might be able to improve upon these models and what difficulties arise. For these models I will mainly look at the assumptions and the characteristics of addiction they are able to explain. I will not go into detail on the differing policy implications of these models as that would require a much more extensive look at the inner workings of the models and this is also greatly dependent on the multiple assumptions that go into the models. Rather I will just look at how good a job they do on explaining addiction (and thus on predicting part of the behav-iour that addicts will display). These models might be used for prediction too but that would require much more evaluation and empirics than space constraints allow for here. After going over the mod-els I will pinpoint some of their flaws. The chapter ends with some digging at the assumptions under-lying the models. It thus provides a concrete case to study the role of neuroeconomics in model guidance and selection as proposed by Bernheim (2009).
4.1. (Beyond) rational addiction
In A Theory of Rational Addiction, Gary Becker and Kevin Murphy (1988) model addictive behaviour with fairly standard rational choice theory. Before the publication of their landmark study, addiction was regarded as an uneasy problem and mostly ignored. It was seen as irrational behaviour and ir-reconcilable with rational choice theory. They however, managed to incorporate several defining characteristics of addiction in an otherwise standard intertemporal decision making model. Their model (simplified here for focus and comprehension) assumes an intertemporal utility maximizing agent. Utility in any time period consists of two goods (c and y in their model) and on past consump-tion of c (thus this represents the addictive good):
[ - 17 - ]
Current utility depends on past consumption of c through the stock of ‘consumption capital’ which changes over time according to:
The stock depreciates with and also depends on expenditures on endogenous depreciation or ap-preciation . These elements taken together, together with a lifespan of T and a constant rate of time preference gives the following utility function:
This utility maximization faces a budget constraint of initial assets and the lifetime stock of capital where r is the interest rate:
Maximizing current utility subject to this budget constraint then gives a rational addiction model where the consumption of an addictive good is a voluntary, rational decision. For the full model and its properties see Becker and Murphy (1988, p. 676-678).
Sixteen years later, Douglas Bernheim and Antonio Rangel (2004) do their own modelling on choice behaviour of addicts. Whereas Becker and Murphy de facto rule out the possibility of mis-taken choices by (implicitly) adhering to classical revealed preference theory, Bernheim and Rangel explicitly use neuroeconomic findings to integrate mistaken choices into a rational decision making model. To do so it is necessary to reject the central tenet of revealed preference theory, namely that choices made always reveal an agent’s underlying preferences, by separating preferences from choice behaviour. A discussion on this follows in section 4.5. Since the model builds on neu-roeconomic findings, it is less familiar, abstract and more intricate than that of Becker and Murphy but it is important dig into its main building blocks. Let me explain them.
At the core of Bernheim and Rangel’s model are psychological and neuroscientific findings re-lating to habituation and agents’ resulting decision making. A large part of this process takes place through a neural mechanism which Bernheim and Rangel dub the hedonic forecasting mechanism (HFM). This mechanism spans multiple neural mechanisms and brain areas and it links cues in the environment of an agent to the predicted hedonic experiences (Schultz, Dayan & Montague 1997 and Schultz 1998, 2000). It thus involves predicting the future hedonic experience from information, signals and situations which the agent currently faces.
[ - 18 - ]
By acting directly on the forecasting mechanism or by rerouting this process and fulfilling a similar function, addictive substances can widen the gap between the forecast of the hedonic experi-ence and the actual hedonic experiexperi-ence. The mesolimbic dopamine system can be said to act like, or mimic such an HFM by producing a dopamine response when presented with a cue. The dopamine release following a cue creates a strong behavioural impulse. The strong behavioural impulse to seek a reward (following the habituation by drug usage) thus acts on the HFM, predicting a large reward that is not proportional to the real expected hedonic experience. It is in this context that neuroscien-tists Kent Berridge and Terry Robinson talk about the wanting process (resembling the reward pre-diction) and the liking process (consisting of the experience itself). The wanting can be thought of as a craving for something in the future. The liking process is the actual level of enjoyment experienced during the future consumption or event. See for example Berridge 1996, Berridge and Robinson 1998 and Robinson and Berridge 1993. This gap between the predicted and actual reward is why Bernheim and Rangel speak of forecasting errors and systematic mistakes. The following point illustrates this concept further (adapted from Bernheim & Rangel 2004, p1562). When an Englishman in New York looks right but fails to look left when crossing the road and is subsequently hit by a car, this choice can hardly be said to reveal his preferences. Through repeated reinforcement and learning, an Eng-lishman has gotten used to look right before crossing the road. When faced with changing circum-stances the resulting choice of only looking right does not reveal a true preference of the agent. If asked he would most likely state that his objective is to safely cross the street. So this example illus-trates that observed decision making cannot be unequivocally be linked to an agent’s preferences. In much the same manner, they content, our brain causes systematic errors. Even though the brain also exhibits a cognitive control function to balance impulses from the HFM, the impulse by the false pre-diction (of a high reward) is so strong that it becomes hard for this controlling function to override the created desire to act on the impulse.
