The Role of Temporal Normalization in Preference Instability
Student: Urša Bernardič, MSc Brain and Cognitive Sciences, University of
Amster-dam
Supervisor: dr. M.P. (Maël) Lebreton
UvA- coassessor: dr. J.B. (Jan) Engelmann
ABSTRACT
In a dynamic environment, our choices are never made in a isolation. Understanding how the brain computes value in this dynamic environment is, therefore, a basic question to researchers at the intersection of neuroscience, psychology and economics. Values, which quantify the expected rewards associated with any choice or action, are the theoretical determinants of choices and preferences. Hence, understanding value processing is cru-cial for understanding a decision-making process. A major challenge in decision theory is to account for the instability of preferences across time and context. The aim of this pre-liminary study was to apply computational principles derived from the electrophysiologi-cal non-human animal study of temporal normalization –alos referred to as range adapta-tion-, to understand suboptimalities in human choice behavior. The results indicated that temporal normalization does occur under some circumstances.
INTRODUCTION
Consider the response of a professor, whose goal is to fairly select the best student be-tween applicants for an internship project. In particular, would (s)he make a fair and op-timal choice if one of the student applications is evaluated just after the professor grad-ed exams from worse students (e.g. low grading context) while the other is considergrad-ed/ evaluated just after the professor graded exams from excellent students (e.g. high grad-ing context)? The same question arises even in daily-life contexts: can a shop/restaurant/ airline seduce you into buying an expensive product by offering similar but more expen-sive alternatives?
In decision theory, a decision maker (a person, a company, or a society) chooses what to do from a set of alternative choices. Different disciplines in behavioural science have ap-proached human decision-making from a variety of perspectives. It is common to distin-guish between normative decision theories, which seek to yield prescriptions about what decision makers are rationally required to do (see Pascal, 1670; Bernoulli, 1954; Samuel-son, 1938), and descriptive theory which seeks to explain and predict how people
actual-ly make decisions (e.g. experimental psychology). From the normative point of view, the
only way a purely economic scientist could infer what things are worth to people (prefer-ence ordering) is by observing their choices (Samuelson, 1938). Moreover, most normative theories in economics consider decision problems as maximization problems. In this view, a decision maker first computes, for each available course of action, a decision variable or utility, which combines the positive and negative aspects of that option. Afterwards a decision maker chooses the option with the highest utility. Therefore, as noted by some others (e.g. Samuelson, 1973), the mere notion of utility “is simply a highly convenient way of compactly representing binary preferences”. Utility, in neoclassical economics
and normative decision making models, can be defined by a series of rational axioms, which assumes that preferences are internally coherent and consistent with a stable and context-independent utility function (Luce, 1959). Hence, expected utility theory (EUT) assume that choices are based on an internal, absolute utility scale which should not be affected by other available alternatives, set size, or temporal history (Stephens & Krebs, 1986; Von Neumann and Morgenstern, 1944). Interestingly, that has been deemed the greatest advantage (because it provides clarity when choices are rational) as well as greatest disadvantage (because irrational choices are common enough that they cannot simply be dismissed) of EUT (Hartner, 2014).
In contrast with normative theories, a sizeable body of empirical evidence in psychology suggests that people do not have well defined attitudes (psychologists’ preferred term for the imputed sum of the positive and negative aspects). Researchers found that individual preferences are often unstable, created in the moment and value representation is pro-foundly affected by environment (Norton et al., 2004), set size (Iyengar & Lepper, 2000; Boatwright & Nunes, 2001; Bertrand et al., 2010), emotions (Strack et al., 1988), and context (Shafir et al., 2002; Bateson et al., 2003; Tversky, 1972; Louie. Glimcher & Webb, 2015; Tsetsos et al., 2012). Thus, real agents tend to violate the axioms of EUT common-ly, which is a major problem for EUT as a realistic model of human behavior (Kahneman and Tversky 1979).
The last decade has seen the development of neuroeconomics, which unites the knowl-edge from behavioral economics with experimental neuroscience and psychology (Hart-ner, 2014). Its key methodological innovation is to use well-defined theories in eco-nomics, like EUT, and explain them with algorithmic models of decision-making and their relevant neural mechanisms (Glimcher, 2010). Therefore insights from neuroeconomics might help to investigate the neural mechanisms where EUT goes astray, i.e., when agents violate its axioms and hopefully offer to better predict and explain those viola-tions (Hartner, 2014).
