Neuronal origins of reduced accuracy and biases in economic choices under sequential offers

Two monkeys participated in the experiments. In each session, they chose between two juices labeled A and B, with A preferred. Offers were represented by sets of colored squares on a monitor, and animals indicated their choice with a saccade. In each session, two choice tasks were randomly interleaved. In Task 1, offers were presented simultaneously (Figure 2A); in Task 2, offers were presented in sequence (Figure 2B). A cue displayed at the beginning of the trial revealed to the animal the task for that trial. Offers varied from trial to trial, and we indicate the quantities offered in any given trial with q_A and q_B. An ‘offer type’ was defined by two quantities [q_A, q_B], and the same offer types were used for the two tasks in each session. For Task 2, trials in which juice A was offered first and trials in which juice B was offered first are referred to as ‘AB trials’ and ‘BA trials’, respectively. The first and second offers are referred to as ‘offer1’ and ‘offer2’, respectively.

Figure 2

Download asset Open asset

The data set included 241 sessions (101 from monkey J, 140 from monkey G; see Methods). Sessions lasted for 217–880 trials (mean ± std = 589 ± 160). For each session, we analyzed choices in the two tasks separately using probit regressions. For Task 1 (simultaneous offers), we used the following model:

where Choice B = 1 if the animal chose juice B and 0 otherwise, Φ was the cumulative function of the standard normal distribution, and q_A and q_B were the quantities of juices offered on any given trial. From the fitted parameters a₀ and a₁, we derived measures for the relative value of the two juices ρ_Task1 = exp(−a₀/a₁) and for the sigmoid steepness η_Task1 = a₁. Intuitively, the relative value was the quantity ratio q_B/q_A that made the animal indifferent between the two juices, and the sigmoid steepness was inversely related to choice variability.

For Task 2 (sequential offers), we used the following model:

where δ_order,AB = 1 for AB trials and 0 otherwise, and δ_order,BA = 1 − δ_order,AB. In essence, AB trials and BA trials were analyzed separately but assuming that the two sigmoids were parallel. From the fitted parameters a₂, a₃, and a₄, we derived measures for the relative value of the two juices ρ_Task2 = exp(−a₂/a₃), for the sigmoid steepness η_Task2 = a₃, and for the order bias ε = 2 ρ_Task2 a₄/a₃. Intuitively, the order bias was a bias favoring the first or the second offer. Specifically, ε < 0 indicated a bias favoring offer1; ε > 0 indicated a bias favoring offer2. We also defined relative values specific to AB trials and BA trials as ρ_AB = exp(−(a₂ + a₄)/a₃) and ρ_BA = exp(−(a₂ − a₄)/a₃). Of note, the order bias was defined such that

The experimental design gave us the opportunity to compare choices across tasks independently of factors such as selective satiation or changes in the internal state. The relative values measured in the two tasks were highly correlated (Figure 2E, F). At the same time, our analyses revealed three interesting phenomena. First, for both animals, sigmoids measured in Task 2 were significantly shallower compared to Task 1 (Figure 2G, H). In other words, presenting offers in sequence reduced choice accuracy. Second, in Task 2, both animals showed a consistent order bias favoring offer2 (Figure 2I, J). Third, in both animals, relative values in Task 2 were significantly higher than in Task 1 (ρ_Task2 > ρ_Task1), and this effect increased with the relative value (Figure 2E, F). In other words, the ellipse marking the 90% confidence interval for the joint distribution of relative values laid above the identity line and was rotated counterclockwise compared to the identity line.

To further investigate the differences in relative values measured across tasks, we quantified them separately in AB trials and BA trials in each monkey. We thus examined the relation between ρ_Task1 and ρ_Task2,AB and, separately, that between ρ_Task1 and ρ_Task2,BA (Figure 3). In both animals and in both sets of trials, the ellipse marking the 90% confidence interval was rotated counterclockwise compared to the identity line. Furthermore, the ellipse measured for BA trials was higher than that for AB trials. We quantified these observations with an analysis of covariance (ANCOVA) using the presentation order (AB and BA) as a covariate and imposing parallel lines (Figure 3C, F). In both animals, the two regression lines were significantly distinct (difference in intercept >0, p ≤ 0.002 in each animal). This result confirmed the presence of an order bias favoring offer2 in Task 2. Concurrently, in both animals the regression slope was significantly >1 (p ≤ 0.04 in each animal; ellipse rotation). This result indicated that the animals had an additional bias favoring juice A in Task 2, and that this bias increased as a function of the relative value ρ. We refer to this phenomenon as the preference bias.

Figure 3

Download asset Open asset

The following sections present a series of results on the neuronal origins of these behavioral phenomena. We begin by discussing the computational framework for the analyses.

Economic choice is thought to entail two stages: values are assigned to the available offers and a decision is made by comparing values. Importantly, in our tasks and in most circumstances, choices elicit an ensemble of mental operations taking place before, during, and after the computation and comparison of offer values. Upstream of valuation, choices examined here entail the sensory processing of visual stimuli and the retrieval from memory of relevant information (e.g., the association between color and juice type). Downstream of value comparison, the decision outcome must guide a suitable motor response. In addition, performance in Task 2 requires holding in working memory the value of offer1 until offer2, remembering the decision outcome for an additional delay, and mapping that outcome onto the appropriate saccade target (Figure 2B). In principle, choice biases could emerge at any of these computational stages. Likewise, each of these mental operations could be noisy and thus contribute to choice variability.

Neuronal activity in OFC does not capture all of these processes. However, previous work indicates that neurons in this area participate both in value computation and value comparison. In the framework proposed here (Figure 1), sensory and limbic areas feed offer value cells, where values are integrated. In turn, offer value cells provide the primary input to a neural circuit constituted by chosen juice cells and chosen value cells, where decisions are formed. Finally, the decision circuit is connected with downstream areas, such as lateral prefrontal cortex, engaged in working memory and in transforming choice outcomes into suitable action plans. This scheme reflects the anatomical connectivity of OFC and other prefrontal regions (Carmichael and Price, 1995a; Carmichael and Price, 1995b; Petrides and Pandya, 2006; Saleem et al., 2014; Takahara et al., 2012); it is motivated by neurophysiology results from OFC (Ballesta et al., 2020; Rich and Wallis, 2016) and connected areas (Cai and Padoa-Schioppa, 2014; Sasikumar et al., 2018); and it is consistent with computational models of economic decisions (Friedrich and Lengyel, 2016; Rustichini and Padoa-Schioppa, 2015; Solway and Botvinick, 2012; Song et al., 2017; Yim et al., 2019; Zhang et al., 2018).