The actual model build on top of this neural basis is fairly intricate and I will not go into great detail here. It broadly consists of the following. A rational decision maker (DM) starts out in a cold mode for which she makes a decision on her lifestyle activity .Different lifestyle activities repre-sent different probabilities of encountering a cue and reprerepre-sent different intrinsic enjoyment. After the lifestyle choice she divides her resources between an addictive substance ( ) and a non-addictive good ( ). When the DM is faced with a cue she is nudged into the hot mode. The probability of encountering a cue depends on lifestyle activity and an exogenous state of nature . This state of nature in turn is drawn with some probability measure (µ) from a state space (Ω). Past usage of addictive substances sensitizes a user through bringing her in a higher addic-tive state ( . Note that in state 0 an agent has never used the substance but this does not rule out that she might rationally choose to do so. In the hot mode, exercising cognitive control
[ - 19 - ]
cannot overcome the impulse to use generated by the distorted signal from the HFM: . The attractiveness of the drugs is represented by and as we can see this depends on the probability of encountering a cue, on the level of sensitization to the drug through past usage, on the lifestyle activity and on the exogenous state of nature. When the user is nudged in the hot mod she will always use, even if she (rationally) prefers not to do so. ‘Use among addicts is potentially a mistake; experience with an addictive substance sensitizes the user to envi-ronmental cues that subsequently trigger mistaken use’ (Bernheim & Rangel 2004, p1567). If she does not face such a cue, she can rationally decide on whether to use or abstain. The agent from the outset aims to maximize her hedonic payoff . This thus depends on the consumption of the addictive and non-addictive good, on the lifestyle activity and on the state dependent level of well being s (e.g. addicts may suffer from deteriorated health). She then maximizes a value function (containing more parameters than listed here) subject to her income and the price of the addictive and non-addictive good. The takeaway message is that Bernheim and Rangel present a model which incorporates stochastic shocks (µ) that can trigger mistaken usage of an addictive substance. When they set the probability of such stochastic shocks to 0 they end up with a variation of a purely ra-tional addiction model.
Further discussion on these models will follow below but for now I would like to point out one potential issue that economists might have with Bernheim and Rangel. Their model makes clear that through sensitization an agent might be nudged into a hot mode where preferences and choices might no longer align. But given the fact that an agent is forward looking she could expect such a nudge to occur with a certain probability. Given that the occurrence of the stochastic shock causes this nudge, she would be able to assign a probability function to the chance of this shock happening. In such a case the potentiality of becoming sensitized to a drug and being nudged into a mode where current preferences and choices might misalign at t>0 is just a choice made at t0. Is this really a
mis-taken choice then or a gamble that did not pay off? This is an issue that Bernheim and Rangel might have given more attention.
Setting aside this objection let us look at how both models can explain characteristics typical of addiction. They do differ in the breadth of phenomena they can explain so let me go through them to see how both models fair against each other and to other scientific findings on addiction, addicts and their behaviour.
[ - 20 - ]
4.2. Reinforcement and tolerance
A typical trait of addictive substances is that an increase in its current consumption, ceteris paribus, increases future consumption of it. Or in other words, current consumption is positively related to past consumption. This reinforcement or adjacent complementarity is almost a tautological defini-tion of addicdefini-tion but important to explicate nonetheless.