To do so, neuroeconomists defined a construct of subjective value (SV), which represents the mean firing rates in action potentials per second of specific populations of neurons which predict choices of agents (though stochastically) (Glimcher, 2011). As such, SV re-lates well with the two-stage model of decision making: valuation (participants assign, store and retrieve SV or utilities) and choice/preference mechanism (participants com-pare these values and chooses one of the options). One of the interesting empirical ques-tions in neuroeconomics has been whether SV exists, and whether neuroscientist could find a neural firing rate pattern or blood-oxygen-level dependent (BOLD) signal correlat-ed with utility (Hartner, 2014). A growing body of evidence show that there are cell popu-lations and brain regions, which encode for SV in ventromedial prefrontal (VMPFC) and
orbitofrontal cortex (OFC) (Hare et al., 2009; Knutson et al., 2007; McClure et al., 2004; Paulus and Frank, 2003; Plassman et al., 2007). Moreover, such brain valuation systems are found to be generic, which means that different types of SV (of food, trinkets, faces, houses, painting, music, etc.) are automatically encoded in the same brain areas (Lebre-ton et al., 2009; 2012; Plassmann, O’Doherty & Rangel, 2007; Chib et al., 2009; Mc-Namee, Rangel & O’Doherty, 2013). These findings lead to an interesting and fundamen-tal question: Can properties and limitations of this biological implementation in our brains (on the level of neural encoding), explain irrationalities in the behavior?
One property of neural encoding, first put forward by Horace Barlow, is known as the
ef-ficient coding hypothesis. The efef-ficient coding hypothesis appeals to two fundamental
limitations to neural response: 1) the limited capacity of neuronal information processing systems and 2) their consequent need for efficiency (Barlow, 1961; Louie & Glimcher, 2012). At the level of single neurons, efficient coding requires that the input-output func-tion adjusts the responses of sensory systems to the regularities of their input and there-fore employ all activity levels equally in response to the distribution (see Rangel & Clithero, 2012). If the sensitivity is set too low, high levels of the stimulus feature will be indistinguishable as the response function saturates and, vice versa if the sensitivity is set too high, low levels of the stimulus feature cannot drive responses (Louie & Glimcher, 2012). This strategy ensures that all possible neural response levels are used with equal frequency (Louie & Glimcher, 2012).
However, such optimization is an inefficient strategy if a sensory system is hard-wired to only the global, long-term average statistics of the environment (Louie & Glimcher, 2012). More specifically, it cannot efficiently transmit information if the short-term, local statistics vary and if the response function does not adjust appropriately (Louie & Glim-cher, 2012). For example, if a hypothetical neuron signals the value of widely disparate worth (recall our professor situation), then only a very limited portion of its dynamic fir-ing rate could be devoted to any given value (e.g. excellent or bad students) and it will be very hard for the neuron to signal consistently the preferred student. However, exper-imental data from across sensory systems suggests (Schwartz & Simoncelli, 2001; Shapley & Enroth Cugell, 1984, Barbour & Wang, 2003; Rabinowitz et al., 2011; Padoa-Schioppa; 2009) that the sensory system adapts to local statistics, such as mean and the variance of stimulus distributions. One of the proposed mechanisms by which neural systems could instantiate context dependence is normalization, which implements relative coding of the represented parameter, and discards information about absolute magnitude (Louie & Glimcher, 2012).
Above findings suggest that value coding systems therefore face two challenges: 1) exter-nal constraints of different statistics of reward distributions in the natural world and 2)
internal constraints of neuronal and physiological status (Louie & Glimcher, 2012). There-fore, under normalization the same objective value can get different “encoded values” depending on the context. For example let’s consider our professor situation at the be-ginning again. Under the absolute coding hypothesis, the response of a professor and con-sequently, the firing rates in the neuron, which encodes for the SV at the time of evaluat-ing the decision, are the same in both low and high grade context. By contrast, under the normalized coding hypothesis, the response and firing rate in the neuron described here, are higher in the lower reward context.
Furthermore, normalization in neural encoding is usually studied in two different con-texts: 1) spatial and 2) temporal normalization. Spatial (also known as divisive normaliza-tion) where perception is usually manipulated by spatial context effects (e.g. interaction of different stimulus features at an instant time. Or in terms of decision-making, the rel-ative preference between two given alternrel-atives change with the quantity or quality (attributes) of other alternatives at the time of decision)(see Louie, Khaw, Glimcher, 2015; Carandini & Heeger, 2012). On the other hand, temporal normalization study the response to a sustained presentation of a stimulus in the recent past and often refers to range-adaptation (Louie & Glimcher, 2012). To conclude, the normalization model pro-poses that aspects of different choice context from both, spatial and temporal paradigm, will influence choice efficiency through the underlying mechanism of value representa-tion, which will cause preference instability. But is there any evidence for temporal nor-malization?