Of note, both offer value and chosen value cells encode subjective values. However, in the framework of Figure 1, offer value cells express a pre-decision value, while chosen value cells express a value emerging during the decision process. Conversely, the activity of chosen juice cells captures the evolving commitment to a particular choice outcome. In this framework, suitable analyses of neuronal activity may reveal whether particular choice biases emerge during valuation, during value comparison, or in subsequent computational stages.

Other things equal, choices under sequential offers (Task 2) were significantly less accurate than choices under simultaneous offers (Task 1; Figure 2). We first investigated the neural origins of this phenomenon.

The primary data set examined in this study included 183 offer value cells, 160 chosen juice cells, and 174 chosen value cells (see Methods). Comparing neuronal responses across tasks, we noted that offer value signals in Task 2 were significantly weaker than in Task 1. Figure 4A, C illustrates one example cell. In both tasks, this neuron encoded the offer value B. However, the activity range (see Methods) measured in Task 2 was smaller than that measured in Task 1. This effect was also observed at the population level. For this analysis, we pooled offer value cells associated with juices A and B, and with positive or negative encoding (see Methods). For Task 1, we focused on the post-offer time window; for Task 2, we focused on post-offer1 and post-offer2 time windows, pooling trial types from both windows. For each cell, we imposed that the response be significantly tuned in these time windows in each task, and we quantified the mean activity and the activity range (Δr, see Methods). At the population level, the mean activity did not differ significantly across tasks (p = 0.6, t-test; p = 0.4, Wilcoxon test, Figure 4D). In contrast, the activity range was significantly lower in Task 2 compared to Task 1 (Δr_Task2 < Δr_Task1; p = 0.06, t-test; p = 0.02, Wilcoxon test Figure 4E). In other words, offer value signals were weaker in Task 2 compared to Task 1.

Figure 4 with 1 supplement see all

Download asset Open asset

The activity of offer value cells is causally related to choices (Ballesta et al., 2020). Furthermore, for given value range and mean activity, the activity range determines the neuronal signal-to-noise ratio. Indeed, we previously found that decreases in the encoding slope of offer value cells due to range adaptation reduce choice accuracy (Conen and Padoa-Schioppa, 2019; Rustichini et al., 2017). Along similar lines, we inquired whether the difference in choice accuracy measured across tasks (Figure 2G, H) might be explained, at last partly, by differences in neuronal activity range (Figure 4E). We thus examined the relation between the difference in sigmoid steepness (Δη = η_Task2 − η_Task1) and the difference in activity range (Δr = Δr_Task2 − Δr_Task1). The two measures were positively correlated (Spearman r = 0.2, p = 0.01; Pearson r = 0.3, p = 0.003; Figure 4F). In other words, the drop in choice accuracy observed in Task 2 compared to Task 1 correlated with weaker offer value signals. A similar analysis of chosen value cells found that the activity range Δr was reduced in Task 2 compared to Task 1. However, this effect and the drop in choice accuracy were not significantly correlated (Figure 4—figure supplement 1).

In conclusion, the lower choice accuracy measured in Task 2 compared to Task 1 correlated with weaker offer value signals in OFC. Thus, this behavioral phenomenon emerged, at least partly, during valuation.

The next series of analyses focused on the neural origins of the order bias (ε). Since this phenomenon pertains only to choices under sequential offers, we included in the analyses an additional data set recorded in the same animals performing only Task 2 (see Methods).

In the framework of Figure 1, we first inquired whether the order bias emerged during valuation. If this was the case, for any given good, offer value cells should encode a higher value when the good is presented as offer2. To test this hypothesis, we pooled offer value cells associated with the two juices. For each neuron, ‘E’ indicated the juice encoded by the cell and ‘O’ indicated the other juice. We thus refer to EO trials and OE trials. For any given cell, we compared the response recorded in EO trials (post-offer1 time window) with the response recorded in OE trials (post-offer2 time window). If the order bias emerged during valuation, the mean activity and/or the activity range should be higher for the latter (Figure 5—figure supplement 1A). Contrary to this prediction, across a population of 128 cells, we did not find any systematic difference in mean activity or activity range (Figure 5—figure supplement 1B, C). Furthermore, the difference between the activity parameters measured in OE and EO trials did not correlate with the order bias (Figure 5—figure supplement 1D). In conclusion, assigned values did not depend on the presentation order.

We next examined whether the order bias emerged during value comparison. If so, the bias should be reflected in the activity of both chosen juice and chosen value cells (Figure 1). For chosen value cells, the hypothesis might be tested noting that in post-offer1 and post-offer2 time windows these neurons encoded the value currently offered independently of the juice type (Table 1). Thus, the activity measured in these time windows in AB and BA trials provided neuronal measures for the relative values of the two juices. More specifically, for each chosen value cell, we derived the two measures ρ^neuronal_AB and ρ^neuronal_BA for AB trials and BA trials, respectively (Figure 5A; see Methods). We also defined the difference Δρ^neuronal = ρ^neuronal_BA − ρ^neuronal_AB. We recall that the order bias (ε) was essentially equal to the difference between the relative values measured behaviorally in BA and AB trials (Equation 3). Thus, assessing whether the activity of chosen value cells reflected the order bias amounts to testing the relation between Δρ^neuronal and ε.

Figure 5 with 1 supplement see all

Download asset Open asset

We conducted a population analysis of 96 chosen value cells. Confirming previous results (Padoa-Schioppa and Assad, 2006), neuronal and behavioral measures of relative value were highly correlated. Similarly, the two neuronal measures of relative value, ρ^neuronal_AB and ρ^neuronal_BA, were correlated with each other (Figure 5B). Most importantly, the difference Δρ^neuronal and the order bias ε were significantly correlated across the population (Spearman r = 0.3, p = 0.007; Pearson r = 0.2, p = 0.02; Figure 5C). Hence, session-to-session fluctuations in the activity of chosen value cells correlated with fluctuations in the order bias.

