In our first experiment 48 participants completed the foraging game online. We manipulated risk of predation between subjects, with 24 participants in the distraction condition, and 24 in the hunted condition. We also modulated task difficulty in a number of other ways. In separate blocks, target selection was based on either a single colour feature or on a conjunction of colour and shape, as in our previous work (Kristjánsson et al., 2014). Across trials, we also varied the velocity with which the wolf objects moved, to increase or decrease the risk they posed. Finally, for half of the participants in each predation group we varied the behaviour of the wolf objects. For those in the “pack” condition, all 4 predator objects moved with independent, linear trajectories, irrespective of the position of the sheep object. For those in the “lone” wolf condition, one of the 4 wolf objects always changed direction to follow the current position of the sheep object. The other 3 wolves moved with independent linear trajectories, as in the pack condition. Again, this manipulation was included to increase the potential risk posed by the predators.
Methods
Participants
All 48 participants were recruited online from https://prolific.co. They were required to be fluent readers of English, within a specified age range (18–40 years) and to have not taken part in previous related studies. Demographically, they were located in different countries, with different native languages, variously employed or in full-time study, aged from 18 to 40 years (M = 27.1 years, SD = 5.7), and 21 were female. For their participation in the experiment they were paid a flat rate of £3.75, based on an estimated session time of 30 min.
Ethics & data protection
The research team were unaware of and had no access to the personal identity of the participants. In addition to the implied consent—given that participants were recruited through a voluntary, professional service—a full information sheet and consent form was presented prior to data collection. Participants were given the option of downloading these documents for later reference. They were required to confirm that they had read and understood the nature of the experiment and the data that would be collected and to explicitly confirm their informed consent for participation. These online procedures conform to the Ethics and Data Protection guidelines of the University of Malta.
Power analysis
The basic group size (N = 12) was determined prior to data collection and was chosen to directly match recent studies from our group where within-subject differences in run behaviour had been successfully measured (Thornton et al., 2019, 2020). To further verify that this sample size would provide sufficient power to detect the within-group feature/conjunction foraging patterns of interest, we conducted an a priori power analysis using the “Bias and Uncertainty Corrected Sample Size” (BUCSS) toolbox described by Anderson et al. (2017). BUCSS uses the reported F values and sample size from previous factorial studies—rather than derived estimates of effect size—to generate necessary sample sizes for planned studies. Here, we chose the previous study from our group (Thornton et al., 2020) that most closely matched the current within-group factorial design. Specifically, we chose a 2 (Target: feature/conjunction) × 5 (Foraging Tempo) repeated measures analysis of variance conducted on run length with a sample size of 11, focusing our a priori analysis on the main effect of Target, F(1,10) = 40.0, p < 0.001, MSE = 6.3, \({\eta }_{p}^{2}\)= 0.8. We used this F value, along with the sample size and alpha parameters from Thornton et al. (2020) as input to the BUCSS ss.power.wa function. We chose custom settings of assumed alpha for the planned study = 0.05, level of assurance = 0.95, and desired power of 0.8. We specified the main within-subject factors from the current experiment—2 (Target) × 5 (Wolf Velocity)—and identified the main effect of Target as the effect of interest. This analysis yielded a minimum sample size of 11 participants, closely approximating our initial choice.
Online protocols
All data for the current study were collected online. Participants were directed to a dedicated URL on the https://maltacogsci.org domain and were taken through a series of webpages that provided instructions, obtained consent and ran the experimental trials. Anonymous data was transferred automatically on a trial-by-trial basis to a secure server for later download and processing. As participation was remote, we could not control the specific laptop/desktop machines that were used, nor the monitor hardware/settings. We did exclude the use of mobile devices, as this version of our foraging task was designed not to respond to touch-based technology. We have previously run the basic foraging task in a desktop environment (Thornton et al., 2019), and while we anticipated some consequences on overall patterns of run behaviour related to reduced response selection speed (Thornton et al., 2020), these would be constant across the current manipulations of interest.