One of the pioneering experimental works providing support for temporal normalization of value signals in value-based decision making was carried out by Tremblay and Schultz (1999). Given the demonstration that OFC neurons encode for different SVs, Tremblay and Schultz (1999) examined whether those value representations are dependent on the other available rewards in a choice situations. In a Pavlovian reward task with value range constant across blocks, which had different range and mean reward, they found that neural responses are consistent with value normalization. Hence, discrimination be-tween pairs appears to be based on the relative preferences for different rewards, rather than physical, absolute, unchanging properties of the reward itself.
In a similar vein, Padoa-Schioppa (2009) confirmed the role of temporal dynamics on how different reward possibilities were presented. The author reanalysed two large datasets (Padoa-Schioppa and Assad, 2006; Padoa-Schioppa and Assad, 2008) and showed that neu-rons in OFC encode value in a linear way. However, this encoding exhibits range adapta-tion: the slope of the relationship between firing rate and value changes under different value ranges conditions (∆V). For example an offer value of 2 will have lower value in high-range conditions (∆V=10) than in low-value condition (∆V=2). Hence, neurons
acti-vated range within high and low-conditions, remains constant. Therefore, the encoding of values adapts to the local temporal context (range adaptation), but not to the immediate context (menu invariant) (Padoa-Schioppa, 2009). Overall, the normalization model, found in different species and for different contexts, proposes that structural aspects of choice context will influence choice efficiency through the underlying mechanism of val-ue representation (Glimcher, 2011).
Inspired by previous discussed electrophysiological non-human animal studies which found temporal normalization on the neuronal level, the aim of this preliminary study was to investigated whether such temporal normalization (as a range adaptation) can give us better understanding of the process behind irrational decisions, also on the level of hu-man behavior. In order to link normalized value coding to choice behavior, we incorporat-ed into our task design items from two different domains (food and trinkets: -which refers to an inexpensive item non-food item). We used thees items in two tasks: a rating, and a choice task (Fig. 1A & B). Critically, we manipulated the underlying distribution of values in the rating task (Fig. 1C), and assessed the consequence of this rating, on the stability of preferences in the subsequent choice task. As suggested by the generality of encoding SV in the same brain area (Lebreton et al., 2009; 2012, Plassmann, O’Doherty & Rangel, 2007, 2007; Chib et al., 2009; McNamee, Rangel & O’Doherty, 2013., 2013), the aim of our design was to test whether such temporal normalization is domain specific or generic.
EXPERIMENTAL PROCEDURES Participants
This study was approved by the Ethics Review Board at Faculty of Social and Behavioral Science, University of Amsterdam. Eleven participants, 6 males (aged: 28.3 ± 10.9 years) and 5 females (aged: 23.2 ± 3.3 years), were asked to participate in this study. Partici-pants were told that the goal of the experiment was to study food and trinket prefer-ences.
Stimuli
Participants made decisions regarding two classes of goods: 160 non-food or trinket items (e.g. office/house/free-time material,etc.) and 160 food items (e.g. vegetables/fruits/ snacks,etc.), for a total of 320 pictures. The pictures were chosen to cover a large vari-ety of tastes and styles and were gathered from an online shop Mercator d.d. We con-trolled that size and resolution was matched between pictures. To reduce perceptual un-certainty regarding the content of the picture every picture had a short description∂ for what exactly participants are rating (e.g. name and quantity).
These pictures were selected from the set of 400 pictures used in a pilot behavioral study, where 6 participants rated their willingness to pay (WTP) and their familiarity for the items. This was primarily done 1) to remove uncertainty considerations from the WTP computation as much as possible and ensure that participants knew the items and 2) to estimate the value of the items, to manipulate the distribution of item value presented in the rating task. Therefore, 320 stimuli were assigned to the four sessions, such that each session employed 80 new stimuli, equally divided between each picture category (40 food and 40 trinkets).
Behavioral Tasks
All tasks were programmed on a PC using the Cogent 2000 (Wellcome Department of Imaging Neuroscience, London, UK) library of Matlab functions for stimuli presentation. The pilot behavioral study (N=6) consisted of 4 sessions of two computerized tasks: a WTP-value-rating and a familiarity task. Each session consist of 100 value rating and 100 familiarity ratings for two categories of items: trinkets and food. WTP ratings and famil-iarity ratings were used to select 320 pictures which were ranked as most familiar and had the smallest standard deviation in WTP ratings. Averaged item WTP ratings were also used to attribute items to blocks of varying “value distributions” in the rating task (Fig. 1C).