The model of Becker and Murphy is able to explain this through the investment like property of consuming an addictive good. With current consumption, future consumption might go up as the marginal utility of the addictive good is positively related to its past consumption (Becker & Murphy 1988, p681).
Another approach is offered by Bernheim and Rangel. In their model of addiction, a user is sensitized to the environmental cues that potentially trigger mistaken usage of the substance. Mis-taken here must be understood as misMis-taken according to the user’s own stated preferences. The cue acts on the HFM whereby the predicted reward increases. This sensitization makes a user more sen-sitive to the cue which could lead to actual usage of the substance.
Both models are thus able to explain the reinforcement effect of addiction. Even though the model of Bernheim and Rangel is grounded in a physical process, economists are likely to be equally happy with both models taken the as-if assumption of revealed preference theory into account4.
A second element of addiction that Becker and Murphy are able to capture is that of toler-ance or habituation whereby the utility of current consumption decreases as the level of past con-sumption goes up. This thus means that a higher level of current concon-sumption is needed vis-à-vis past consumption to create an equal level of utility, or that more consumption is needed to obtain the same hedonic experience. Rather simply, this is done through assuming that consuming an addic-tive good is negaaddic-tive for future utility, or that given the levels of consumption, utility is negaaddic-tively correlated to past consumption of the addictive substance. This in turn implies that current con-sumption of the good should increase to reach an equal level of utility. Use doth breed a habit indeed (Becker and Murphy 1988, p675).
As Bernheim and Rangel focus on the reward prediction, rather than the hedonic experience of substance abuse, they do not have much to say about (and do not model) the way in which sub-stance usage can cause tolerance to its effects. They simply accept that the hedonic experience di-minishes with past usage but this finding is not further incorporated in their model, nor does it influ-ence their results or the mechanisms they describe. It can be argued however (and they seem to do so themselves) that the hedonic experience is not central to the mechanism of addictive behaviour.
4
[ - 21 - ]
The fact that tolerance builds up is more of an effect of addiction than that it is a causal factor. It is a causal factor in so far that it might create increasing levels of consumption over time, but it does not cause a user to consume the addictive substance in subsequent periods, unlike reinforcement. That being said it is obviously an empiric phenomenon which is interesting to capture and the rational addiction model is able to do so whereas the neuroeconomic model of Bernheim and Rangel does not.
4.3. Instances of rational and non-rational behaviour
The biggest difference in assumptions and in the power of the models to explain real world observa-tions, lie in the way the concept of rationality is dealt with. Remember that the central claim of re-vealed preference theory is that agents act as-if they are fully rational because those assumptions seem to lead to models that are powerful in predicting consumer behaviour. However, as we will see some of the observed characteristics of addicts cannot be explained with the rational model alone. But let us take a closer look at rational behaviour first.
As Becker and Murphy apply rather straightforward rational choice theory in A Theory of
Ra-tional Addiction, they assume that decision makers have stable preferences over time and that the
choices they make reveal those preferences. This entails that preference orderings are preserved from t0 to t1, from t1 to t2 etcetera. In such a time-consistent model, consumers will maximize their
current utility which consists of the sum of discounted utility over all future time periods. This stan-dard intertemporal decision making imposes that consumers are fully rational and forward looking. This assumption might seem odd as few would label addiction as a rational action but evidence of forward looking behaviour of addicts can indeed be found.
In 1964 the U.S. Surgeon General (the head of the public health service) issued an influential report on health and smoking. This report unequivocally stated that smoking causes lung cancer. Despite earlier warnings and indications, hearing this causal link from America’s ‘top doctor’ could be considered new information for a large part of the population. Under rationality assumptions, a valid reason to modify your choice behaviour would be to obtain new information. Such a reduction was indeed found as a 34 to 45 per cent decrease in smoking was observed following the eleven years after the Surgeon General’s report (Ippolito, Murphy & Sant 1979). Clearly smokers changed their behaviour incorporating the new information, thus confirming the rational addiction model of Becker and Murphy. Another instance of rational behaviour can be found in anecdotal evidence of forward looking behaviour, even amongst serious heroin addicts. In The Fix (2000) Michael Massing describes how heroin users abstain from heroin usage and get clean with the intent to renew the original high
[ - 22 - ]
of the drug when returning to using it. Considering the serious (short term) withdrawal effects of abstaining from heroin use as an addict, this indicates a rather striking level of rational planning. Fu-ture hedonic rewards are weighed against short term physical and mental discomfort and the former is deemed to outweigh the latter. A final indication of forward looking behaviour can be found in Gruber and Koszegi (2001, p1293). In their paper they point towards evidence that current cigarette sales increases and current consumption decreases with (credibly) announced future price increases. To maximize utility over time it makes sense to stock up cigarettes before they increase in price. This behaviour convincingly points towards some degree of rational intertemporal decision making.
