Mar, 20 2020, 5-10min read
Thomas, A. W., Molter, F., Krajbich, I., Heekeren, H. R., & Mohr, P. N. (2019). Gaze bias differences capture individual choice behaviour. Nature human behaviour, 3(6), 625. doi.org/10.1038/s41562-019-0584-8
In this study, we investigated whether the previously reported group-level link between gaze allocation and choice behaviour in small choice sets (see, Krajbich et al. 2010 & 2011) exists on the level of the individual and whether its strength is variable across individuals.
To this end, we analyzed four published simple choice datasets: Krajbich 2010, Krajbich 2011, Folke 2016, Tavares 2017. These four datasets span 118 individuals, two choice set sizes (tow- and three-alternative), and two choice domains (value-based and perceptual choice) (see Fig. 1 for an overview).
Figure 1. In all experiments, participants were instructed to choose the best out of two (a, d) or three (b, c) items (that is, the item they would like to eat most in value-based tasks or the item most similar to a target stimulus that was presented every five trials in the perceptual task). Value-based experiments included a valuation task before the main choice task, whereby participants indicated their liking of each item (either by rating scale or willingness-to-pay). All choices were made without time restrictions. The choice task in c used a gaze-contingent presentation, whereby items were only revealed when the participant’s gaze was directed to an item’s location on the screen. Experiments used real snack food items instead of illustrations.
We first studied general behavioural differences between the individuals in the data on the following three metrics: participants’ mean response time (RT); mean probability of choosing the best item (we defined the best item either as the item with the highest liking rating or willingness-to-pay in the value-based choice tasks, or the item with the smaller angular distance to the target in the perceptual choice task); and influence of gaze allocation on choice probability (defined as the mean increase in choice probability for an item that was looked at longer than the others, after correcting for the influence of the item’s value or angular distance on choice probability).
Figure 2. Individual differences in the three studied behavioural metrics and their associations. a–c: Distributions of individuals’ mean RT (a), gaze influence (mean increase in choice probability for an item that is fixated longer than the others, after correcting for the influence of item value) (b) and probability of choosing the best item (c) per dataset.
We found that participants differed considerably in all three behavioural metrics (Fig. 2). Notably, 98% of the participants also showed positive scores on gaze influence measure, indicating an overall positive relationship between gaze allocation and choice probability. We also probed the relationship between the behavioural metrics, to better understand the dynamics of individual differences. We did not find any association between participants’ probability of choosing the best item and their RTs (Fig. 2d) and no association between participants’ gaze influence and RT (Fig. 2e). However, participants’ probability of choosing the best item from a choice set decreased with increasing individual gaze influence measures (Fig. 2f).
The four datasets strongly differed on the three behavioural metrics. Yet, these differences between datasets cannot be attributed to the effect of choice domain (perceptual versus value-based) or set size (two versus three items) alone, as original tasks also differed in other aspects (for example, different stimuli in value-based versus perceptual tasks, different number of trials and different presentation format). For this reason, we refrain here from interpreting these differences between datasets further.
The behavioural and eye-tracking data suggested substantial variability in the extent to which gaze affects participants’ choice behaviour (Fig. 2b). To provide conclusive quantitative evidence for or against the presence of a mechanism that biases choices depending on the distribution of gaze on the level of the individual, we adopted a computational modelling approach: We fitted and compared two variants of the gaze-weighted linear accumulator model (or GLAM; for details on the GLAM, see this page) to the RT, choice, and gaze data of each participant. The first is a full GLAM variant (with gaze bias) allowed the gaze bias parameter γ to vary freely between individuals. The second is a no-gaze-bias GLAM variant, whereby the gaze bias parameter γ was fixed to 1 (resulting in no influence of gaze on the accumulation process). Overall, the full GLAM fitted 109 out of 118 (92%) participants better than the no-gaze-bias variant. Within each dataset, the data of 79% (Krajbich 2010), 97% (Krajbich 2011), 100% (Folke 2016) and 100% (Tavares 2017) of the participants were better described by the full GLAM (according to the WAIC measure; Fig. 3a).
Figure 3. Individual relative model comparison between the full GLAM and a restricted no-gaze-bias GLAM variant. a: Across datasets, the response behaviour of most individuals is better described by the full model with a gaze bias (given by the lowest score on the WAIC). b: Individual WAIC differences between the full and restricted GLAM variant. Negative differences indicate better fits of the full model. Note that the y axis in b is truncated to better show small differences; the lowest WAIC difference was -400.64.
Similarly, in an out-of-sample prediction (in which we fitted both model variants to the even-numbered trials of each individual and then predicted choices and response times for all odd-numbered trials) both model variants performed well in predictions individuals’ response times (Fig. 4a) and probability of choosing the best item (Fig. 4b). Yet, only the full GLAM also accurately predicted captured empirical choice patterns that are driven by gaze and not solely by the values of the items (Fig. 4c).
Figure 4. Individual out-of-sample predictions of behavioural metrics for all odd-numbered trials. a–c: The full model with gaze-bias variant accurately predicts individuals’ mean RT (a), probability of choosing the best item (b) and influence of gaze on choice probability (c). d–f: The no-gaze-bias variant also accurately predicts individuals’ mean RT (d) and probability of choosing the best item (e), but it fails to accurately capture individuals’ influence of gaze on choice probability (f). Model predictions are simulated using parameter estimates obtained from individual fits on even-numbered trials.
Lastly, the individual parameter estimates of the GLAM from the even-numbered trials also correlated with the three behavioral metrics in the odd-numbered trials (Fig. 5): v (velocity parameter) estimates scaled logarithmically with the participants’ mean RT (Fig. 5a), while γ (gaze bias) estimates correlated with the strength of participants’ gaze influence on choice probability (Fig. 5b) and probability of choosing the best item (Fig. 5c).
Figure 5. Associations between individuals’ response behaviour in the odd-numbered trials and the model parameters estimated from the even-numbered trials. a: log-transformed mean RTs decrease with increasing log-transformed v estimates. b: Behavioural gaze influence decreases with increasing γ estimates. c: Individuals’ probability of choosing the best item increases with increasing γ estimates.
Taken together, these findings indicate an active role of gaze allocation in the decision process for the individual: On the behavioural level, participants exhibited an overall positive relationship between gaze and choice (with longer gaze increasing choice probability). Similarly, in a structured model comparison, the response behaviour of the majority of participants in all four datasets was also better captured by a model with gaze bias than by one without. Yet, the strength of the association between gaze allocation and choice was highly variable across individuals.
Interestingly, the strength of the association was predictive of individuals’ probability of choosing the best item from a choice set (stronger gaze biases were associated with more choices that were inconsistent with item values; Fig. 5 c). This relationship can be explained as follows: the gaze bias parameter lets the model bias the choice process according to the distribution of gaze between items. That is, with a strong gaze bias, the model’s predictions are strongly dependent on the distribution of gaze, and a gaze distribution that is random with respect to the value of the items then leads to more random choices. Conversely, the model’s predictions are independent of gaze when no gaze bias is present. The model then neglects gaze and predicts choices solely driven by the values of the item. These findings thereby identify another source of variability among individuals’ ability to choose the best item from a choice set, which is often attributed to differences in generic accumulation noise parameters obscuring further insight into the mechanisms driving these individual differences.