Choice History Biases in Dyadic Decision-Making

doi:10.21203/rs.3.rs-4375984/v1

Download PDF

Article

Choice History Biases in Dyadic Decision-Making

https://doi.org/10.21203/rs.3.rs-4375984/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

How do we interact with our environment and make decisions about the world around us? Empirical research using psychophysical tasks has demonstrated that our perceptual decisions are influenced by past choices, a phenomenon known as the “choice history bias” effect. This decision-making process suggests that the brain adapts to environmental uncertainties based on history. However, the use of single-subject experiment task design is prevalent across the work on choice history bias, thus limiting the implications of the empirical evidence to individual decisions. Here, we explore the choice history bias effect using a dual-participant approach, where dyads perform a shared perceptual decision-making task. We first consider two extreme hypotheses: the participant either treats his/her partner’s decision as his/her own or simply ignores the partner’s decision. We then use a statistical modeling approach to fit generalized linear models to the choice data in a series of steps. Our best-fitting model suggests the participant has a choice repetition bias that spans several trials in the past, compatible with previous single-participant studies. Yet, there is also a dyadic influence on decision-making where both the participant’s own and partner’s last responses indicated a choice alternation bias. The results reject the hypothesis that the participant ignores the partner’s decision, in line with the idea that perceptual decision-making is not solely an individualistic decision process, though the partners’ decisions are treated differently from their own decisions.

Earth and environmental sciences/Environmental social sciences/Psychology and behaviour

Biological sciences/Neuroscience/Cognitive neuroscience/Perception

In daily life, people perceive and process uncertain sensory information to make decisions that lead to useful actions. For example, medical professionals examine X-ray scans to determine signs of abnormalities, or badminton players judge whether the shuttle during a double-player match landed inside or outside the court line. This ability to make perceptual judgments is central to human cognition (Shadlen & Kiani, 2013). In particular, it involves mapping noisy sensory information as input and transforming it into decision responses as output. As such, classical psychophysical methods are often used to describe this perceptual process to understand cognition better (Busse et al., 2011; Donner et al., 2009).

Notably, extensive work on perceptual processing has demonstrated that past choices influence current decisions, a phenomenon referred to as “choice history bias” (Abrahamyan et al., 2016; Urai et al., 2017). This effect has been shown using perceptual tasks such as a two-alternative forced-choice (2AFC) task in which participants are asked to discriminate the direction of motioning visual stimuli (Murphy et al., 2014; Urai et al., 2019). A choice history bias effect is also found when the stimuli presented on successive trials are uncorrelated (Fründ et al., 2014). Such empirical evidence suggests the history bias effect persists as a suboptimal decision-making process in which the brain adapts to environmental uncertainties (Glaze et al., 2015). Therefore, perceptual decisions are influenced by experiment trial history even when the task is not adaptive.

Consistent across the studies on the choice history bias effect is the use of single-subject designs independently of social settings. For instance, Abrahamyan and colleagues examined the adaptability of choice history bias using data collected from individuals across three laboratories (Abrahamyan et al., 2016). Urai et al. analyzed choice data collected from multiple perceptual experiments across different sensory modalities conducted at the level of single subjects (Urai et al., 2019). Nonetheless, in reality, people are not isolated decision-makers. Rather, people often interact with others and integrate existing information available to them, such as when looking at maps together to navigate physical surroundings. In this scenario, one can be influenced by the social cues of others or their own bias in the decision process. Therefore, when standard psychophysical task designs do not account for interaction, insights drawn from these works remain limited to individual decisions.

Research on joint attention implicated the influence of social cues and shared attention on perceptual judgments. For example, Seow and Fleming experimented to test whether perceptual sensitivity depends on social context (Seow & Fleming, 2019). By asking participants to detect low-contrast Gabor patches, they discovered that participants' detection performance improved when the perception was shared with an avatar. This indicates that individuals consider the visual perspective of others when making perceptual judgments. Experimental work by Wahn and colleagues used joint visual-spatial tasks and linear modeling analyses to investigate how social factors, e.g., information about the co-actor’s actions or performance feedback, might account for group benefits (Wahn et al., 2017, 2023). The result of their stepwise modeling approach showed an accurate prediction of collaborative benefits and contributed towards understanding joint action in social cognition. Thus, perception and action are not solely individualistic processes but can be shaped by the dyadic nature of human interactions (Böckler et al., 2012; Gallotti & Frith, 2013; Knoblich & Sebanz, 2006; Vesper et al., 2016).