The model by Bernheim and Rangel too allows for rational behaviour as an agent in cold mode adheres to a value maximizing function which includes activities and their expected payoff (Bernheim & Rangel 2004, p1567). However, their model is further able to explain features of sub-stance abuse that undermine a purely rational theory of addiction. Without going into great detail let me list those characteristics and why their model can account for them occurring.
The first one is the observed contradiction between the stated preference to quit and not ac-tually succeeding in quitting. Note that this is a different situation from addicts who rationally enter rehabilitation to renew their high as the former users state an intention to follow through with ab-staining. Barring new information or physical restrictions, this cannot be explained under the RPT model. The usual solution is to discard the stated preferences as being cheap talk. Only choices re-veal an agent’s true preferences. A closely related second characteristic of addiction is that addicts often ascribe a loss of control over their addiction. Substance abuse is categorized as a mistake and not as a conscious decision. ‘[H]e was powerless to resist, even though he knew while it was happen-ing that it was a disastrous course of action for him’ (Goldstein 2001, p249). Bernheim and Rangel however, can account for this behaviour through their distinction of a cold mode and hot mode ad-dict. Under cold mode conditions an agent can state her preferences which she is unable to follow through on when encountering a cue that nudges her into a hot mode. There is something to be said for discarding cheap talk as relying solely on stated intentions might not reveal true underlying pref-erences either. An agent might have an interest in concealing these intentions. Yet this position can-not explain why someone states a sincere preference to abstain, starts rehab, experiences discom-forting (costly) withdrawal effects, only to relapse after having gone through most of the discomfort in the process. To me these phenomena need to be taken serious in any model then and not be ig-nored as cheap talk as fully rational models (implicitly) do.
The second characteristic Bernheim and Rangel stress is that of cue triggered recidivism. Given that no new information comes to light when you hang around old friends that are using, you should not change your behaviour. But environmental cues play an important role in craving and relapse (Robinson & Berridge 2003) and successful tactics for abstaining include avoiding cues that
[ - 23 - ]
trigger a relapse. Precomitment, which Bernheim and Rangel categorize as a separate characteristic, is the behaviour of an addict who acknowledges the fact that current preferences and future choices might not align. Such precomitment could for example consist of entering a rehab clinic to make sure no cues can appear. Since Bernheim and Rangel’s model has the cue triggered hot mode at its base, it is easy to see why this empirical observation and their model overlap. Becker and Murphy do ac-count for precomitment by entering rehab but they cannot acac-count for the empirical observation that it is best to avoid old environments in order to stay clean.
A final phenomenon of addiction is that successfully abstaining can be greatly aided by cogni-tive and behavioural therapy. In these therapies the agent can be taught to cope with cues and with the craving resulting from the skewed hedonic reward prediction. Cognitive therapy aims to create new associations and neural connections that severe the old link between the cues and the craving. Behavioural therapy can teach cue avoidance in order to stay in control. Based on the assumptions of the working of the HFM in the Bernheim and Rangel model it becomes clear why such therapy might help addicts to abstain from using. By gradually reducing the gap between the skewed predicted hedonic reward and the actual hedonic experience, an agent can regain control of this Pavlovian re-action.
The neuroeconomic model then seems to be able to explain more of the observed behaviour of addicts than does the revealed preference theory model. But so far I have been putting one model against the other to see how they compare. I will now try to apply some direct criticism to both mod-els and their assumptions to see how they fare in that respect.