Figure 1. Experimental Task (A&B) and Manipulation (C&D) Overview. In the rating task
(A), participants had to move a cursor on a scale from 0€ to 5€, to indicate their willing-ness to pay for the item, which could be a food or trinket. In the subsequent choice task (B), participants had to state which item they preferred between two of those in the same category (food or trinket). Given trials are illustrated from left to right, with dura-tions in milliseconds. (C) Design of the 4 different condidura-tions of value-rating task. Unbe-knownst to participants, the mean of the distribution of value is manipulated between the 2 blocks (halves) of the task for one category (food or trinket) and randomly selected for the other category. Moreover to cancel the effect of which items (with small or big values) are presented first, we presented in condition 1 & 3 manipulated items with small value first and later items with big values and vice versa in condition 2 & 4, first big val-ues and later small valval-ues items. (D) Task design of the binary choice. To evaluate the impact of our manipulation on valuation consistency, the task was designed with 50% within- (both items for the same block) and 50% between-blocks choices (or items for a different blocks).
The main experiment (N=11) consisted of 4 sessions of two computerized tasks: a WTP-value-rating and a binary choice task. Each session consist of 80 WTP-value-rating and 80 bina-ry choice trials for two categories of items: trinket and food. Each trial started with 0.5s fixation, after a picture was displayed on the screen and participants had to rate their
(50%)
WTP on a scale, graduated between 0 to 5 Euros, in price intervals of .50€ (Fig. 1A). Par-ticipants could move the cursor by pressing a left or right arrow keys on the keyboard. Rating was self-paced and participants were instructed to press a spacebar to validate the amount indicating their WTP and go to the next trial. The initial position of the cur-sor on the scale was randomized to avoid confounding the ratings with the movements they involve.
Critically, unbeknownst to participants, the underlying distribution of the items was ma-nipulated, creating two “blocks” with different valuation “contexts”. In this first value-rating task, we manipulated the underlying mean of the item values between the two blocks: e.g. block 1 with items which have low value and block 2 with items which have high value. At the session level, the 2 (items: trinket and food) x 2 (manipulations: high versus low block, and low versus high block) within-subject design allowed us to design 4 different conditions (Fig. 1C). Conditions were counterbalanced across subjects.
The average duration of rating session (n=80 pictures) was around 6 to 7 minutes. After each rating task session, participants were instructed that the task would change, and new rules were explained, after that a choice task session was performed.
On each trial of the choice task, participants were presented with pairs of items drawn from the rating task, and asked to indicate which item they prefer (Fig. 1B). At the trial level, a similar procedure was implemented for both the rating and choice tasks: after the fixation for 0.5s, two pictures were displayed side by side (Fig. 1B). The relative posi-tion of the two pictures on the screen was also randomized. Participants were asked to indicate a preference self-paced by pressing left or right arrow keys on the keyboard, and validate their response with pressing a spacebar. Critically, the couple of items can be drawn from the same manipulation-block of the rating task: this is a within block (WB) comparison: e.g. both items are either from block 1 or block 2 (Fig. 1D). Alternatively, the couple of items can be drawn from the different blocks of the rating task: this is a between-block (BB) comparison: e.g. one item from block 1 and second item from block 2. Moreover we tried to control temporal valuation distance with designing WB/BB choice pairs in easy or hard comparison. Therefore, 50% of the WB choice situations implement-ed choices between pictures rankimplement-ed n and pictures rankimplement-ed n+10 (easy comparison) in the rating task, and 50% implemented choices between pictures ranked n and pictures ranked n+1 (hard comparison) in the rating task. In the BB choice pairs 50% of the implemented choices were ranked n and pictures ranked n+30 (easy comparison), and 50% of the BB choice situations implemented choices between pictures ranked n and pictures ranked n +10 (hard comparison). Hence, half of the choices are equivalent in term of temporal dis-tance in the WB and BB conditions. In each session of choice task participants performed 40 WB comparison and 40 BB comparison and the average duration of choice session (n=80 comparison) was 5,4 min.
Statistical Analysis
The link between item valuation (WTP) and binary choices was modelled using a logistic choice function. Given the ratings of any two items VA and VB the probability (P) of
choos-ing left (A) over right (B) was estimated uschoos-ing a logistic function (φ):
P(A)= φ (VA - VB)
The logistic function ( φ(x) ) is:
where e represents the natural logarithm base, ß0 represents the intercept, and ß1 the
slope of the logistic function. In order to avoid any potential confound in ß1, due to
dif-ferences in the distribution of our independent variable x, we z-scored independent vari-able (x) per condition and per participant:
where µ represents the mean and σ the standard deviation of VA - VB for all trials per subject.
Hence our model is:
Obtained model parameters averaged per condition (between sessions), and per subject. Qt the population level, we used repeated measure ANOVAs . We considered three signifi-cance levels: 0.05, 0.01, and 0.001. Data distribution was assumed to be normal but this was not formally tested. All statistical analyses were performed with Matlab Statistical Toolbox (Matlab R2015b, The MathWorks, Inc., USA).