Here, building on the existing body of literature that highlights the role of social interaction in understanding cognition, we aim to explore how the choice history bias effect might be modulated in a social context. Specifically, we examine the participants' behaviors while they take turns performing a perceptual task with stimuli presented in a random sequence with their dyadic partner. Our research objective is to determine whether perceptual decision-making is more of an individualistic (independent of the co-actor’s action) or collective (contingent on the co-actor’s action) process despite the co-actor’s actions being irrelevant to the present decision. For this, we formulate and test competing hypotheses that reflect separate assumptions regarding choice history bias in a social context. The first hypothesis states that each participant treats her/his partner’s decision as her/his own. This suggests that the choice history bias is actor-independent, i.e., the history effect on perceptual decision-making is not limited to a specific actor in the dyad but relates to the combined sequence of decisions. The second hypothesis is that the participant ignores her/his partner’s decisions. This assumes the decision is solely influenced by the participant’s own trial history. In other words, there is no dyadic choice history bias but only individual choice history bias effect. Furthermore, intermediate models that combine these two extremes in different fashions are conceivable. To test these competing hypotheses, we use a statistical modeling approach. Specifically, we examine the fit of generalized linear models corresponding to the different hypotheses to trial-by-trial choice response in a series of steps. The goal is to arrive at a model that best fits the behavioral data and, in turn, explains the extent to which the hypotheses are supported. Here, we test which hypothesis best fits our observations.

Participants

Seventy-eight individuals, grouped in thirty-nine dyads, were recruited for the present study. Twelve participants (six dyads) were excluded from performing the main experimental task due to exceptionally poor performance during the practice block. This leaves 33 dyads, or 66 individuals (N = 66, 44 females, 21 males, one non-binary, M = 25 years old, SD = 5 years). All participants had normal or corrected-to-normal vision without a history of neurologic or psychiatric illnesses. All participants provided written informed consent before the experiment. The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of the University of Osnabrück.

Experiment Protocol

A speeded random dot motion (RDM) discrimination task was used (Fig. 1a). The task involved viewing a cloud of moving dots (each motioning randomly and simultaneously towards left or right) and determining whether they coherently move rightward or leftward by pressing the two colored buttons (blue = right, yellow = left) on the custom keyboard (Black Box Toolkit USB Response Pads [URP48/URPVK], blackboxtoolkit.com) accordingly. Before the participants began the task, the experimenter gave both written and verbal instructions on the experiment procedure. The experimenter also demonstrated how to make a response by using the keyboard buttons. Participants were instructed to perform the task as quickly and as accurately as possible. In addition, participants were instructed to fixate their eyes on the center of the stimulus presented as a green cross when performing the task.

All stimuli were created in Python (version 3.9.2) using the Psychophysics Toolbox version 2021.1.3 (Peirce et al., 2019). The stimuli's leftward and rightward movement directions were equiprobable and randomly selected across trials. The dots (N = 328) were white with a size of 3x3 pixels, circular aperture of 5° diameter, speed of 9.95°/s, and density of 16.70 dots/degree². They were presented against a black background. The coherence of the stimuli, defined as the proportion of dots moving in the signal direction, was pre-determined. For instance, at a coherence level of 0.5, half of all the dots moved in the trial’s direction, which was set to either 0 or 180 degrees (leftward or rightward) on every frame. These dots constituted the “signal dots''. The remaining half was referred to as the “noise dots”, where each followed a random but constant direction on each frame. In order to direct the participant’s gaze at the stimuli and keep any involuntary eye movements or drift to a minimum for extended time periods, a bullseye fixation cross was used (Thaler et al., 2013). For every trial, the fixation cross color changed from green to either blue or yellow for 700ms post-response to indicate the acting participant’s choice response (yellow if responded ‘left,’ blue if responded ‘right’). During this feedback interval, the dots were stationary. The participant’s partner also saw such feedback information in the main experiment.

In general, our task design and procedure closely replicated established work on perceptual decision-making, particularly that of Murphy et al., 2014 where they quantified the decision-making parameters. Our study consisted of two sessions: testing and the main experiment. In the testing session, the participants individually and separately performed the practice block followed by the titration block. The practice block consisted of 40 trials of moving dots at a fixed coherence of 0.4. The titration block consisted of 240 trials with randomly selected dot coherences (0, 0.05, 0.1, 0.2, 0.4, 0.8, 40 trials each). While one participant of a given dyad was performing the practice and titration blocks, the other participant was instructed to wait quietly outside the experiment room. There was a short break between the practice and the titration blocks. If the participants did not achieve an accuracy level of 75% during practice, they were afforded another opportunity (3 practices in total) to repeat the practice block before proceeding to the titration block. In the titration block, the individual coherence threshold was estimated from a psychometric function fit to yield a goal accuracy level of 75% (Murphy et al., 2014). If the participants failed to reach the goal accuracy or higher, they were excluded from participating in the main experiment, and the experiment was aborted.

The main experiment consisted of 10 blocks with 100 trials each. In the main experiment, two participants of a dyad sat in two separate experimental rooms to perform the task. They alternated randomly to respond to the stimuli presented on a 24-inch-wide Dell U2412M monitor with a resolution of 1920 x 1200 pixels and a refresh rate of 60 Hz at a viewing distance of 60 cm. On each trial, only one dyad member was assigned to respond to the stimulus. The viewing distance was measured from the participant’s eye to the center of the monitor. The participants self-adjusted the chair’s height to view the center of the stimulus comfortably and placed their index fingers on the custom keyboard to make a response. The chair was fixed to the floor with the help of the experimenter. After every two blocks of the experiment, the experimenter measured the participant’s viewing distance again to keep the viewing distance equal.