Equipment
As the current study was run online, we could not control the precise display conditions or equipment used. The online task was custom written in JavaScript so that it would run via web browsers opened on any laptop or desktop machine. Several recent review papers have indicated that the display and response timing of native JavaScript is capable of producing data that is comparable to lab-based testing (e.g., Bridges et al., 2020; Miller et al., 2018; Pronk et al., 2019). The code ensured that browsers were switched to full-screen mode, so that only the foraging display appeared centered on the screen. Checks within the code identified the physical frame rate of the display and capped the effective update rate to 60 Hz. To minimize possible mouse versus trackpad differences in response times, participants were allowed to move the cursor with either, but observers were required to press the spacebar to register a response. We have used this technique previously to equate response demands across input modalities (Thornton et al., 2019).
Foraging stimuli
Figure 1 shows the initial moment of a typical trial. Stimuli appeared on a grey canvas region (800 × 600 pixels) that was always centred on an otherwise blank, full screen. Participant used their regular mouse/trackpad to control the position of the cursor, that was visualised as a sheep (64 × 80 pixels). Each trial also contained 4 wolf objects (70 × 94 pixels) that had to be avoided or ignored, depending on the predation group of the participant. Target and distractor items (20 pixels) were randomly distributed on a trial-by-trial basis within a regular 10 × 8 virtual grid. During Feature foraging, the 40 targets were yellow and blue disks and the 40 distractors were red and green disks. During Conjunction foraging the 40 targets were red disks and green squares and the 40 distractors were green disks and red squares. In our previous work, we have found no effects of counter-balancing stimulus categories, and used a fixed mapping in the current task to simplify the online protocols.
Wolf behaviour
At the start of each trial, the 4 wolf objects were positioned as seen in Fig. 1, at the corners of the dot grid. They immediately began to move, initially converging on the centre of the screen. For wolves that were programmed to move on independent linear trajectories, a new direction was repeatedly chosen from the full 360° range after a period of between 1.7 and 3.3 s. These general motion characteristics were modelled on previous dynamic tasks from our group (e.g., Thornton et al., 2014, 2019) where further methodological details can be found. The lone wolf, if present, changed direction at 20 Hz to converge on the current location of the sheep object. For all wolves, if they arrived at the edge of the dot grid, their direction reversed. Wolf objects did not bounce when colliding with each other, but simply passed through. In the hunted condition, if a wolf overlapped with the sheep object, this terminated the trial. Across trials, the velocity of the 4 wolf objects was either 30, 42, 54, 66 or 78 pixels/s, with 3 repetitions of each velocity randomly distributed across the 15 trials of each condition.
Design
Overall the study involved a 2 (Predation: Hunted/Distracted) × 2 (Wolf Behaviour: Pack/Lone) × 2 (Target: Feature/Conjunction) × 5 (Wolf Velocity) factorial design, with the first two factors between subjects and the second two as within subject factors.
Task
On each trial, the goal was to cancel all of the target items as quickly as possible by placing the sheep on top of them using the mouse, and then pressing the spacebar. Once selected in this way, items disappeared from the screen. If a distractor item was mistakenly selected, the trial ended. Participants in the hunted condition were required to avoid contact with any of the wolves. For hunted participants, if the sheep object overlapped with any of the wolves, the trial would also end. For participants in the distracted condition, wolf objects could be ignored. A trial would be successfully completed after all 40 targets were cancelled. The game was thus an exhaustive search task, with no opportunity to leave a trial when target prevalence reduced. At the end of each trial an appropriate success or error feedback message was displayed, and the next trial was initiated by pressing a “continue” button. To complete a block of each experimental condition, 15 correct trials were required.