Further insights on the order bias came from the analysis of chosen juice cells. Again, for each neuron, E and O indicated the juice encoded by the cell and the other juice, respectively. A previous study found that the baseline activity of chosen juice cells recorded in OE trials immediately before offer2 was negatively correlated with the value of offer1 (i.e., the value of the other juice) – a phenomenon termed circuit inhibition (Ballesta and Padoa-Schioppa, 2019). If the decision is conceptualized as the evolution of a dynamic system (Rustichini and Padoa-Schioppa, 2015; Wang, 2002), circuit inhibition sets the system’s initial conditions and is thus integral to value comparison. In this account, the evolving decision is essentially captured by the activity of chosen juice cells in OE trials, which reflects a competition between the negative offset set by the value of offer1 (initial condition) and the incoming signal encoding the value of offer2. If so, the intensity of circuit inhibition should be negatively correlated with the order bias.

We tested this prediction as follows. First, we replicated previous findings and confirmed the presence of circuit inhibition in our primary data set (Figure 6A). We then focused on a 300-ms time window starting 250 ms before offer2 onset. For each chosen juice cell, we regressed the firing rate against the normalized offer1 value (see Methods). Thus, the regression slope c₁ quantified circuit inhibition for individual cells. Across a population of 295 chosen juice cells, mean(c₁) was significantly <0 (p = 5 × 10⁻⁶, t-test; p = 9 × 10⁻⁸, Wilcoxon test; Figure 6B). Third, we examined the relation between circuit inhibition (c₁) and the order bias (ε). Confirming the prediction, the two measures were significantly correlated across the population (Spearman r = 0.1, p = 0.02; Pearson r = 0.1, p = 0.02; Figure 6C). In other words, stronger circuit inhibition (more negative c₁) corresponded to a weaker order bias (smaller ε).

Figure 6

Download asset Open asset

In conclusion, the order bias did not originate before or during valuation. Analysis of chosen juice cells and chosen value cells indicated that the order bias emerged during value comparison (decision stage).

When offers were presented sequentially (Task 2), both monkeys showed an additional preference bias that favored juice A and was more pronounced when the relative value of the two juices was larger (Figure 3). Our last series of analyses focused on the origins of this bias.

First, we inquired whether the preference bias emerged during valuation. If this was the case, one or both of the following should be true: (1) offer value A cells encoded higher values in Task 2 than in Task 1 and/or (2) offer value B cells encoded lower values in Task 2 than in Task 1. Furthermore, these putative effects should increase as a function of the relative value. To test these predictions, we examined the tuning functions of offer value cells. For each cell group (offer value A, offer value B), we pooled neurons with positive and negative encoding. For Task 1, we focused on the post-offer time window; for Task 2, we focused on post-offer1 and post-offer2 time windows, pooling trial types from both windows. Indicating with b₀ and b₁ the tuning intercept and tuning slope (see Methods, Equation 8), we computed the difference in intercept Δb₀ = b_0,Task2 − b_0,Task1 and the difference in slope Δb₁ = b_1,Task2 − b_1,Task1 for each cell. We then examined the relation between these measures and the relative value ρ across the population, separately for each cell group. Contrary to the prediction, we did not find any correlation between neuronal measures (Δb₀, Δb₁) and the behavioral measure (ρ) for either offer value A or offer value B cells (Figure 7—figure supplement 1). Thus, the preference bias did not seem to emerge at the valuation stage.

We next examined chosen value cells. As discussed above, their activity provided a neuronal measure for the relative value (ρ^neuronal), which reflected the internal subjective values of the juices emerging during value comparison. In principle, ρ^neuronal might differ from the relative value derived from choices through the probit regression (ρ^behavioral) because choices might be affected by systematic biases originating downstream of value comparison (Figure 1). In the light of this consideration, we examined the relation between the neuronal measure of relative value in Task 2 (ρ^neuronal_Task2, see Methods) and the behavioral measures obtained in the two tasks (ρ^behavioral_Task1, ρ^behavioral_Task2). We envisioned two possible scenarios (Figure 7A). In scenario 1, the preference bias reflected a difference in values across tasks. In other words, the subjective values of the juices in the two tasks were different and such that the relative value of juice A was higher in Task 2 than in Task 1. If so, ρ^neuronal_Task2 should be statistically indistinguishable from ρ^behavioral_Task2 and systematically larger than ρ^behavioral_Task1. In scenario 2, the subjective values of the juices were the same in both tasks and the preference bias reflected some neuronal process taking place downstream of value comparison. If so, ρ^neuronal_Task2 should be statistically indistinguishable from ρ^behavioral_Task1 and systematically smaller than ρ^behavioral_Task2.

Figure 7 with 1 supplement see all

Download asset Open asset

The results of our analysis clearly conformed with scenario 2 (Figure 7B). For each chosen value cell, we computed ρ^neuronal_Task1 in the post-offer time window and ρ^neuronal_Task2 in the post-offer2 time window. Across the population, the two measures were statistically indistinguishable (p = 0.3, t-test; not shown). We then regressed ρ^neuronal_Task2 onto ρ^behavioral_Task1. The linear relation between these measures was statistically indistinguishable from identity. Separately, we regressed ρ^neuronal_Task2 onto ρ^behavioral_Task2. In this case, the regression slope was significantly <1 (p = 0.02). This result is quite remarkable. It shows that the chosen value represented in the brain in Task 2 was equal to that inferred from choices in Task 1, and significantly different from that inferred from choices in Task 2. This fact implies that the preference bias was costly for the monkey, as it reduced the value obtained on average at the end of each trial (see Discussion).

In summary, the preference bias did not reflect differences in the values assigned to individual offers (offer values). Furthermore, insofar as the activity of chosen value cells reflects the decision process (Figure 1), the preference bias did not seem to emerge during value comparison. So how can one make sense of this behavioral phenomenon? At the cognitive level, the preference bias might be interpreted as due to the higher demands of Task 2. When the two saccade targets appeared on the monitor, information about values was no longer on display (Figure 2B). If at that point the animal had not finalized its decision, or if it had failed to retain in working memory the decision outcome, the animal might have selected the target associated with the better juice (juice A). Such bias would have been especially strong when the value difference between the two juices was large. In this view, the preference bias would reflect a ‘second thought’ occurring after value comparison, in some trials.