(1)
(2)
(3)
RESULTS
We constructed several analyses to establish the validity of the experimental design. Be-cause the value distribution of food and trinket items were generated from average pic-ture rankings (WTP) obtained in a pilot study, we verified that WTP ratings of our stimuli in our study were indeed successfully implementing the value-distribution manipulations between the blocks of our four conditions (Fig. 1C). Using paired-wise t-test we found that ratings were significantly impacted manipulation(i.e. shift in value), both when trin-kets were the manipulated category (in first condition t(10) = -3.86, p = 0.003, second session t(10) = 4.94, p < 0.001), and when food items were the manipulated category (third session t(10) = -4.81, p <0.001, fourth session t(10) = 7.29, p < 0.001). As expect-ed, there were no statistically significance for the non-manipulated category (i.e. con-stant value) for neither trinkets or food (Fig. 2).
Figure 2. Ratings (WTP) comparison according to conditions. In first two conditions
trin-kets were manipulated and food was not manipulated and contrary, in third and fourth condition food was manipulated and trinkets were not manipulated. Both, manipulated trinkets and food have significant difference between block1 and block2. Moreover, with first and third condition we tested small vs big block, and in second and fourth condition we tested big vs small block. As expected, there is no statistically significance for the non-manipulated items, neither for trinkets nor for food. Error bars represent intersub-ject SEM. Note. *=p < .05, **=p < .01, ***=p <.001
Figure 3. Results comparison according to the arrangement of within block and between
block pairs between non manipulated (NM) and manipulated condition (M). (3A) WTP av-erage value ratings between WB and BB pairs, avav-eraged separately for non manipulated (NM) and manipulated task (M) session. (3B) The slope of the logistic choice function of within- (pictures are selected from the same block) vs between-block (couple of items are selected from the different blocks), for the both categories of items. Error bars rep-resent intersubject SEM. Note: NS.
Our expectations were that the more items participants rate, the more adapted to the range (mean) ratings become. Therefore, we hypothesised this might cause that the lo-gistic choice function will be higher in BB than WB manipulated comparison. And this ef-fect could be domain specific (to just manipulated items) or it should have efef-fect on both, food and trinkets (generic).
To assess that, we estimated separate logistic model for (WB) and between blocks (BB) choices, for both trinkets and food items (Fig. 1 D). We used two-way repeated-measures ANOVA to test for the effect (Fig. 3) of the value distribution manipulatec (M) vs (NM), on the slope of the logistic choice function for WB vs BB choices. However, we only found a significant main effect of our Manipulation (F(1,10) = 15.88, p = 0.003). We did not find a significant effect of the WB vs BB effects (F(1,10) = 1.15, p = 0.309), nor a significant interaction between our two factors (M vs NM x WB vs BB) (F(1,10) = 2.9, p = 0.119). Next, we explored whether this effect was similar for both domains (food and trinkets), and whether manipulation order (small or big block presented first) affect the manipula-tion. We therefore extracted the slope of the logistic choice function in all our condi-tions.
Figure 4. Slope of logistic choice function comparison according to 4 different session
conditions. Error bars represent intersubject SEM. Note. *=p < .05, **=p < .01, ***= p<.001 Analyses showed that the slope is significantly lower in the manipulated versus non-ma-nipulated condition when the trinkets are manon-ma-nipulated (for Condition 1, t(10) = -3.28, p = 0.008 and in Condition 2, t(10) = -2.37, p = 0.034) (Fig 3; Condition 1 & Condition 2). This observation was replicated in conditions irrespective of whether the block of high (vs low) value trinkets was presented first in the rating task (condition 1), or second (condi-tion 2) (Fig. 3). However, we did not find any statistical difference in Condi(condi-tion 3 and 4 (Fig. 3), where food was manipulated and trinkets were not manipulated. The results therefore indicate that ratings are less predictive of choices in manipulation condition for trinkets, as hypothesized in our operationalization of temporal normalization, however this do not hold for food items.
To conclude, our preliminary results suggest that manipulation (as range adaptation) sig-nificantly decrease the predictive power of rating on choices, as indexed by the slope of the logistic function. However, this seems to be specific to manipulations of the trinkets items, and could not be generalized to our food items. Therefore, we could not further test whether normalization is generic, as food items did not response to our manipula-tion. Nonetheless, given a small sample size (n=11), which clearly limit the power of the current study, more behavioral data has to be collected before making any definite claim.