Procedure
Participants self-selected the experiment via their https://prolific.co account, and were then directed to the URL of the experiment starting page at https://maltacogsci.org. Here they were shown an introductory screen containing the name of the experiment and identifying the Department of Cognitive Science, University of Malta, as the institution conducting the study. To proceed, participants were asked to navigate to the next page which contained a detailed information and consent form. In order to proceed to the experiment itself, they were required to explicitly confirm their consent. A final screen then provided a reminder of the instructions and that 15 trials of the first condition would follow. After 15 successful trials, a new instruction screen provided details of the target mapping for the conjunction condition. Participants needed to complete 15 of those trials in order to finish the experiment. Block order was fixed, as this factor had not been found to qualitatively affect the pattern of foraging results in our previous work (see Thornton et al., 2019 for a detailed discussion) and in an online context, having the less demanding task first was useful from a familiarisation standpoint.
Data analysis
Our primary dependent variable was the average number of runs. As noted above, a “run” corresponds to the sequential selection of targets of the same category. With 40 targets divided into 2 categories, the number of runs on a given trial could vary between 2 and 40. We also examined other dependent variables which have proven sensitive measures of foraging behaviour. These included inter-target times (the time elapsed in milliseconds between two successive target selections) and inter-target distances (the distance in pixels between two successive target selections). On each trial, we also assessed the distance between the sheep object and the closest wolf. This latter measure—Wolf Distance—can provide an indication of whether hunted participants are risk taking or risk averse, with respect to the predator objects. Lastly, search organization was assessed by calculating the “best-r” (Woods et al., 2013) that assesses the degree to which target selections were pursued orthogonally (either horizontally or vertically). We calculated the correlation coefficient r1 between the x coordinates of all targets in a trial with the order in which they were selected, and the correlation coefficient r2 between the y coordinates of all targets in a trial with the order in which they were selected. The best-r corresponds to the higher of these two correlation coefficients.
All dependent variables were analysed using the same 2 (Predation: Hunted/Distracted) × 2 (Wolf Behaviour: Pack/Lone) × 2 (Target: Feature/Conjunction) × 5 (Wolf Velocity) mixed ANOVA with the first two factors as between subjects and the second two as within subjects, repeated measures. Full details of all analyses can be found in the Open Science Framework (OSF) supplementary material associated with this paper at https://osf.io/jwn8f/, with the text reporting the main factors of interest.
Results
Figure 2 summarises the main findings in terms of the interaction between Predation (Distracted/Hunted) and Target (Feature/Conjunction) for each of the dependent variables. Panel a shows that when target identification was easy (Feature condition), both groups of participants switched frequently between target categories, with the number of runs approaching half the total targets (i.e., 20), indicating random selection. Increasing the difficulty of target selection (Conjunction condition) led to a general drop in the number of runs, giving rise to a main effect of Target, F(1,44) = 235.5, p < 0.001, ηp2 = 0.58. Of most interest however, is the nature of the Predation × Target interaction, F(1, 44) = 7.18, p = 0.01, ηp2 = 0.04. Specifically, the reduction in the number of runs when target selection becomes more difficult is more pronounced for the distracted participants than the hunted participants, the opposite of the pattern we had predicted. Aside from the simple main effect of Predation, F(1,44) = 10.7, p = 0.002, ηp2 = 0.09, there were no other significant effects in the analysis of run patterns (see OSF supplementary materials for full descriptive statistics and ANOVA details).
Turning to the additional dependent measures, the only other Predation × Target interaction occurred for best-r, F(1, 44) = 5.7, p = 0.021, ηp2 = 0.11. As can be seen in Panel b, while search organisation was reduced for both groups of participants during conjunction foraging, distracted participants initially had more regular patterns during the less-demanding feature condition. Panels c–e show the expected main effects of Target for inter-target distances F(1,44) = 69.5, p < 0.001, ηp2 = 0.61, inter-target times, F(1,44) = 39.9, p < 0.001, ηp2 = 0.48 and number of selection errors, F(1,44) = 12.1, p = 0.001, ηp2 = 0.03. That is, participants moved greater distances, selected more slowly and made more selection errors during conjunction than feature foraging.