At the beginning of the main experiment, the dyads underwent sound familiarization trials. They were trained to recognize their own and their partner’s tones as cues to respond in a given trial. The two distinct tones in the sound familiarization trials were musical notes “C” at octave 5 and “F” at octave 4. Each note had a duration of 0.5 seconds and was played to the participants 5 times. Lexical instructions accompanied the playing of the tones: “When you hear this, it’s your turn to respond” and “When you hear this, your partner will respond.”

Note that while the testing session consisted of lexical feedback on the response correctness, i.e., a green “Correct” or a red “Incorrect” word was presented below the stimulus after every trial, such feedback was absent during the main experiment. In addition, following the decision interval of 1500 ms after stimulus onset as used in Murphy et al., 2014, we also set lexical warnings for response time < 100ms (“Too Fast”) and > 1500ms (“Too Slow”). Similarly, during the main experiment, lexical warnings “Partner Too Slow” and “Partner Too Fast” were indicated to the observing participant in the dyad.

We used an adaptive approach to achieve an overall target of 75% mean accuracy. In other words, while the participants were simultaneously presented with stimuli moving in identical directions, the stimuli difficulty level was tailored to each participant based on his/her behavioral data. We implemented such an adaptive procedure because our extensive pilot tests, which used a logistic Weibull function to estimate the individual psychometric curve, showed notable misalignment in the accuracy outcomes between the titration block and the main experiment. The adaptive approach was implemented by fitting a drift-diffusion model (PyDDM) to the mean reaction time (RT) and accuracy data from the titration block of the experiment (Palmer et al., 2005; Shinn et al., 2020). This proportional-rate diffusion model used a maximum likelihood procedure to estimate each participant's psychometric and chronometric functions. The psychometric curve parameterization was derived from the product of the drift rate and decision bound. The drift-diffusion model was re-fitted to the behavioral data after every experiment block to re-align the task difficulty with the participant’s behavior during the main experiment. Unlike the logistic Weibull function that only fitted the accuracy data, the drift-diffusion model used more data, including the response, RT, and accuracy, and accounted for the entire titration block. The adaptive procedure enabled task difficulty adjustments throughout the main experiment, thus minimizing large deviations from the targeted accuracy outcomes.

After completing the experiment, the participants were asked to complete questionnaires on their demographic data and how well they know their partner on a 100-point scale. The testing session for each participant lasted about 30 minutes. The main experiment took around 2 hours; therefore, the entire experiment took about 3 hours for each pair of participants. Overall, the experiment set-up followed closely past empirical work that the interaction between the dyads is solely within the perceptual task.

Method of data analysis

Terminologies and variable coding. Here, we describe the variable names and terminologies used throughout the study. As illustrated in Fig. 1b, the participant currently performing the trial was described as the “acting participant” in the “active trial”. In contrast, the participant who was not currently performing is described as the “observing participant”. In discussing the experiment trial history, the last response given by the acting participant is referred to as the “own last response”. In contrast, the last response performed by the observing participant was referred to as the “partner last response”. To be exact, the former refers to the person acting now, what was his or her last response, while the latter refers to the person not acting now, what was his or her last response. +1 and − 1 were used in the coding of the identity of the participant (+ 1 = own; -1 = “partner”) as well as the choice response (+ 1 = right; -1 = left). For logistic regression modeling estimation, the dependent variable is the participants’ response choices, which were bounded between 0 and 1, reflecting the probability of selecting rightward, rising from 0 (left) to 1 (right).

Generalized linear modeling. To test our hypotheses, we fitted a series of generalized linear models (GLMs) with logit link function to quantify the influence of trial history on choice behavior. Following prior work on modeling choice history biases (Braun et al., 2018; Urai et al., 2019), we used the Akaike Information Criterion (AIC) values (Akaike, 1974) for formal model comparisons. Alongside the AIC values, we also used the model’s accuracy and mean squared error (MSE) to assess the model’s performance. The binomial logistic regression estimated the probability of selecting the right response based on the weighting of both sensory (i.e., current stimulus) and nonsensory parameters (e.g., past trial responses). The model distinguished response biases, such as when the participants preferred to repeat or switch their choice response. We chose to use effect coding for easier interpretation of the coefficient estimates as they directly indicate the difference in the mean outcome variable between the two levels of the predictor variables.

Exploratory data analysis

First, we explored the number of recorded responses. Out of the 33,000 responses recorded, there are 16,302 left responses and 16,698 right responses. The data were balanced with an approximately equal distribution of trials completed by each participant (range = 337 ~ 522 trials, M = 471 trials, SD = 30). No missing response was observed in the participants.