To test this intuition, we turned to the activity of chosen juice cells. As noted above, in Task 2, the evolving decision was captured by the activity of these neurons recorded in OE trials immediately before and after offer2 onset (Figure 8A). More specifically, the state of the ongoing decision was captured by the distance between the two traces corresponding to the two possible choice outcomes (E chosen, O chosen). For any neuron, we quantified this distance with a receiver operating characteristic (ROC) analysis, which provided a choice probability (CP). In essence, CP can be interpreted as the probability with which an ideal observer may guess the eventual choice outcome based on the activity of the cell. For each chosen juice cell, we computed the CP at different times in the trial. Across the population, mean(CP) exceeded chance level starting shortly before offer2, consistent with the above discussion on circuit inhibition. We then proceeded to investigate the origins of the preference bias.

Figure 8

Download asset Open asset

We reasoned that, at the net of noise in measurements and cell-to-cell variability, CPs ultimately quantify the animal’s commitment to the eventual choice outcome. If the preference bias emerged late in the trial – perhaps after target presentation, if animals had not already finalized their decision – the intensity of the preference bias should be inversely related to the animals’ commitment to the eventual choice outcome measured earlier in the trial. In other words, there should be a negative correlation between the preference bias and CPs computed at the time when decisions normally take place (shortly before or after offer2 onset). Our analyses supported this prediction. To quantify the preference bias intensity independent of the juice pair, we defined the preference bias index (PBI) = 2 (ρ_Task2 − ρ_Task1)/(ρ_Task2 + ρ_Task1). We then focused on four 250-ms time windows before offer1 (control window), before and after offer2 onset, and before juice delivery (Figure 8B–E). Confirming our predictions, CP and PBI were significantly anti-correlated immediately before and during offer2 presentation, but not in the control time window or late in the trial (Figure 8F–I).

In conclusion, our results indicated that the preference bias did not emerge during valuation or during value comparison. Conversely, our results suggest that the preference bias emerged late in the trial, as a ‘second thought’ process that guided choices when decisions were not finalized based on offer values alone.

Early economists proposed that choices between goods entail the computation and comparison of subjective values. However, the concept of value is somewhat slippery, because values relevant to choices cannot be measured behaviorally other than from choices themselves. This circularity problem haunted generations of scholars, dominating academic debates in the 19th and 20th centuries. In the end, neoclassic economic theory came to reject (cardinal) values and to rely only on (ordinal) preferences (Niehans, 1990; Samuelson, 1947). In other words, standard economics is agnostic as to whether subjective values are computed at all. The construction of standard economic theory was a historic success, but it came at a cost: the theory cannot explain a variety of biases observed in human choices (Camerer et al., 2003; Kahneman and Tversky, 2000; Lichtenstein and Slovic, 2006). In this perspective, neuroscience results showing that neuronal activity in multiple brain regions is linearly related to values defined behaviorally (Amemori and Graybiel, 2012; Cai et al., 2011; Cai and Padoa-Schioppa, 2012; Hosokawa et al., 2013; Jezzini and Padoa-Schioppa, 2020; Kable and Glimcher, 2007; Kim et al., 2008; Lak et al., 2014; Levy et al., 2010; Louie and Glimcher, 2010; Padoa-Schioppa and Assad, 2006; Pastor-Bernier et al., 2019; Plassmann et al., 2007; Shenhav and Greene, 2010), constitute a significant breakthrough. They validate the concept of value and effectively break the circularity surrounding it. Indeed, a neuronal population whose activity is reliably correlated with values measured from choices (behavioral values) may be used to derive independent measures of subjective values (neuronal values). In most circumstances, neuronal values and behavioral values should be (and are) indistinguishable. However, in specific choice contexts, the two measures might differ somewhat. When observed, such discrepancies indicate that choices are partly determined by processes that escape the maximization of offer values. If so, suitable analyses of neuronal activity may be used to assess the origins of particular choice biases.

These considerations motivated the analyses conducted in this study. In our experiments, monkeys chose between two juices offered simultaneously or sequentially. Choices under sequential offers were less accurate, biased in favor of the second offer (order bias), and biased in favor of the preferred juice (preference bias). It is generally understood that good-based economic decisions take place in OFC (Cisek, 2012; Padoa-Schioppa, 2011; Rushworth et al., 2012) and that the encoding of decision variables in this area is categorical in nature (Hirokawa et al., 2019; Onken et al., 2019; Padoa-Schioppa, 2013). Earlier studies had identified in OFC three groups of neurons encoding individual offer values, the chosen juice and the chosen value. Furthermore, choices under simultaneous or sequential offers were found to engage the same groups of cells (Shi et al., 2022a). Notably, the variables encoded in OFC capture both the input and the output of the decision process. This observation and a series of experimental and theoretical results lead to the hypothesis that the cell groups identified in OFC constitute the building blocks of a decision circuit (Padoa-Schioppa and Conen, 2017). More specifically, we hypothesize that offer value cells provide the primary input to a circuit formed by chosen juice cells and chosen value cells, where decisions are formed. In this view, different cell groups in OFC may be associated with different computational stages: offer value cells instantiate the valuation stage; chosen value cells reflect values possibly modified by the decision process; and chosen juice cells capture the evolving commitment to a particular choice outcome. In this framework, we examined the activity of each cell group in relation to each behavioral phenomenon.

Our results may be summarized as follows. (1) Other things equal, neuronal signals encoding the offer values were weaker (smaller activity range) under sequential offers than under simultaneous offers. The reason for this discrepancy is unclear, but this neuronal effect was correlated with the difference in choice accuracy measured at the behavioral level. In other words, the drop in choice accuracy observed under sequential offers originated, at least partly, at the valuation stage. (2) The order bias did not correlate with any measure in the activity of offer value cells. However, the order bias was negatively correlated with circuit inhibition in chosen juice cells – a phenomenon seen as key to value comparison (Ballesta and Padoa-Schioppa, 2019). Furthermore, session-to-session fluctuations in the order bias correlated with fluctuations in the neuronal measure of relative value derived from chosen value cells. These findings indicate that the order bias emerged during value comparison. (3) The preference bias did not have any correlate in the activity of offer value cells or chosen value cells. Moreover, the preference bias was inversely related to a measure derived from chosen juice cells and quantifying the degree to which the decision was finalized when offer values are ‘normally’ compared (i.e., upon presentation of the second offer). These findings indicate that the preference bias emerged late in the trial. As a caveat, the hypothesis discussed above, linking different cell groups in OFC to specific decision stages, awaits further confirmation.