4.4. Critique on and discussion of both models
Roughly two lines of critique can be applied to the models. One would be to find inconsistencies within the theory. The other would be to use other studies and findings to point to inconsistencies or lacunas in the model. I will employ both sorts of critique here.
The first challenge to the rational addiction model comes from literature which links decision making deficits to vmPFC activation which is not on par with healthy control subjects. Damage to the vmPFC has been associated with decision making deficits in real life settings involving rewards, pun-ishment and uncertainty. Specifically, decision makers with these lesions seem to largely ignore fu-ture consequences whether they were positive or negative (Bechara, Damasio & Damasio, 1994 and Bechara et al., 2000) and are instead much more motivated by short term prospects. Remember that the BOLD signal variance in the vmPFC in the Falk studies was successful in explaining a large portion of the actual reduction in smoking behaviour. Bechara et al. (2001) furthermore find that substance
[ - 24 - ]
dependent decision makers resemble patients with bilateral lesions of the vmPFC significantly closer than healthy subjects. That is, the average vmPFC of substance dependent decision makers closer resembles the vmPFC of patients with bilateral lesions than the average vmPFC of healthy subjects resembles these patients’ vmPFC. Mitchell (1999) uses both behavioural observations and personal-ity traits to show the relative impatience of smokers vis-à-vis those who never smoked. The ques-tionnaire for example asked about self-control and deference and the choice based behaviour tests showed steeper discounting rates for smokers. Though nicotine of course differs in its chemic prop-erties and in the exact effects, its neurologic effect can and has been linked to other addictive sub-stances (e.g. Corrigall, Franklin, Coen & Clarke, 1992, Hyman & Malenka, 2001 and Nestler & Malenka, 2004). Of course this leaves the question of causality wide open. Is addiction (partially) caused by a flawed vmPFC or is a flawed vmPFC (partially) responsible for addiction? The causal ef-fect of smoking on impatience is explored by Bickel, Odum and Madden (1999). They find that smok-ers are more impatient with regard to both never smoksmok-ers and ex-smoksmok-ers. Now it could be a selec-tion effect whereby ex-smokers have a different discount rate from the smokers irrespective of the current smoking. They are ex-smokers for a reason. Maybe only those smokers with a low discount rate are able to stop successfully. Standing up to that objection is the fact that ex-smokers are those that smoked at least 20 cigarettes per day for at least 5 years. They were thus heavy smokers before stopping. It then appears as if nicotine is at least causally involved in measures of impatience. More research should be able to shed light on the causal effect of addictive substances on vmPFC activity. It seems then that the dependence on nicotine at least partially causes a deviation from fully healthy vmPFC functioning and that is causes myopic or short-sighted behaviour where the future is not fully taken into account. Even though Becker and Murphy allow for weak rationality whereby a high dis-count rate causes some if this behaviour, hyperbolic disdis-counting or myopic behaviour cannot be ac-counted for in their model. True myopic behaviour ignores some of the future consequences and this could easily create dynamic inconsistencies allowing preference reversal.
As the discount rate is a measure of impatience of agents, it seems as if the discount rate is not a stable one across time but seems to be at least influenced by the act of smoking. A way to ac-count for this conjecture and for the findings of Bickel et al. (1999) is explored by Orphanides and Zervos (1998). They allow for a discount rate that is positively related to the consumption of an ad-dictive substance. With time the stock of total consumed adad-dictive substances increases and the discount rate increases proportionally to the size of this stock. Consuming more addictive substances thus leads to a higher discount rate and more impatient or impulsive behaviour. Their model then would be able to describe a consumption pattern that would look irrational with only standard dis-counting. This allowance for hyperbolic discounting implies that the existence of brain mechanisms which cause the observed consumption patterns of addicts does not necessarily exclude the
model-[ - 25 - ]
ling of a rational consumer who exhibits rational preferences. We have seen though that the elegant but simple model of Becker and Murphy alone does not suffice to explain these findings.