DISCUSSION
A highly replicated finding in experimental decision sciences (behavioral economics, psy-chology, neuroeconomics) is that preferences are not stable, which violates one of the fundamental assumptions of normative theories. Namely axioms, such as transitivity, which suggest that decisions reflect absolute valuations assigned to individual options and cannot be biased by the context in which they were formed (Ariely, Loewenstein & Pr-elec, 2003). Although various mechanisms have been proposed for such effects (Rangel & Clithero, 2012; Louie & Glimcher, 2012), to the best of our knowledge little research has paid attention to the effect of value temporal normalization (as a range adaptation) on behavioral preference stability. We investigateed such effects of temporal normalization on preference instability, by sequentially presenting items in a rating task, in two differ-ent conditions, manipulated and non-manipulated (regarding to the mean of value distri-bution) (Fig. 1C & 2). Then, participants had to indicate binary preferences between coupes of items, which could be sampled from blocks of similar (within block) or differ-ent (between block) valuation contexts (Fig. 1D & 3).
In this study we did not observe normalization per se, therefore we assessed difference in normalization through the logistic choice model, which accounts for (see methods, (4)). The results did not show significant difference in normalization between or within blocks for manipulated or non-manipulated contexts. Thus, it might be that our hypothesis is wrong, or that the design or model, which controls for temporal normalization might not be accurate or optimal to examine the effect of temporal normalization on preference instability.
A possible explanation for the latter might be, that our model should take into account not just mean of certain value distribution, but also variance, as results from sensory sys-tem adapt to both local statistics. Moreover, it was out of the scope of this research to investigate how normalization parameters evolve over time. Although it is out of the scope of this report, we believe that restricting analyses to WB-BB of different -hard and easy choices of our data, could offer carefully controlling for the confounding effects of temporal distance on temporal normalization. Moreover, the link between ratings and preferences could also be inferred from decision times. According to well-described drift diffusion models, decisions take longer if alternatives have closer values, therefore the difference in WTP ratings might predict the response time (RT) in the choice task. Thus, the reanalysis of our data could also binned trials into two bins (easy trials and hard tri-als) and fit the response time (RT) by applying the Ratcliff drift diffusion model (see Rat-cliff and McKoon, 2008), which has recently been shown to accurately account for RT dis-tributions in the context of value-based decision making (Mormann et al., 2010).
In our current design we used a scale from 0€ to 5€ in price increments of .50€. Although even less precise scales where used in similar studies (Plasmann, O'Doherty & Rangel, 2007), we acknowledge that biasing effects might be harder to observe on such coarse value intervals (Kahneman, Schkade & Sunstein, 1998). Importantly, such a ratings from a scale from 0€ to 5€ is also bounded, whereas our intrinsic values may not be. To be more precise, our intrinsic values are not just positive, but also negative (i.e. imagine the veg-etable you really do not like). Therefore non-monatery liking scales might be more useful for such investigations, as they offers us to make positive, but also negative ratings of our values. To push this idea even further, Payne and colleagues (1999) suggested that intrin-sic values are unlikely to be represented in memory in dollar terms (Payne et al., 1999). All together, this suggest that our model instead of real, intrinsic values uses simplified computations of the mean and variance of ratings and not values. These simplifications could introduce noise into our model which might be one of possible explanations, why we did not observe the effect of temporal normalization. Another important limitation is that our study used auction, economic (WTP) design, but we did not sell items according to Becker-DeGroot_Marschak auction (Becker et al., 1964), which is commonly used auc-tion instituauc-tion of market transacauc-tions in the laboratory (Plassmann, O’Doherty & Rangel, 2007).
Many questions remain unanswered. In particular, our analysis showed that the slope is significantly lower in the manipulated compared to non-manipulated condition, however this hold just for trinkets. This is consistent with our hypothesized operationalization of temporal normalization, which indicates that ratings for trinkets are less predictive when the range context is manipulated compared to non-manipulated. One possible explana-tion for that could be that preferences are not only a point estimate, but may also have an uncertainty/confidence/precision parameter. Hence, this could lead us to the ques-tion: how bias our preferences are by contextual factors? In this perspective, real agents might be more precise/confident when rating food items, which we see/buy on the week basis. Or in terms of temporal normalization, it might be that our range is already quite stable (we have more knowledge about it), and therefore it is harder to manipulate it. In contrast we do see/buy trinkets items on a less regular basis, than food. Hence, we might be less precise and therefore more susceptible for manipulation as our knowledge of the range is more uncertain. This idea is also consistent with the informal debriefing after the end of experiments, where participants mentioned that they were ‘familiar with items, however [I] had more difficulties stating prices for trinkets’.
To push this idea even further, recent findings by Lebreton and colleagues (2015), provid-ed an interesting evidence that confidence ratings are indeprovid-ed encodprovid-ed in the same brain area as SV of different goods; VMPFC. Therefore, the next research question could be, whether confidence ratings could give us better insight in the role of temporal
normaliza-tion. To be more specific, our results could suggest that less precise valuation might be more accessible for temporal normalization, however this is out of the scope of this re-search and future studies are needed.