Two of our findings are particularly relevant to the distinction between behavioral values and neuronal values. First, the activity of offer value cells did not present any difference associated with the presentation order or with the juice preference. Second, relative values derived from chosen value cells under sequential offers differed significantly from behavioral measures obtained in the same task, and were indistinguishable from behavioral measures obtained in the other task (simultaneous offers). Thus, the order bias and the preference bias highlighted significant differences between neuronal and behavioral measures of value. These observations imply that the order bias and the preference bias emerged downstream of valuation. Importantly, they also imply that the two choice biases imposed a cost to the animals, in the sense that they reduced the (neuronal) value obtained on average in any given trial. Notably, it would be impossible to draw such conclusion based on choices alone.

To our knowledge, this is the first study to investigate the origins of choice biases building on the distinction between behavioral values and neuronal values. At the same time, some of our results are not unprecedented. Earlier work showed that human and animal choices are affected by a bias favoring, on any given trial, the same good chosen in the previous trial (Alós-Ferrer et al., 2016; Goodwin, 1977; Padoa-Schioppa, 2013; Schoemann and Scherbaum, 2019; Senftleben et al., 2021). The origins of this phenomenon, termed choice hysteresis, are hard to pinpoint based on behavioral evidence alone. However, previous analysis of neuronal activity in OFC revealed that choice hysteresis is not reflected in the encoding of offer values. Conversely, choice hysteresis correlates with fluctuations in the baseline activity of chosen juice cells, which is partly influenced by the previous trial’s outcome (Padoa-Schioppa, 2013). Thus, similar to the order bias, choice hysteresis appears to emerge at the decision stage.

We have noted that the three behavioral phenomena described here were detrimental to the animals. This point bears a few comments.

Let us consider, in a very general sense, choices between two goods A and B, taking place in two possible conditions 1 and 2. We may refer to the subjective values of the two goods in the two conditions as V_A,1, V_A,2, V_B,1, and V_B,2. Lets also assume the presence of a choice bias such that in condition 1 the animal consistently chooses A, while in condition 2 it consistently chooses B. In broad strokes, that might be for two reasons. Either (1) values differ across conditions such that V_A,1 > V_B,1 and V_A,2 < V_B,2, or (2) the value of each good remains unchanged across conditions (V_A,1 = V_A,2 and V_B,1 = V_B,2) and choices are affected by some other process, downstream of value computation. If so, in one of the two conditions the animal consistently chooses the lower value. In other words, the choice bias is detrimental to the animal. Coming to our experiments, the analysis of neuronal activity indicates that the subjective values of offered juices were independent of the choice task, and independent of the presentation order in Task 2. Thus, scenario (2) held true with respect to both the order bias and the preference bias. Consequently, both biases were detrimental to the animals.

With respect to the preference bias, one question is whether the bias affected choices in Task 1 or in Task 2 (in principle, there could be a bias favoring the unpreferred juice in Task 1). The fact that ρ^behavioral_Task1, ρ^neuronal_Task1, and ρ^neuronal_Task2 were all indistinguishable from each other while ρ^behavioral_Task2 differed from them (Figure 7) argues for the latter understanding.

It is interesting to speculate whether the choice biases documented here might benefit the animal in some more general sense. For example, one might wonder whether the cost imposed by the preference bias was lower than the metabolic cost the monkey would have incurred to increase its performance level and avoid that bias. If so, the preference bias would be, in fact, ecologically adaptive. Addressing this question would require quantifying the metabolic cost of increasing performance in the same value units used for the juices – a challenge open for future studies. However, independent of that assessment, our present results indicate that the putative metabolic cost of increasing performance in the task did not explicitly enter the decision process. If metabolic cost affected behavior, it did so in a meta-decision sense.

Similar considerations hold for the difference in choice accuracy measured across tasks. The fact that sigmoid functions are not infinitely steep (i.e., the presence of choice variability) means that on some trials the animal chooses the lower value. In fact, one can quantify the loss in expected payoff as a function of the sigmoid steepness (Constantinople et al., 2019; Rustichini et al., 2017). That sigmoids were shallower in Task 2 means that the average payoff was lower in that task – a detriment to the animal. Again, it is interesting to speculate whether weaker offer value signals recorded in Task 2 might also benefit the animal in some way, perhaps by reducing cognitive or metabolic costs. This question remains open for future studies. Importantly, such costs did not explicitly enter the decision process; if they affected behavior, they did so in a meta-decision sense.

The past two decades have witnessed a lively interest for the neural underpinnings of choice behavior. In this effort, a significant breakthrough came from the adoption of behavioral paradigms inspired by the economics literature, in which subjective values derived from choices are used to interpret neural activity. Without renouncing this approach, here we took a further step, showing that the decision process sometimes falls short of selecting the maximum offer value, and that choices are sometimes affected by processes taking place downstream of value comparison. In other words, behavioral values and neuronal values sometimes differ. These results might seem uncontroversial, but they have deep implications for economic theory and beyond. Looking forward, the framework developed here, in which the computation and comparison of offer values are central, but choices can also be affected by other processes accessible through neuronal measures, may help understand the origins of other choice biases.

This study presents new analyses of published data (Shi et al., 2022a). Experimental procedures for surgery, behavioral control, and neuronal recordings have been described in detail. Briefly, two male monkeys (Macaca mulatta; monkey J, 10.0 kg, 8 years old; monkey G, 9.1 kg, 9 years old) participated in the study. Under general anesthesia, we implanted on each animal a head restraining device and an oval chamber (axes 50 × 30 mm) allowing bilateral access to OFC. During the experiments, monkeys sat in an electrically insulated environment with their head fixed and a computer monitor placed at 57 cm distance. The gaze direction was monitored at 1 kHz using an infrared video camera (Eyelink, SR Research). Behavioral tasks were controlled through custom written software based on Matlab (v2016a; MathWorks Inc). The code is available online (Hwang et al., 2019; Shi et al., 2022b; https://monkeylogic.nimh.nih.gov).