We processed and analyzed the reaction time (RT), performance accuracy, and coherence level data. Only trials with RT greater than 0.1 seconds and less than 1.5 seconds were included in the subsequent analysis, resulting in a removal of 1795 trials out of 33,000 trials (5.44% of total trials). In the main task, the average RT was 0.87 seconds (SD = 0.25 seconds), and the average performance accuracy was 73.6% (SD = 5.87%). The individual coherence threshold level range was 0.20 ~ 0.23, with a mean of 0.21 (SD = 0.012). Figure 2 shows the participant’s mean RT, accuracy, and adapted coherence level changes throughout the main experiment. Pearson's product-moment correlation between the accuracy and stimuli coherence shows a significant positive but weak correlation (r = 0.12, p < .001). The correlation between RT and accuracy was negative and significant (r = -0.23, p < .001); furthermore, the correlation between coherence and RT was negative and significant (r = -0.18, p < .001). The consistency in the mean RT and performance with a slightly decreasing trend for stimuli coherence suggested the adaptive procedure worked reasonably well. This enabled the behavioral data to be treated as stationary.

Modeling choice history biases

In general, the modeling approach started with an unbiased baseline model, followed by including the actor’s identity and choice history data. The modeling steps were hypothesis-driven, and the objective was to arrive at an arguably less complex model that is simpler to interpret and performs well. In the following, we report the results of each modeling step and how we arrived at the best-fitting model.

First, we examined the Task-Only model, where only the current trial stimulus was included as a predictor to model the participants’ performance without additional constraints, such as what the previous response was or the identity of the previous trial actor. This model served as a baseline model and assumed the estimation of the current response depends solely on the current trial stimulus, which was the actual task. The results showed a statistically significant effect of the current stimulus on the current response (β = 1.04, 95% CI [1.01, 1.06], SE = 0.01, p < .001). The accuracy of the model’s prediction was 73.41%, with an MSE of 0.193. The accuracy value was computed by comparing the model’s predicted probabilities (in values 0s and 1s) against the true labels of the outcome variable. The MSE value was calculated as the average squared difference between the predicted and the true values. The threshold for transforming the predicted probabilities into predicted labels was set at 0.5. If the predicted probability for a rightward response exceeded 0.5, we interpreted it as a prediction for class 1 (rightward response). Otherwise, the predicted probabilities below 0.5 were classified as a prediction for class 0 (leftward response). The predicted probability for a rightward response given a rightward-moving stimulus derived from the model’s estimates was 73.8%, which aligned with our intention of the task design to yield a goal of about 75% accuracy performance. The significant and positive association between the predictor and the response suggested the participants followed task instructions and behaved as they should.

To investigate the influence of past trial responses, we built on the baseline model to additionally include the variable “own last response”, which coded for the last response by the acting participant. Note that the variable included only the last response and did not account for the number of trials that may have passed since the same participant last acted. In line with the second hypothesis, which posits that the participant ignores his partner’s response, this Own History model assumes any of the partner’s past choices as irrelevant but allows modeling of the own history bias. The results of the model showed that the participant’s own last response positively and significantly predicted the current response (β = 0.15, 95% CI [0.12, 0.18], SE = 0.01, p < .001), while the current stimulus variable remained significant as seen in the baseline model (β = 1.04, 95% CI [1.01, 1.07], SE = 0.01, p < .001). The model exhibited an AIC value of 31334.72 (Δ = -118.39 units from the Task-Only model). The AIC measures the model’s goodness of fit and complexity by penalizing additional parameters. The observed decrease in AIC values thus suggested the additional variable led to a better fit. The model's accuracy remained at 73.4% with an MSE value of 0.192. Note, however, that the accuracy value was computed based on predictions thresholded at 0.5, thus not fully reflecting the variations in the model’s factual prediction values. The results suggested a repeat bias based on the acting participant’s last response.

Given the participant exhibited a choice history bias, we extended the previous Own History model to include the partner’s last response. This step tested how the partner’s history response influences the model’s prediction. This Own & Partner History model did not account for the number of trials that had passed since the same or different participant (acting or observing participant) last acted and assumed distinct contributions of the participant’s last choice response versus that of their partner’s. Therefore, the model further tested the second hypothesis on whether there is a dyadic choice history or only individual choice history bias. The results showed that both the current stimulus (β = 1.04, 95% CI [1.01, 1.07], SE = 0.01, p < .001) and the participant's own last response (β = 0.15, 95% CI [0.12, 0.18], SE = 0.01, p < .001) had a significant positive effect on the current response. In contrast, the last response made by the participant's partner showed a significant negative association with the current response (β = -0.04, 95% CI [-0.07, -0.02], SE = 0.01, p < .01). The accuracy of the model was 73.4% with an MSE value of 0.192. The AIC value, however, decreased compared to the Own History model (AIC = 31326, ΔAIC= -8.32). The results indicated that including the partner’s last response slightly improved the model’s predictive performance. This suggested that the participant tended to repeat his last decision but did not ignore his partner’s last decision. Nevertheless, the influence of the partner’s last response on switching the choice to be made was relatively slight compared to the participant’s own last response.