As for the neuroeconomic model by Bernheim and Rangel, they focus on the sensitization to the cue of addictive substances and not so much on the cognitive control function that the paragraph above touches upon. Their cue-triggered addiction model, models cue sensitization for particular goods. One key assumption in their model is that addicts have rational preferences in a cold mode, but can be nudged into a hot mode due to environmental cues. In a hot mode their decision making does not align with their preferences during the cold mode. This is thus a transitory phase and not a fixed underlying trait of the decision maker that always manifests itself. What Bechara et al. (2001) find though is that the neural data from addicts which accounts for impaired decision making vis-à-vis healthy controls is also at play during other choice environments. In these environments of mone-tary rewards and predictions, the hot-mode would not be entered as no cue related to the addiction is present. In other words, the impaired decision making seems to be a global problem and not a local problem only in the context of the specific addiction. It appears that this problem partly lies with the cognitive control function which for addicts seems to be weaker in general. Bernheim and Rangel acknowledge this fact when they say: ‘Since poor cognitive control increases the likelihood of becoming addicted, it should not be surprising that the typical addict exhibits other self-control prob-lems’ (Bernheim & Rangel 2004, p1565). The assumption that addiction only concerns a narrow do-main is true in itself (addicts are addicted to specific substances, not all at the same time). However, the role of the cognitive control function could be better explicated in the model and the susceptibil-ity towards becoming addicted is certainly an important factor to be accounted for deserving of more attention in an addiction model.
A finding that is harder to explain than the ones up until now for a standard model is illus-trated by Falk et al. (2011), but becomes more salient in Falk, Berkman and Lieberman (2012). It is related to the characteristics of addiction that Bernheim and Rangel list. In the 2012 study which had a focus group like setup. The BOLD signal in the vmPFC during the showing of different anti-smoking advertisements was predictive of its effectiveness for the general population when aired, whereas self-related reports on those advertisements by the subjects were not. This is again a challenge for the standard model. The self-reports of subjects (and experts) on the advertisements’ effectiveness did not line up with its observed effectiveness in the population, whereas the neural data was predic-tive of its effecpredic-tiveness. When we assume that the general population maps to the focus group, it turns out that a process which is not accessible to the agent herself does steer her choice behaviour. Clearly this is a problem for a rational choice theory which models a consumer with perfect informa-tion and a full understanding of the factors driving choices. No such understanding seems to exist in some real life situations.
[ - 26 - ]
Besides the question of whether self-reports are a good source of knowledge in the first place, positive economics have another answer ready. Any individual consumer is not required to adhere to strict rationality requirements, but behaves as-if she is fully rational and maximizes her utility. This as-if assumption has been introduced by Friedman in The Methodology of Positive
Eco-nomics (1953) and has been a standard reply in ecoEco-nomics ever since. He argues that an agent does
not need to be fully rational for her to act as-if she is. He draws the analogy with a billiard player. Even though he does not know the exact laws of physics, prediction of his shots is perfectly possible by assuming he does. In the same vein it can be argued that a rigorous understanding of how the information from the advertisements influences choice is not needed, as long as the axioms of con-sumer choice are not violated. As these do not appear to be violated in the Falk studies, rational ad-diction theory cannot be discarded through this neuroeconomic evidence.
Another finding by Bickel et al. (1999) however, proves problematic for all rational addiction models (that I know of). As stated before, rational choice theory posits that consumers do not violate the preference axioms, a key one of which is that preferences are non-transitive. If A is (revealed) preferred to B and B is preferred to C, then A must be preferred to C. This condition prevents cycles of choice, and thus opportunities for arbitrage. What Bickel et al. find though is that smokers dis-count cigarettes more heavily than an equivalent monetary award even though they are aware of the
fact that they represent the same monetary amount.