However, in terms of predation, these measures only gave rise to two significant effects. First, as shown in Panel c, hunted participants generally moved greater distances between selections than distracted participants, giving rise to a main effect of Predation for inter-target distance, F(1,44) = 5.4, p = 0.025, ηp2 = 0.11. Second, there was a significant Predation × Wolf-behaviour (Pack/Lone) interaction for inter-target times, F(1, 44) = 6.2, p < 0.017, ηp2 = 0.12. While full details of this pattern are given in the OSF supplementary materials, we note that the effect appears to be driven by the distraction condition, where selection speed was significantly slower in the lone wolf condition than the pack condition, (p < 0.05, Tukey HSD). Such slowing likely arises when the “ignored” lone wolf approaches the sheep and occludes possible target items. In contrast, the rate of target responses increased slightly for hunted participants in the lone wolf condition, although post-hoc comparisons with the pack condition were not significant, (p = 0.92, Tukey HSD).
Figure 3 confirms that participants were taking action to avoid being eaten, with the distance to the nearest wolf object at the time of selection, being consistently greater for hunted than for distracted participants, giving rise to a main effect of Predation on Wolf Distance, F(1, 44) = 38.1, p < 0.001, ηp2 = 0.35. Remaining with this dependent measure, while neither the Wolf behaviour (Pack/Lone) nor the Wolf Velocity manipulations showed any predation-foraging patterns with respect to the number of runs, there was a clear impact on Wolf Distance. Specifically, there were significant main effects and two-way interactions (see OSF supplementary material) which were in turn qualified by the Predation × Wolf Behaviour × Wolf Velocity interaction, F(4, 176) = 4.1, p < 0.01, ηp2 = 0.01, shown in Panel b of Fig. 3. The fairly linear increase in distance seen for both the distracted and hunted groups during the pack condition could be an artefact, reflecting the greater distance travelled by the higher speed wolves. However, during the lone wolf condition, the two lines diverge. For hunted participants, Wolf Distance continues to linearly increase, indicating attempts to avoid being eaten. For distracted participants, the opposite pattern occurs as the “ignored” lone wolf converges on the sheep, and does so more effectively at higher speeds. This pattern provides direct evidence that hunted participants were taking active measures to increase the gap between themselves and the wolves.
Returning to the run patterns shown in Fig. 2a, it is clear that there is considerable variation in performance, particularly in the conjunction condition. A consistent finding in many previous studies from our group, and other labs, has been the existence of subsets of individuals who continue to forage randomly under conjunction conditions (e.g., Clarke et al., 2018; Jóhannesson et al., 2017; Kristjánsson et al., 2014; Tagu & Kristjánsson, 2020; Thornton et al., 2020) The foraging patterns of such individuals is clear to see in the raw data, by plotting the run length for each trial as a function of condition (see Kristjánsson et al., 2014; Fig. 4). Here we provide the equivalent individual plots in OSF supplementary material. As a more concise summary, however, Fig. 4 shows how our 48 participants would be categorised according to whether more than 50% of their conjunction trials are random (switch focused) or non-random (run focused) using a Bonferroni-corrected one-sample runs test (for more details, see Kristjánsson et al., 2019). It is immediately clear that Predation has a large impact on such categorisation, with switch focused foraging being much more prevalent for hunted than the distracted participants. In the General Discussion, we further discuss the possible causes and consequences of such individual foraging behaviour.
Finally, while our main analysis has focused on between-group comparisons, it is also useful to look specifically within the hunted participants. We performed a median split based on the overall number of times participants were eaten by the wolves, to produce a low-surviving “food-focused” group (M_eaten events = 13.9, SD = 5.8) and a more successful “wolf-focused” group (M_eaten events = 4.7, SD = 1.6), Welch t(14.1) = 5.4, p < 0.001. We examined performance across the same dependent variables used in the main analysis to explore whether success in avoiding the wolves related to other aspects of foraging behaviour, but there were no clear interactions with this survival variable (see OSF supplementary material for full details). We note that as the “food-focused” individuals would have initiated many more trials than the “wolf-focused” group—trials were terminated with each collision—this additional time and effort does not appear to have affected run behaviour. This is important as it suggests that overall time-on-task—which would have been considerably longer for hunted than distracted participants—is unlikely to affect foraging patterns.