The results thus far indicate an influence of the acting and observing participant’s last response on choice behavior; however, it is unclear whether this choice history bias effect reaches further back in trial history for both acting and observing participants. Therefore, in the next step, we developed a family of models (Trace History model), which considered more and more decisions further into the past, to investigate the influence of the past response as the number of the last trial lag increases. These models are indexed up to the fifth lag, reflecting the maximal lag considered for the acting (own) and observing (partner) participant. The indexing notation (i, j) in the Trace History model represents the number of the last trial lags for the acting and observing participants, respectively. As such, the Own & Partner History model is identical to the Trace History (1,1) model. Including the participant’s own second-last choice response results in the Trace History (2,1) model. Following this, we included the same data for the dyadic partner, leading to the Trace History (2,2) model. In this order, we repeated the modeling steps until we reached the Trace History (5,5) model, where both the own and partner's five last trials are considered. From this series of model fitting, we observed the inclusion of the variables for the own last trial data from lag 2 until lag 5 showed a consistent statistical significance (p < .001), reduction in the AIC values (Mean ΔAIC = -176.80), constant accuracy value of 73.41% with slight decrease in the MSE values (Mean ΔMSE = -0.013). Nevertheless, the partner’s trial history response data beyond the last one did not systematically improve the model’s predictive performance. Specifically, the model that showed statistically significant variables with the lowest AIC and MSE is the Trace History (5,1) model (AIC = 30619; MSE = 0.187) in comparison with the rest of the family of models fitted. The Trace History (5,1) model exhibited estimates for each of the acting participant’s trial lag of β₁ = 0.10, 95% CI [0.08, 0.13]; β₂ = 0.23, 95% CI [0.20, 0.25]; β₃ = 0.17, 95% CI [0.14, 0.20]; β₄ = 0.15, 95% CI [0.12, 0.18]; β₅ = 0.12, 95% CI [0.10, 0.25], each estimate with a SE = 0.01 and a p < .001. In contrast, the partner's last response variable showed a negative estimate of -0.04, 95% CI [-0.06, -0.01], SE = 0.01, p < .01. The stimulus variable showed an estimate of 1.08, 95% CI [1.05, 1.11], SE = 0.01, p < .001. Figure 3 presents a schema illustrating the statistically significant log-odds estimates observed in the Trace History (5,1) model for each variable (own and partner) at different lag indices. This schema highlights the influence of the own trial history compared to that of the partner in predicting the participant’s choice behavior. Taken together, the results suggested only the participant’s own choice bias effect traces further back in trial history, with a small dyadic influence observed in which the partner’s last response predicted a switching of choice response.

Having examined the role of the acting and observing participant’s last five decisions in predicting choice response, we investigated whether a simpler model that assumes exponential decay of past choices could provide an equally good fit. For this, we developed the Joint Weighted History model that fitted the acting and observing participants’ trial history responses in a way that accounted for a memory-decaying effect. The Joint Weighted History model builds on the Own & Partner History model which additionally included two variables that combined the participant’s and the partner’s responses, respectively. These responses were weighted as a function of the lag to the current trial. This approach conceptually aligned with past work that computed a “history kernel” to quantify the effect of stimuli and responses from past trials on the current choice processes (Urai et al., 2017). There, positive and negative weights were assigned to each previous stimulus and choice to indicate a tendency to repeat or alternate. Then, every set of seven previous trials was convolved with exponentially decaying functions sensitive to the changes in history data due to time, e.g., more distant trials having less impact (Fründ et al., 2014). Here, the exponentially weighted moving averages for the own and the partner trial responses were computed with a loss factor gamma (γ) of 0.8. The γ value determined how quickly the influence of the older data points decayed. This value approximated for a window size of 5 last trial lags (1 / (1 - γ) = 5), where the last decision was assumed to receive the highest weight and exponentially diminished as it moved further back in trial history. The results of the Joint Weighted History model showed the current stimulus (β = 1.08, 95% CI [1.05, 1.11], SE = 0.01, p < .001) and the weighted own trial history variable (β = 0.95, 95% CI [0.88, 1.02], SE = 0.04, p < .001) had a significant positive effect on the current response. In contrast, the partner's last response variable showed a significant negative estimate (β = -0.04, 95% CI [-0.07, -0.00], SE = 0.02, p < .05). Notably, during model fitting, the participant's own last response variable was corrected into a negative estimate (β= -0.18, 95% CI [-0.21, -0.14], SE = 0.02, p < .001). However, the additional weighted partner trial history variable did not suggest significance. The model improved its predictive power as indicated by a decrease in the AIC value compared to the Own & Partner History model (AIC = 30616, ΔAIC = -710.02). The model's accuracy is 73.41% and an MSE of 0.262. Further removing the weighted partner trial history variable led to improvement in the model’s performance as observed in the 2 units decrease in the AIC (AIC = 30614, ΔAIC = -2), along with a prediction accuracy of 73.41% and an MSE of 0.187. This reduced model, named the Individual Weighted History model, therefore exhibited more of a balance between model complexity and explanatory power. Specifically, it showed significantly positive estimates for the current stimulus (β = 1.08, 95% CI [1.05, 1.11], SE = 0.01, p < .001) and the weighted own trial history variable (β = 0.95, 95% CI [0.88, 1.02], SE = 0.04, p < .001). The partner's last response variable, however, showed a significant negative estimate (β = -0.04, 95% CI [-0.06, -0.01], SE = 0.01, p < .01). Similarly, the own last response variable exhibited a significant negative estimate (β= -0.18, 95% CI [-0.21, -0.14], SE = 0.02, p < .001). Together, the simpler Individual Weighted History model indicated a choice bias in which the participant tended to repeat based on his own trial history. However, the participant was also more likely to switch after his own last response, similar to the partner’s last response.