[ - 27 - ]
Now it is easy to see that this creates an inconsistency. Ignoring transaction costs and liquidity con-straints, say a smoker would just prefer a pack of cigarettes over 5 euro’s now. That is, the present value of a pack of cigarettes is just over 5 euro’s. The monetary amount and the cigarettes are both discounted. So at some point in the future the present value of the 5 euro’s in cash might be only 4 euro’s. Receiving 5 euro’s now is equivalent to receiving 4 euro’s x days from now. With cigarettes discounted with a higher rate, their present value at the same x days from now is lower than 4 euro’s. Let’s say it is 3 euro’s. Whereas in the present a pack of cigarettes is preferred over the 5 euro’s, in the future the monetary equivalent is preferred to the pack of cigarettes. Clearly this vio-lates the requirement of dynamic consistency. As it is a strict requirement in rational choice theory, something goes awry here. Perhaps a rational model would be able to incorporate these findings too but I do not see an easy or straightforward way to do so. The preference of many economists for rational choice theory stems from its power in explaining a wide range of phenomena while staying elegant and simple at its core. If the model must be patched with band-aids to account for findings such as these, it loses that attractive elegance. To be clear, I am not discarding the rational model here, but other models such as that by Bernheim and Rangel do gain in attractiveness by such find-ings as those that reject one of the axioms underlying strict rational theory.
Bernheim and Rangel do not make assumptions on the full rationality of addicts and there-fore do not run into the problem posited by the Falk studies, namely that data which is unavailable to the agent or outside observers does steer her choices. Since they moreover make use of the cue-triggered hot mode decision maker, they do not need to adhere to the consistency requirement for every decision of the agent. To the contrary, they explicitly incorporate mistaken choices into their model, and the dynamic inconsistency displayed here could serve as a justification for the fact that such mistaken choices do indeed occur. This does not mean however that the issue of mistaken choices is wholly uncontroversial as touched upon in section 4.1.
4.5. Choice, information and revealed preferences
A few issues that underlie the models of Bernheim and Rangel and Becker and Murphy have already been touched upon. The as-if assumption of the rational addiction model for example or Bernheim and Rangel’s contention that preferences cannot be unequivocally be deducted from observed choices. The discussion on choices, preferences and information could do with some more attention however.
[ - 28 - ]
It is hard to argue on the possibility of mistaken choices with the position of Gul and Pesen-dorfer as they refer back to the foundations of (micro)economics that exclude the existence of such choices a priori. If preferences are defined in terms of choices, the two can naturally not diverge. If that is the starting point of a model there is no use in pointing towards examples of where they do diverge. All one does in that case is to talk about two different notions of preferences. This seems a tug of war over the semantics of ‘preferences’. Let me circumvent this stalemate by taking a look at the notion of full information.
The availability of full information is a requirement in order to allow for a rational choice to be made. It is interesting to dig into what exactly is meant by full information. Is it just the principle that all information is available to an agent if she wishes to look for it? Or is it so that she should also be able to process all this information in the correct way? The bounded rationality concept intro-duced by Herbert Simon questions the idea that an agent would ever be able to process all informa-tion relevant to any decision. Agents necessarily filter the amount of informainforma-tion that goes into deci-sion making. What is brought up by the neuroscientific findings that Bernheim and Rangel introduce is the way in which the information processing might be skewed. The question then is whether we should be happy if the information is in principle available, or whether a requirement on the correct processing of this information should also enter rational decision making. Let me elaborate.
Rational choice theory spends little time on the combination of information and the human processing of this information. Through its abstraction it implicitly assumes that the agent is perfectly capable of processing the available information. The agent is really a perfect calculating machine where the information is not filtered, distorted or skewed by what goes on inside the brain. This as-sumption of a calculating machine works quite well in everyday consumer choices. But with addiction this model is operating in a very different environment. What we have seen with Bernheim and Rangel is that the brain might systematically skew information by the influence of drugs on the he-donic forecasting mechanism, causing expected hehe-donic reward and actual hehe-donic reward to di-verge. To me it seems that this is important to take into account. What economists are really inter-ested in is deducting preferences from observed choices in order to predict behaviour or calculate the impact of policy on welfare. If the actual decision making process has information as an input that does not align with how an objective outside observer would see this information, the decision making process is not supplied with full information.
For Gul and Pesendorfer the moment of distortion of information seems to matter in whether it is to be taken seriously or not. Here is a thought experiment to tease out why I would maintain that there is a certain amount of arbitrariness to this. Let us imagine the following. Person A would like to buy a car, she likes red cars over blue ones. She finds an online advertisement of a nice car with a photo of it that shows a red car. She transfers the money and only then to her dismay finds