In summary, irrational decision making has been puzzling economists and psychologist for decades. An important part in neuroeconomics research agenda is to understand how the brain evaluates goods, and how other cognitive, neuronal, emotional and visceral pro-cesses affect the computation of economic value. Hence, fMRI studies and computational modelling might give us better, and especially more precise insights into the effect of temporal normalization of SV. Hence, we believe future fMRI studies should also try to investigate how, not just where, this cognitive valuation process is implemented in the brain. Underlying even more precise temporal mechanisms on lower, neuronal level will help us answer how neural circuits implement the relative information coding in contex-tual modulation. Finally, combining findings from different: 1) computational, 2) behav-ioral and 3) neuronal levels are a key goals in linking and understanding spatiotemporal normalization mechanisms.
Although relative (rather than absolute) valuation is a good news for marketing con-sulants, it is a challenge for normative theories which are schooled in traditional notion that WTP for goods reflects stable, inherent and known value. However, we hope that understanding the normalization and its qualitative features will contribute towards a building code for measuring constructed preferences. And last, but not least - help us (students, professors, politicians, companies, etc.) make better decisions!
SUPPLEMENTAL INFORMATION
Experimental instruction and ANOVA table.
ACKNOWLEDGEMENTS
Many thank to dr. Maël Lebreton for his great ideas, and willingness to always offer an invaluable advice and support throughout the entire internship. We would like to thank dr. Jan Engelmann for his helpful suggestions for research proposal agreement. We are also grateful to the Mercator d.d. for sharing pictures from their online shop.
REFERENCES
Abitbol, R., Lebreton, M., Hollard, G., Richmond, B.J., Bouret, S., Pessiglione, M., 2015. Neural Mechanisms Underlying Contextual Dependency of Participantive Values: Converg-ing Evidence from Monkeys and Humans. J. Neurosci. 35, 2308–2320. doi:10.1523/JNEU-ROSCI.1878-14.2015
Ariely, D., Loewenstein, G., Prelec, D. (2003). “Coherent arbitrariness”; stable demand curves without stable preferences”. Technology, 73.
Barlow, H. B. (1961). Possible principles underlying the transformations of sensory mes-sages.
Bateson M, Healy SD, Hurly TA. Context-dependent foraging decisions in rufous humming-birds. Proc Biol Sci. 2003;270(1521):1271–1276. [PMC free article]
Bernoulli, D. (1954). Exposition of a new theory on the measurement of risk. Econometri-ca: Journal of the Econometric Society, 23-36.
Boatwright, P., & Nunes, J. C. (2001). Reducing assortment: An attribute-based approach. Journal of marketing, 65(3), 50-63.
Chib, V.S., Rangel, A., Shimojo, S., O’Doherty, J.P., 2009. Evidence for a Common Repre-sentation of Decision Values for Dissimilar Goods in Human Ventromedial Prefrontal Cor-tex. J. Neurosci. 29, 12315–12320. doi:10.1523/JNEUROSCI.2575-09.2009
Glimcher, P. W. (2010). Foundations of neuroeconomic analysis. Oxford University Press. Hare, T. A., Camerer, C. F., & Rangel, A. (2009). Self-control in decision-making involves modulation of the vmPFC valuation system. Science, 324(5927), 646-648.
Hartner, D. F. (2014). From desire to subjective value: what neuroeconomics reveals about naturalism. Erasmus Journal for Philosophy and Economics, 7(1), 1.
Iyengar, S. S., & Lepper, M. R. (2000). When choice is demotivating: Can one desire too much of a good thing?. Journal of personality and social psychology, 79(6), 995.
Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica: Journal of the Econometric Society, 263-291.
Lebreton, M., Abitbol, R., Daunizeau, J., Pessiglione, M., 2015. Automatic integration of confidence in the brain valuation signal. Nat. Neurosci. 18, 1159–1167. doi:10.1038/nn. 4064
Lebreton, M., Jorge, S., Michel, V., Thirion, B., Pessiglione, M., 2009. An Automatic Valu-ation System in the Human Brain: Evidence from Functional Neuroimaging. Neuron 64, 431–439. doi:10.1016/j.neuron.2009.09.040
Lebreton, M., Kawa, S., d’Arc, B.F., Daunizeau, J., Pessiglione, M., 2012. Your Goal Is Mine: Unraveling Mimetic Desires in the Human Brain. J. Neurosci. 32, 7146–7157. doi: 10.1523/JNEUROSCI.4821-11.2012
Louie, K., Glimcher, P.W., 2012. Efficient coding and the neural representation of value. Ann. N. Y. Acad. Sci. 1251, 13–32. doi:10.1111/j.1749-6632.2012.06496.x
Louie, K., Glimcher, P. W., & Webb, R. 2015. Adaptive neural coding: from biological to behavioral decision-making. Current opinion in behavioral sciences, 5, 91-99.