In summary, we performed a stepwise modeling procedure guided by our hypotheses and selected the Individual Weighted History model as the best-fitting one based on the model selection criteria. This model assumed exponential decay and achieved a better fit with few parameters. The model specifies an effect of the task stimulus, the last response by the acting, as well as that of the observing participant, and the exponentially weighted moving averages of the participant’s own trial history responses on the choice to be made. Figure 4a shows a coefficient plot for the variable estimates. Figure 4b shows the estimated marginal means (the predicted probability values) for the response at the level of the exponentially weighted moving averages of the participant's own trial history predictor while holding the other predictors constant. As the average weighted response increases from − 1 to + 1, the predicted likelihood of repeating the same response increases, indicating that one’s own cumulative choice history leads to a stronger bias in the participant’s choice decision. Figure 4c, d, e also shows the model’s predicted probabilities for the response at the levels of the other variables in the model, including the trial stimulus and the own and partner’s last response with their associated standard errors. Lastly, Fig. 5 summarizes the regression steps to model choice history biases in dyadic decision-making. The model indicated that while the participant had a repeat bias that spans several trials in the past, he tended to alternate the choice of his own last response. The dyadic partner’s last response also influenced the participant to alternate his choice, albeit to a lesser extent. Thus, the participant did not ignore the partner’s decision as stated in the second hypothesis, rather, he acknowledged the partner's decision by not following it, similar to his own last response.

In the present study, we investigated the choice history bias effect in a social context. Using a stepwise statistical modeling approach, we tested the extent to which the modeling results support our two proposed hypotheses. Comparison between the models led to the Individual Weighted History model that best fits the observed data. The model indicated a significant influence of the partner’s last response on the participant’s decision. This rejected the second hypothesis that the participant ignores his partner’s decision. However, the model also showed the partner’s last response predicted a tendency towards a bias for choice alternation. Thus, this also rejected the first hypothesis that the participant treats his partner’s decision as his own.

The current study is exploratory and has limitations. One limitation is that the participants sat in separate experimental rooms and did not share the same peripersonal space. This physical separation may have limited the interaction effect typically accounted for in a shared space (Knoblich & Sebanz, 2006). The participants did not communicate but only observed each other’s responses on each trial. While this factor was part of the experiment design, it could reduce a sense of social presence that influences perceptual judgments (Bahrami et al., 2012). On the flip side, these design choices allowed a clean and unambiguous analysis of the dyadic decision process.

The modeling of decisions tracing further back in history indicated that the participant has a choice repetition bias that spans several trials in the past. This is consistent with the literature that sequential perceptual choices depend on the current sensory information and the acting individual’s own trial history responses (Fründ et al., 2014). We also found that including more and more responses further into the past improved the model’s predictive performance, with the corresponding positive estimates diminishing in weight. This is in line with the findings of Braun et al., 2018 and Urai et al., 2017 where the choice bias effect lasted up to seven previous trials, and the computation of the participant’s history weights as a function of lags exhibited decaying profiles. Taken together, the participant accumulated his own past trial responses into a bias for the choice to be made.

Despite the evidence suggesting the participants relied on their own choice history, our findings also indicate a dyadic influence. In the best-fitting model, the partner’s last trial response variable showed a negative estimate, implying a tendency to alternate in the choice to be made. This rejects the second hypothesis that the participant ignores his partner’s past decision. Such findings align with the research on shared perception (Gallotti & Frith, 2013; Seow & Fleming, 2019), where how others perceive the visual stimulus can serve as an available source of information for the individual who is making the decision. However, while the participants acknowledged their partners’ decisions, they did not necessarily adhere to them. This behavior also rejected the first hypothesis, which posits that each participant treats his/her decision as his/her own. Instead, the participants appeared to treat their partners’ decisions differently from their own. Previous research by Bahrami et al. suggested that the lack of communication as well as feedback between pairs of participants could result in no build-up of collective benefit (Bahrami et al., 2012). Furthermore, empirical work by Cole and colleagues also challenged the notion that people spontaneously adopt the perspective of others. Their work showed that perspective-taking also could occur when two actors perceived different stimuli (Cole et al., 2016). Together, our findings indicate a dyadic influence in choice history, suggesting that participants relied on their own trial history while also considering their partners’ decisions, though not following them.