In each session, the animal chose between two juices labeled A and B (A preferred) offered in variable amounts. Trials with two choice tasks, referred to as Task 1 and Task 2, were pseudorandomly interleaved. In both tasks, offers were represented by sets of colored squares displayed on the monitor. For each offer, the color indicated the juice type and the number of squares indicated the quantity. Each trial began with the animal fixating a large dot. After 0.5 s, the initial fixation point changed to a small dot or a small cross; the new fixation point cued the animal to the choice task used in that trial. In Task 1 (Figure 2A), cue fixation (0.5 s) was followed by the simultaneous presentation of the two offers. After a randomly variable delay (1–1.5 s), the center fixation point disappeared and two saccade targets appeared near the offers (go signal). The animal indicated its choice with an eye movement. It maintained peripheral fixation for 0.75 s, after which the chosen juice was delivered. In Task 2 (Figure 2B), cue fixation (0.5 s) was followed by the presentation of one offer (0.5 s), an interoffer delay (0.5 s), presentation of the other offer (0.5 s), and a wait period (0.5 s). Two colored saccade targets then appeared on the two sides of the fixation point. After a randomly variable delay (0.5–1 s), the center fixation point disappeared (go signal). The animal indicated its choice with a saccade, maintained peripheral fixation for 0.75 s, after which the chosen juice was delivered. Central and peripheral fixation were imposed within 4–6 and 5–7 degrees of visual angle, respectively. Aside from the initial cue, the choice tasks were nearly identical to those used in previous studies (Ballesta and Padoa-Schioppa, 2019; Padoa-Schioppa and Assad, 2006).

For any given trial, q_A and q_B indicate the quantities of juices A and B offered to the animal, respectively. An ‘offer type’ was defined by two quantities [q_A q_B]. On any given session, we used the same juices and the same sets of offer types for the two tasks. For Task 1, the spatial configuration of the offers varied randomly from trial to trial. For Task 2, the presentation order varied pseudorandomly and was counterbalanced across trials for any offer type. The terms ‘offer1’ and ‘offer2’ indicated, respectively, the first and second offer, independently of the juice type and amount. Trials in which juice A was offered first and trials in which juice B was offered first were referred as ‘AB trials’ and ‘BA trials’, respectively. The spatial location (left/right) of saccade targets varied randomly. The juice volume corresponding to one square (quantum) was set equal for the two choice tasks and remained constant within each session. It varied across sessions between 70 and 100 μl for both monkeys. The association between the initial cue (small dot, small cross) and the choice task varied across sessions in blocks. Across sessions, we used 12 different juices (and colors) and 45 different juice pairs. Based on a power analysis, in most sessions the number of trials for Task 2 was set equal to 1.5 times that for Task 1.

Neuronal recordings were guided by structural MRI scans (1 mm sections) obtained before and after surgery and targeted area 13 m (Ongür and Price, 2000). We recorded from both hemispheres in both monkeys. Tungsten single electrodes (100 µm shank diameter; FHC) were advanced remotely using a custom-built motorized microdrive. Typically, one motor advanced two electrodes placed 1 mm apart, and 1–2 such pairs of electrodes were advanced unilaterally or bilaterally in each session. Neural signals were amplified (gain: 10,000) bandpass filtered (300 Hz to 6 kHz; Lynx 8, Neuralynx), digitized (frequency: 40 kHz) and saved to disk (Power 1401, Cambridge Electronic Design). Spike sorting was performed offline (Spike2, v6, Cambridge Electronic Design). Only cells that appeared well isolated and stable throughout the session were included in the analysis.

In each session, choice patterns were analyzed using probit regressions as described in the main text (Equations 1 and 2). For convenience, we repeat here the equation only for Task 1.

Here, Φ indicates the cumulative function of the standard normal distribution. This model is referred to as the ‘log value ratio’ model. For Task 1 (simultaneous offers), the probit fit provided measures for the relative value ρ_Task1 and the sigmoid steepness η_Task1. For Task 2 (sequential offers), the probit fit provided measures for the relative value ρ_Task2, the sigmoid steepness η_Task2, and the order bias ε. Subsequent analyses of neuronal activity relied on these behavioral measures.

To test the robustness of our findings, we conducted a series of control analyses. First, we fitted a probit using a ‘value difference’ model, defined as follows:

Second, we fitted a logit using a log value ratio model:

Third, we fitted a logit using a value difference model:

Each of these fit provided measures for each of the parameters characterizing choices in the two tasks (ρ_Task1, ρ_Task2, etc.). For each session and for each model we obtained an R². We then compared different models by computing the distribution of BIC across sessions for each pair of models. We generally found that log value ratio models provided a better fit compared to value difference models, consistent with theoretical considerations (Padoa-Schioppa, 2022). We also found that logit models provided a better fit compared to probit models, although measures of relative value, sigmoid steepness, and order bias were very similar and highly correlated. For consistency with previous studies, we report the results of neuronal analyses based on neuronal measures derived from Equations 1 and 2. However, all our results held true using measures derived from logit regressions.

Notably, Equation 2 describes two parallel sigmoids. In a control analysis, we relaxed this assumption and fitted choices in AB and BA trials with two independent sigmoids. Analyzing neuronal activity based on measures derived from this analysis did not substantially alter any of the results.

Finally, we defined the order bias as ε = 2 ρ_Task2 a₄/a₃. This definition is particularly convenient for the present analyses because ε equals the difference ρ_BA – ρ_AB (Equation 3). Alternative and valid definitions include ε = a₄ and ε = a₄/a₃. Control analyses showed that using these definitions did not substantially alter any of the results.

The present analyses build on the results of a previous study showing that both choice tasks engage the same groups of neurons in OFC (Shi et al., 2022a). Here, we briefly summarize those findings.

The original data set included 1526 neurons (672 from monkey J, 854 from monkey G) recorded in 306 sessions (115 from monkey J, 191 from monkey G). For each neuron, trials from Task 1 and Task 2 were first analyzed separately using the procedures developed in previous studies (Ballesta and Padoa-Schioppa, 2019; Padoa-Schioppa and Assad, 2006). For Task 1, we defined four time windows: post-offer (0.5 s after offer onset), late-delay (0.5–1 s after offer onset), pre-juice (0.5 s before juice onset), and post-juice (0.5 s after juice onset). A ‘trial type’ was defined by two offered quantities and a choice. For Task 2, we defined three time windows: post-offer1 (0.5 s after offer1 onset), post-offer2 (0.5 s after offer2 onset) and post-juice (0.5 s after juice onset). A ‘trial type’ was defined by two offered quantities, their order and a choice. For each task, each trial type and each time window, we averaged spike counts across trials. A ‘neuronal response’ was defined as the firing rate of one cell in one time window as a function of the trial type. Neuronal responses in each task were submitted to an analysis of variance (factor: trial type). Neurons passing the p < 0.01 criterion in ≥1 time window in either task were identified as ‘task-related’ and included in subsequent analyses.