Discussion
Using an online protocol, we replicated our previous findings that increasing attentional demands using a feature/conjunction manipulation leads to less random foraging behaviour (e.g., Kristjánsson et al., 2014; Tagu & Kristjánsson, 2020; Thornton et al., 2019, 2020). The primary goal of this study, however, was to examine whether foraging patterns changed when participants also had to monitor and avoid potential predators. While our simulated risk of predation manipulation clearly affected performance, it was not in the way we had predicted. Rather than showing a reduced tendency to switch between target categories—the expected effect of increased attentional load—hunted participants continued to alternate, using more frequent, shorter runs than the distracted participants. How might we explain this finding?
One possibility is that the simulated “risk” of predation in our task modulated levels of alertness/arousal (Kahneman, 1973; Posner & Petersen, 1990; Sturm & Willmes, 2001; Yerkes & Dodson, 1908), counteracting the costs of having to both select targets and monitor for wolves. The presence or approach of the predator objects may have actually improved “attentional control” (Kane & Engle, 2003; Unsworth & Robison, 2017), allowing hunted participants to switch more frequently and more efficiently between complex target categories. The effects of phasic changes in alertness and arousal are central to recent attempts to explain individual differences in human cognitive performance (Esterman & Rothlein, 2019; Petersen et al., 2017; Unsworth & Robison, 2017) and more generally play a role in standard capacity models of attention (e.g., Kahneman, 1973) and other relevant models of behaviour (e.g., Aston-Jones & Cohen, 2005; Gray, 1990). The specific suggestion here—which we return to in the General Discussion—is that within-trial modulation in levels of arousal/alertness could have a direct impact on the creation, maintenance and selection of WM search templates during foraging.
Two more directly testable alternative explanations also suggest themselves. First, if attention has to be switched back and forth between target selection and wolf monitoring—as suggested by the MOT/Search study of Alvarez et al. (2005)—then maintaining the focus on a single target category may become more difficult or even impossible, raising the likelihood of a switch. Second, being forced to quickly move from one area of the display to another due to the approach of a dangerous wolf, could simply increase the salience of target items in the proximity of the new landing site, overcoming any tendency to use extended run behaviour. In Experiment 2, we designed a task variant that should help to distinguish between these attention switching and avoidance explanations.
Before leaving Experiment 1, however, we should comment on two other aspects of the results. First, in contrast to our original iPad studies (e.g., Jóhannesson et al., 2017; Kristjánsson et al., 2014), we found little evidence of fully exhaustive category selection during Conjunction foraging in the online task, even for participants in the distracted group. That is, while run length clearly increased when target selection was more demanding, few of our distracted participants consistently selected all of one target category before proceeding to the next (see OSF supplementary figures). We note that in a previous study that used a very similar display and response methodology (Thornton et al., 2019) we also found reduced use of exhaustive runs, which we suggested was an indirect consequence of extended inter-target response times. Specifically, when foraging tempo is quite slow—in the current Experiment 1 average inter-target times are all > 600 ms (Fig. 2)—we would thus expect to see a reduced tendency to use extended runs (see also Thornton et al., 2020 for further discussion).
Second, our attempts to modulate risk by increasing Wolf Velocity or changing Wolf Behaviour were largely unsuccessful, at least in terms of their impact on run patterns. Hunted participants did systematically adjust their distance from the nearest wolf as a function of Wolf Velocity and Wolf Behaviour (Fig. 3), but this did not impact run behaviour. In the General Discussion we suggest some additional ways in which the predictability and/or behaviour of predator objects could be modified in order to increase perceived risk.