The total influence of the participant’s own history of decisions was determined by assigning weights to each of the last trial responses, starting from the most recent (lag 1) to the most distant (lag 5) and computing the exponentially weighted moving averages. As such, in the best-fitting model, the own last response variable was included twice. It was included in the weighted own trial history variable for which the exponentially weighted moving averages for the own last trial responses were calculated, and additionally, it was explicitly included as a separate standalone variable in the model. Fitting the model to the choice history data resulted in a negative estimate (β= -0.18) for the isolated own last response variable, indicating an overestimate of the combined influence of the participant’s own weighted average history responses (β = 0.95). In other words, when solely considering the own last response variable in isolation, the model predicted a positive weight, suggesting a choice repetition bias following one’s own last decision. However, when factoring in the combined impact of one’s own history decisions, which accounted for a temporally decaying influence, the effect of the own last response was corrected, and the model returned a negative estimate. This implies a tendency for choice alternation, similar to following the partner’s last response. As a result, the participant seemed to show a degree of adaptability in his choice behavior, where he was responsive to the partner’s past decision and adapted his behavior accordingly. For this, past empirical work using a shared perceptual task to investigate how co-actors influence each other’s attentional focus provided evidence that people are sensitive to other’s attentional relations to the environment (Böckler et al., 2012). Specifically, analyses of RTs showed the participants slowed down when they had to adopt a different attentional focus from that of their own, which induced a selection conflict. Therefore, while the participant showed a bias to repeat, he adapted his decision-making strategy to account for both his personal and partner’s past decisions.

The present study contributes to the literature on the role of social interaction in perceptual decision-making (Bahrami & Frith, 2011; Deroy et al., 2023). Our task design differed from past work on choice history bias by including a second person. We present a model showing that perceptual judgment is not solely individualistic. The participant acknowledged the partner’s most recent decision yet did not treat it as his own. Future work can extend the applicability of our findings to other interaction scenarios. For example, using a perceptual task, a human participant could interact and observe a computer that mimics the participant’s behavior. Specifically, one can examine how choice history bias might emerge or change under three conditions, such as when the computer is consistently correct, consistently wrong, or follows a mixed pattern more similar to naturalistic human decision-making. Given that people typically attribute social qualities to computers (Nass et al., 1994), the extent to which there might be a joint social perception remains unclear. The results carry broader implications of trust in technology and adapting choice decisions to external agents. In conclusion, we have explored the choice history bias effect in dyadic perceptual decision-making, which suggests a more realistic approach to understanding cognition as, in reality, humans are not isolated decision-makers.

Author Contribution

AC, PK, MP, and AH designed the study. AH and MP carried out the main experiments and data analysis under the supervision of AC and PK. VZ performed code reviews and additional analyses. All authors contributed to the discussions of the results. AH wrote the manuscript, and PK provided comments to improve the manuscript.

Data Availability

The code of this project is publicly available on the Open Science Framework: https://osf.io/3v4m8/; DOI 10.17605/OSF.IO/3V4M8.