Following earlier work (Padoa-Schioppa, 2013), neurons in Task 1 were classified in one of four groups offer value A, offer value B, chosen juice, or chosen value. Each variable could be encoded with positive or negative sign, leading to a total of eight cell groups. Each neuronal response was regressed against each of the four variables. If the regression slope b₁ differed significantly from zero (p < 0.05), the variable was said to ‘explain’ the response. In this case, we set the signed R² as sR² = sign(b₁) R²; if the variable did not explain the response, we set sR² = 0. After repeating the operation for each time window, we computed for each cell the sum(sR²) across time windows. Neurons explained by at least one variable in one time window, such that sum(sR²) ≠ 0, were said to be tuned; other neurons were labeled ‘untuned’. Tuned cells were assigned to the variable and sign providing the maximum |sum(sR²)|, where |·| indicates the absolute value. Thus, indicating with ‘+’ and ‘−’ the sign of the encoding, each neuron was classified in one of nine groups: offer value A+, offer value A−, offer value B+, offer value B−, chosen juice A, chosen juice B, chosen value+, chosen value−, and untuned.

Neuronal classification in Task 2 followed the procedures described in a previous study (Ballesta and Padoa-Schioppa, 2019). Under sequential offers, neuronal responses in OFC were found to encode different variables defined in relation to the presentation order (AB or BA). Specifically, the vast majority of responses were explained by one of 11 variables including 1 binary variable capturing the presentation order (AB | BA), 6 variables representing individual offer values (offer value A | AB, offer value A | BA, offer value B | AB, offer value B | BA, offer value 1, and offer value 2), 3 variables capturing variants of the chosen value (chosen value, chosen value A, and chosen value B), and a binary variable representing the binary choice outcome (chosen juice). Each of these variables could be encoded with a positive or negative sign. Most neurons encoded different variables in different time windows. In principle, considering 11 variables, 2 signs of the encoding and 3 time windows, neurons might present a very large number of variable patterns across time windows. However, the vast majority of neurons presented one of eight patterns referred to as ‘sequences’. Classification proceeded as follows. For each cell and each time window, we regressed the neuronal response against each of the variables predicted by each sequence. If the regression slope b₁ differed significantly from zero (p < 0.05), the variable was said to explain the response and we set the signed R² as sR² = sign(b₁) R²; if the variable did not explain the response, we set sR² = 0. After repeating the operation for each time window, we computed for each cell the sum(sR²) across time windows for each of the eight sequences. Neurons such that sum(sR²) ≠ 0 for at least one sequence were said to be tuned; other neurons were untuned. Tuned cells were assigned to the sequence that provided the maximum |sum(sR²)|. As a result, each neuron was classified in one of nine groups: seq #1, seq #2, seq #3, seq #4, seq #5, seq #6, seq #7, seq #8, and untuned (Table 1).

The results of the two classifications were compared using analyses for categorical data. In essence, we found a strong correspondence between the cell classes identified in the two choice tasks (Shi et al., 2022a). Hence, we may refer to the different groups of cells using the standard nomenclature – offer value, chosen juice, and chosen value – independently of the choice task. Based on this result, we proceeded with a comprehensive classification based on the activity recorded in both choice tasks. For each task-related cell, we calculated the sum(sR²) for the eight variables in Task 1 (sum(sR²)_Task1) and eight sequences in Task 2 (sum(sR²)_Task2) as described above. We then added the corresponding sum(sR²)_Task1 and sum(sR²)_Task2 to obtain the final sum(sR²)_final. Neurons such that sum(sR²)_final ≠ 0 for at least one class were said to be tuned; other neurons were untuned. Tuned cells were assigned to the cell class that provided the maximum |sum(sR²)_final|.

In some sessions, one or both choice patterns presented complete or quasicomplete separation – that is, the animal split choices for <2 offer types in Task 1 and/or in Task 2. In these cases, the probit regression did not converge, the resulting steepness η was high and unstable, and the relative value was not unique. This issue affected the classification analyses described above only marginally, but for the present study it was critical that behavioral measures be accurate and precise. We thus restricted our analyses to stable sessions by imposing an interquartile criterion on the sigmoid steepness (Tukey, 1977). Defining IQR as the interquartile range, values below the first quartile minus 1.5*IQR or above the third quartile plus 1.5*IQR were identified as outliers and excluded. Thus, our entire data set included 1204 neurons (577 from monkey J, 627 from monkey G) recorded in 241 sessions (101 from monkey J, 140 from monkey G). In this population, the classification procedures identified 183 offer value cells, 160 chosen juice cells, and 174 chosen value cells. These neurons constitute the primary data set for this study.

Most of our analyses compared choices and neuronal activity across tasks and were restricted to the primary data set. However, some analyses included only trials from Task 2 and quantified the effects due to the presentation order (AB vs. BA). In these analyses, we included an additional data set recorded previously from the same two animals performing only Task 2 (Ballesta and Padoa-Schioppa, 2019). All the procedures for behavioral control and neuronal recording were essentially identical to those described above. Furthermore, behavioral analyses and inclusion criteria were identical to those used for the primary data set. The resulting data set included 1205 neurons (414 from monkey J, 791 from monkey G) recorded in 196 sessions (51 from monkey J, 145 from monkey G). In this population, the classification procedures identified 243 offer value cells, 182 chosen juice cells, and 187 chosen value cells. We refer to these neurons as the additional data set. Importantly, the order bias was also observed in these sessions (Ballesta and Padoa-Schioppa, 2019).

The interquartile criterion was also used to identify outliers in all the analyses conducted throughout this study. In practice, this criterion became relevant only for the analyses shown in Figure 6 and Figure 5—figure supplement 1, as indicated in the respective figure legends.