Louie, K., Khaw, M.W., Glimcher, P.W., 2013. Normalization is a general neural mecha-nism for context-dependent decision making. Proc. Natl. Acad. Sci. 110, 6139–6144. doi: 10.1073/pnas.1217854110
Luce, R.D. 1959. Individual Choice Behavior: A Theoretical Analysis. Wiley. New York McClure, S. M., Laibson, D. I., Loewenstein, G., & Cohen, J. D. (2004). Separate neural systems value immediate and delayed monetary rewards. Science, 306(5695), 503-507. McNamee, D., Rangel, A., O’Doherty, J.P., 2013. Category-dependent and category-inde-pendent goal-value codes in human ventromedial prefrontal cortex. Nat. Neurosci. 16, 479–485. doi:10.1038/nn.3337
Mormann, M. M., Navalpakkam, V., Koch, C., & Rangel, A. (2012). Relative visual saliency differences induce sizable bias in consumer choice. Journal of Consumer Psychology, 22(1).
Norton, M. I., Vandello, J. A., & Darley, J. M. (2004). Casuistry and social category bias. Journal of personality and social psychology, 87(6), 817.
Padoa-Schioppa, C. (2009). Range-adapting representation of economic value in the or-bitofrontal cortex. The Journal of Neuroscience, 29(44), 14004-14014.
Pascal, B. (1910). Pensées (W. F. Trotter, Trans.). London: Dent. (Original work published 1670).
Paulus, M. P., & Frank, L. R. (2003). Ventromedial prefrontal cortex activation is critical for preference judgments. Neuroreport, 14(10), 1311-1315.
Payne, J.W., Bettman, J.R., Schkade, D.A., Schwarz, N., Gregory, R., 1999. Measuring Constructed Preferences: Towards a Building Code, in: Fischhoff, B., Manski, C.F. (Eds.), Elicitation of Preferences. Springer Netherlands, pp. 243–275.
Plassmann, H., O’Doherty, J.P., Rangel, A., 2010. Appetitive and Aversive Goal Values Are Encoded in the Medial Orbitofrontal Cortex at the Time of Decision Making. J. Neurosci. 30, 10799–10808. doi:10.1523/JNEUROSCI.0788-10.2010
Plassmann, H., O’Doherty, J., Rangel, A., 2007. Orbitofrontal Cortex Encodes Willingness to Pay in Everyday Economic Transactions. J. Neurosci. 27, 9984–9988. doi:10.1523/JNEU-ROSCI.2131-07.2007
Rabinowitz, N. C., Willmore, B. D., Schnupp, J. W., & King, A. J. (2011). Contrast gain control in auditory cortex. Neuron, 70(6), 1178-1191.
Rangel, A., Clithero, J.A., (2012). Value normalization in decision making: theory and ev-idence. Curr. Opin. Neurobiol., Decision making 22, 970–981. doi:10.1016/j.conb.
Rangel, A., & Clithero, J. A. (2012). The computation of stimulus values in simple choice. Neuroeconomics: Decision making and the brain, 2, 125-147.
Samuelson, P. A. (1938). A note on the pure theory of consumer's behaviour. Economica, 5(17), 61-71.
Schwartz, O., & Simoncelli, E. P. (2001). Natural signal statistics and sensory gain control. Nature neuroscience, 4(8), 819-825.
Shafir S, Waite TA, Smith BH. Context-dependent violations of rational choice in honey-bees (Apis mellifera) and gray jays (Perisoreus canadensis) Behav Ecol Sociobiol. 2002;51(2):180–187.
Simonson I. Choice based on reasons: The case of attraction and compromise effects. J Consum Res. 1989;16(2):158–174.
Stephens, D.W. & J.R. Krebs. 1986. Foraging Theory. Princeton University Press. Prince-ton, NJ
Strack, F., Martin, L. L., & Stepper, S. (1988). Inhibiting and facilitating conditions of the human smile: a nonobtrusive test of the facial feedback hypothesis. Journal of personali-ty and social psychology, 54(5), 768.
Tsetsos, K., Chater, N., & Usher, M. (2012). Salience driven value integration explains de-cision biases and preference reversal. Proceedings of the National Academy of Sciences, 109(24), 9659-9664.
Von Neumann, J., & Morgenstern, O. (1944). Games and economic behavior. Princeton, NJ.