Abrahamyan, A., Silva, L. L., Dakin, S. C., Carandini, M., & Gardner, J. L. (2016). Adaptable history biases in human perceptual decisions. Proceedings of the National Academy of Sciences, 113(25), E3548–E3557. https://doi.org/10.1073/pnas.1518786113
Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions on Automatic Control, 19(6), 716–723. https://doi.org/10.1109/TAC.1974.1100705
Bahrami, B., & Frith, C. D. (2011). Interacting Minds: A Framework for Combining Process- and Accuracy-Oriented Social Cognitive Research. Psychological Inquiry, 22(3), 183–186. https://doi.org/10.1080/1047840X.2011.573767
Bahrami, B., Olsen, K., Bang, D., Roepstorff, A., Rees, G., & Frith, C. (2012). Together, slowly but surely: The role of social interaction and feedback on the build-up of benefit in collective decision-making. Journal of Experimental Psychology: Human Perception and Performance, 38(1), 3–8. https://doi.org/10.1037/a0025708
Böckler, A., Knoblich, G., & Sebanz, N. (2012). Effects of a coactor’s focus of attention on task performance. Journal of Experimental Psychology. Human Perception and Performance, 38(6), 1404–1415. https://doi.org/10.1037/a0027523
Braun, A., Urai, A. E., & Donner, T. H. (2018). Adaptive History Biases Result from Confidence-Weighted Accumulation of past Choices. Journal of Neuroscience, 38(10), 2418–2429. https://doi.org/10.1523/JNEUROSCI.2189-17.2017
Busse, L., Ayaz, A., Dhruv, N. T., Katzner, S., Saleem, A. B., Schölvinck, M. L., Zaharia, A. D., & Carandini, M. (2011). The Detection of Visual Contrast in the Behaving Mouse. Journal of Neuroscience, 31(31), 11351–11361. https://doi.org/10.1523/JNEUROSCI.6689-10.2011
Cole, G. G., Atkinson, M., Le, A. T. D., & Smith, D. T. (2016). Do humans spontaneously take the perspective of others? Acta Psychologica, 164, 165–168. https://doi.org/10.1016/j.actpsy.2016.01.007
Deroy, O., Longin, L., & Bahrami, B. (2023). Co-Perceiving: Bringing the social into perception. https://www.psycharchives.org/en/item/57b1490f-77f8-4a6b-8362-524e9f85b46c
Donner, T. H., Siegel, M., Fries, P., & Engel, A. K. (2009). Buildup of Choice-Predictive Activity in Human Motor Cortex during Perceptual Decision Making. Current Biology, 19(18), 1581–1585. https://doi.org/10.1016/j.cub.2009.07.066
Fründ, I., Wichmann, F. A., & Macke, J. H. (2014). Quantifying the effect of intertrial dependence on perceptual decisions. Journal of Vision, 14(7), 9. https://doi.org/10.1167/14.7.9
Gallotti, M., & Frith, C. D. (2013). Social cognition in the we-mode. Trends in Cognitive Sciences, 17(4), 160–165. https://doi.org/10.1016/j.tics.2013.02.002
Glaze, C. M., Kable, J. W., & Gold, J. I. (2015). Normative evidence accumulation in unpredictable environments. eLife, 4, e08825. https://doi.org/10.7554/eLife.08825
Knoblich, G., & Sebanz, N. (2006). The Social Nature of Perception and Action. Current Directions in Psychological Science, 15, 99–104. https://doi.org/10.1111/j.0963-7214.2006.00415.x
Murphy, P. R., Vandekerckhove, J., & Nieuwenhuis, S. (2014). Pupil-Linked Arousal Determines Variability in Perceptual Decision Making. PLOS Computational Biology, 10(9), e1003854. https://doi.org/10.1371/journal.pcbi.1003854
Nass, C., Steuer, J., & Tauber, E. R. (1994). Computers are social actors. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 72–78. https://doi.org/10.1145/191666.191703
Palmer, J., Huk, A. C., & Shadlen, M. N. (2005). The effect of stimulus strength on the speed and accuracy of a perceptual decision. Journal of Vision, 5(5), 1. https://doi.org/10.1167/5.5.1
Peirce, J., Gray, J. R., Simpson, S., MacAskill, M., Höchenberger, R., Sogo, H., Kastman, E., & Lindeløv, J. K. (2019). PsychoPy2: Experiments in behavior made easy. Behavior Research Methods, 51(1), 195–203. https://doi.org/10.3758/s13428-018-01193-y
Seow, T., & Fleming, S. M. (2019). Perceptual sensitivity is modulated by what others can see. Attention, Perception, & Psychophysics, 81(6), 1979–1990. https://doi.org/10.3758/s13414-019-01724-5
Shadlen, M. N., & Kiani, R. (2013). Decision Making as a Window on Cognition. Neuron, 80(3), 791–806. https://doi.org/10.1016/j.neuron.2013.10.047
Shinn, M., Lam, N. H., & Murray, J. D. (2020). A flexible framework for simulating and fitting generalized drift-diffusion models. eLife, 9, e56938. https://doi.org/10.7554/eLife.56938
Thaler, L., Schütz, A. C., Goodale, M. A., & Gegenfurtner, K. R. (2013). What is the best fixation target? The effect of target shape on stability of fixational eye movements. Vision Research, 76, 31–42. https://doi.org/10.1016/j.visres.2012.10.012
Urai, A. E., Braun, A., & Donner, T. H. (2017). Pupil-linked arousal is driven by decision uncertainty and alters serial choice bias. Nature Communications, 8(1), 14637. https://doi.org/10.1038/ncomms14637
Urai, A. E., de Gee, J. W., Tsetsos, K., & Donner, T. H. (2019). Choice history biases subsequent evidence accumulation. eLife, 8, e46331. https://doi.org/10.7554/eLife.46331
Vesper, C., Abramova, E., Bütepage, J., Ciardo, F., Crossey, B., Effenberg, A., Hristova, D., Karlinsky, A., McEllin, L., Nijssen, S. R. R., Schmitz, L., & Wahn, B. (2016). Joint Action: Mental Representations, Shared Information and General Mechanisms for Coordinating with Others. Frontiers in Psychology, 7, 2039. https://doi.org/10.3389/fpsyg.2016.02039
Wahn, B., Kingstone, A., & König, P. (2017). Two Trackers Are Better than One: Information about the Co-actor’s Actions and Performance Scores Contribute to the Collective Benefit in a Joint Visuospatial Task. Frontiers in Psychology, 8, 669. https://doi.org/10.3389/fpsyg.2017.00669
Wahn, B., König, P., & Kingstone, A. (2023). Predicting group benefits in joint multiple object tracking. Attention, Perception, & Psychophysics. https://doi.org/10.3758/s13414-023-02693-6

No competing interests reported.

Download PDF

Editorial decision: Revision requested
31 Jul, 2024
Reviews received at journal
23 Jul, 2024
Reviews received at journal
04 Jul, 2024
Reviewers agreed at journal
29 Jun, 2024
Reviewers agreed at journal
28 Jun, 2024
Reviewers invited by journal
24 May, 2024
Editor assigned by journal
18 May, 2024
Editor invited by journal
11 May, 2024
Submission checks completed at journal
09 May, 2024
First submitted to journal
06 May, 2024

You are reading this latest preprint version

Choice History Biases in Dyadic Decision-Making

Status:

Version 1

Abstract

Figures

Introduction

Materials and Methods

Participants

Experiment Protocol

Method of data analysis

Results

Exploratory data analysis

Modeling choice history biases

Discussion

Declarations

Author Contribution

Data Availability

References

Additional Declarations

Status:

Version 1