Several analyses compared the tuning functions recorded in the two tasks. Tuning functions were defined by the linear regression of the firing rate r onto the encoded variable S:

Regression coefficients b₀ and b₁ were referred to as tuning intercept and tuning slope, respectively. Positive and negative encoding corresponded to b₁ > 0 and b₁ < 0, respectively. We also defined the mean activity and the activity range as follows. Indicating with S_max the maximum value of S, the mean activity was defined as r_mean = b₀ + b₁S_max/2. The activity range was defined as Δr = |b₁S_max|, where |·| indicates the absolute value.

For any neuronal response, the tuning was considered significant if b₁ differed significantly from zero (p < 0.05) and if the sign of the encoding was consistent with the cell class (e.g., b₁ > 0 for offer value A + cells). All the analyses comparing tuning functions across tasks were restricted to neuronal responses with significant tuning.

Several analyses relied on neuronal measures for the relative value of the juices (ρ^neuronal) derived from the activity of chosen value cells. In Task 1, these neurons encode the chosen value independently of the juice type. For each neuronal response, we performed a bilinear regression:

where θ₀, θ_A, and θ_B were the regression coefficients, δ_choice,A = 1 if the animal chose juice A and 0 otherwise, and δ_choice,B = 1 δ_choice,A. If the response encodes the chosen value, θ_A should be proportional to the value of a quantum of juice A (uA), θ_B should be proportional to the value of a quantum of juice B (uB), and the ratio θ_A/θ_B should equal the value ratio – that is, the relative value of the two juices. We thus defined

Previous studies showed that this measure is statistically indistinguishable from the behavioral measure ρ^behavioral derived from the probit analysis of choice patterns (Padoa-Schioppa and Assad, 2006).

In Task 2, in the post-offer1 and post-offer2 time windows, chosen value cells encoded the value of the current offer, independent of the juice type (Table 1). For each neuron, we thus performed a bilinear regression for each of the two time windows:

where r₁ and r₂ were their responses recorded in the post-offer1 and post-offer2 time windows, respectively, and θ₁₀, θ_1A, θ_1B, θ₂₀, θ_2A, and θ_2B were regression coefficients. These coefficients provided four neuronal measures of relative value:

In essence, these four measures corresponded to the two time windows (post-offer1 and post-offer2) and to the two presentation orders (AB and BA). Importantly, all these measures were computed conditioned on θ_1A, θ_1B, θ_2A, and θ_2B differing significantly from zero (p < 0.05). The analyses illustrated in Figures 5 and 7 were restricted to neurons satisfying this criterion.

In terms of notation, we often omit the superscript in ρ^behavioral and we indicate behavioral measures simply as ρ (with the relevant subscripts). We use the superscript ‘behavioral’ only when we explicitly compare behavioral and neuronal measures, for clarity. In contrast, for neuronal measures of relative value we always use the superscript ‘neuronal’.

To conduct population analyses, we pooled all chosen juice cells. The juice eliciting higher firing rates was labeled ‘E’ (encoded) and other juice was labeled ‘O’. In Task 2, we thus referred to EO trials and OE trials, depending on the presentation order.

To illustrate the activity profiles of chosen juice cells in Task 2, we aligned spike trains at offer1 and, separately, at juice delivery. For each trial, the spike train was smoothed using a kernel that mimicked the postsynaptic potential by exerting influence only forward in time (decay time constant = 20 ms) (So and Stuphorn, 2010). In Figures 6 and 8A, we used moving averages of 100 ms with 25 ms steps for display purposes.

Under sequential offers, chosen juice cells encode different variables in different time windows (see Table 1). During offer1 and offer2 presentation, these cells encode in a binary way the juice type currently on display. Later, as the decision develops, these neurons gradually come to encode the binary choice outcome (i.e., the chosen juice). We previously showed that the activity of these neurons recorded in OE trials shortly before offer2 is inversely related to the value of offer1 (Ballesta and Padoa-Schioppa, 2019). This phenomenon, termed circuit inhibition, resembles the setting of a dynamic system’s initial conditions and is regarded as an integral part of the decision process (Ballesta and Padoa-Schioppa, 2019).

For a quantitative analysis of circuit inhibition, we focused on a 300-ms time window starting 250 ms before offer2 onset. We excluded forced choice trials, for which one of the two offers was null. For each neuron, we examined OE trials and we regressed the firing rates against the normalized value of offer1:

where ΔV_O was the value range for juice O. The normalization allowed to pool neurons recorded with different value ranges. The regression slope c₁ quantified circuit inhibition for individual cells, and we studied this parameter at the population level.

The activity of chosen juice cells in OE trials captures the momentary state of the decision and thus the evolving commitment to a particular choice outcome. To quantify the momentary decision state, we conducted an ROC analysis (Green and Swets, 1966) on the activity recorded during OE trials. This analysis was conducted on raw spike counts, without kernel smoothing, time averaging or baseline correction. We restricted the analysis to offer types for which the animal split choices between the two juices and we excluded trial types with <2 trials. For each offer type, we divided trials depending on the chosen juice (E or O) and we compared the two distributions. The ROC analysis provided an area under the curve (AUC). For each neuron, we averaged the AUC across offer types to obtain the overall CP (Kang and Maunsell, 2012). The ROC analysis was performed in 100ms time windows shifted by 25 ms. We also conducted the same analysis on four 250 ms time windows, namely pre-offer1 (−250 to 0 ms from offer1 onset), late offer2 (−250 to 0 ms from offer1 offset), early wait (0 to 250 ms after offer2 offset), and pre-juice (−250 to 0 ms before juice delivery) (Figure 8). In Figures 6 and 8B–I, cells were excluded because the Matlab function perfcurve.m failed to converge.

The funders had no role in study design, data collection, and interpretation, or the decision to submit the work for publication.

We thank H Schoknecht for help with animal training, L Snyder for helpful discussions, and Z Balewski, E Bromberg-Martin, K Conen, A Livi, P Natenzon, T Ott, J Tu, and M Zhang for comments on the manuscript. This research was supported by the National Institutes of Health (grant number R01-MH104494 to CPS) and by the McDonnell Center for Systems Neuroscience (predoctoral fellowship to WS).

All the experimental procedures adhered to the NIH Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee (IACUC) at Washington University (protocol number 190931).

© 2022, Shi et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.