Attention deficits linked with proclivity to explore while foraging

Abstract

All mobile organisms forage for resources, choosing how and when to search for new opportunities by comparing current returns with the average for the environment. In humans, nomadic lifestyles favouring exploration have been associated with genetic mutations implicated in attention deficit hyperactivity disorder (ADHD), inviting the hypothesis that this condition may impact foraging decisions in the general population. Here we tested this pre-registered hypothesis by examining how human participants collected resources in an online foraging task. On every trial, participants chose either to continue to collect rewards from a depleting patch of resources or to replenish the patch. Participants also completed a well-validated ADHD self-report screening assessment at the end of sessions. Participants departed resource patches sooner when travel times between patches were shorter than when they were longer, as predicted by optimal foraging theory. Participants whose scores on the ADHD scale crossed the threshold for a positive screen departed patches significantly sooner than participants who did not meet this criterion. Participants meeting this threshold for ADHD also achieved higher reward rates than individuals who did not. Our findings suggest that ADHD attributes may confer foraging advantages in some environments and invite the possibility that this condition may reflect an adaptation favouring exploration over exploitation.

1. Introduction

All mobile organisms forage for resources such as food, water and mates [1,2]. Foraging for food that is difficult to find or extract has been hypothesized to be a major driver of the evolution of intelligence [3–5]. Foraging models have recently been applied to human behaviour in multiple disciplines including cognitive psychology [6], cognitive neuroscience [7,8] and computer science [9,10], validating an approach grounded in mathematical ecology to study the computations and mechanisms that evolved to regulate the tradeoff between exploitation and exploration [11].

Patch foraging is the iterated accept-or-reject decision to stick with a known but potentially depleting option or disengage and search for something that might be better [12]. This description characterizes a wide array of decision contexts, including when to search for resources like food, water, minerals or sexual encounters; internal searches through concepts [13], memory [14] or strategies [15,16]; or searches for abstract resources such as for information [9,17,18] or reputation. These decisions can be solved using simple algorithms derived from optimal foraging theory [1], such as the Marginal Value Theorem (MVT) [19]. These algorithms dictate that individuals should leave a depleting resource patch when local intake rates fall below the average for the environment. Hundreds of species tested—from bees [20] to birds [21] to monkeys [22] to humans [23]—either quantitatively or qualitatively behave in accordance with the predictions of such models, suggesting that evolution long ago settled on a near-optimal solution to the challenge of determining when to abandon a depleting resource and search for a new one.

Whether searching for external, internal or abstract resources, foraging requires balancing a tradeoff between exploring for new options and exploiting known ones. Attention deficit hyperactivity disorder (ADHD) is a neurodevelopmental disorder affecting 11% of children [24] and 4.25% of adults [25]. Symptoms include distractibility as well as restlessness and excessive movement [26]. These characteristics may impact foraging behaviour by promoting a tendency to abandon the current resource patch to search for a new one [11]. This tendency to explore while foraging might extend to other behaviours such as cycling more frequently between information sources in the classroom or sources of stimulation in the home environment [11]. Human populations, such as the Ariaal tribe in Africa, that live nomadic lifestyles favouring exploration are characterized by genetic mutations implicated in ADHD [27], consistent with the hypothesis that this condition may impact foraging decisions in other populations. In addition, the neurological and genetic basis of ADHD has a well-established noradrenergic component [28–30] linked to task focus [31] and foraging behaviour [32]. Given the link between distractible, impulsive behaviour and atypical activity in noradrenergic circuitry, we predict that individuals with ADHD or ADHD-like phenotypes will both leave patches earlier than predicted by foraging models like the MVT and also earlier than individuals lacking these phenotypical characteristics.

We sought to test these predictions empirically by studying the relationship between foraging behaviour and ADHD self-report scores in a large population of adults. We measured ADHD-like symptoms using the Optimal RiskSLIM DSM-5 ASRS Screening Scale (ASRS) [33]. We designed an online incentive-compatible virtual berry-picking game to assess the impact of ADHD symptoms on foraging behaviour. We tested the pre-registered prediction that stronger self-reported ADHD symptoms would be associated with shorter patch residence times, less optimal behaviour and lower cumulative rewards in our online foraging task. In our sample of 457 participants, time in patch, travel time and harvested reward all influenced the decision to continue to forage at a patch or to depart and travel to a new patch. Overall, participants stayed longer in patches than predicted by MVT. We found that participants with elevated scores on the self-report ADHD scale exhibited shorter patch residence times than those whose scores were not elevated. Surprisingly, ADHD screen-positive participants also achieved higher reward rates and their patch-leaving decisions were more closely aligned with those predicted by the MVT, although they still stayed longer in patches than the MVT predicts.

ADHD is thus associated with shorter patch dwell times and higher reward rates in the online foraging environment implemented here. These findings are consistent with systematic amplification of activity within neural circuits involved in exploration, such as the default mode network (DMN) [12,34], and changes in levels of the neuromodulator norepinephrine broadcast from the locus coeruleus [31,35,36], both implicated in individuals with ADHD [28–30,37]. Our findings resonate with field studies highlighting a potential adaptive fit between population variation in genes linked to ADHD and nomadic lifestyles that favour exploration [27], endorsing approaches to variation in human behaviour grounded in evolutionary medicine [38] and computational psychiatry [39].

2. Methods and materials

(a) . Task

We recruited a representative US sample (n = 506) via Prolific (see https://www.prolific.com/) to complete a virtual patchy foraging task programmed in Javascript. All methods were approved by the Internal Review Board of the University of Pennsylvania. After preprocessing, 49 individuals were excluded from final analyses for the following criteria. First, they either failed pre-defined survey attention checks or did not complete every survey. Further, participants were excluded if they completed less than 25 trials (i.e. an exploit or explore decision) throughout the session, not counting incomplete patches (see below). Finally, we excluded participants that were believed to have misunderstood the instructions given their comments (e.g. ‘I don't know if I understood it’ or ‘I did not understand until the last round’). Note that exclusion criteria were slightly changed compared to the pre-registered criteria (see below). Of the final 457 participants (µ_age = 45.63 y.o., σ_age = 16.09), 232 identified as ‘male’, 217 identified as ‘female’, 5 identified as ‘other’, and 3 chose not to provide their gender. The final sample was composed of 71% (n = 326) White Caucasian (non-Hispanic), 14% (64) African-American or Black, 8% (38) Asian, 4% (19) Spanish, Hispanic or LatinX, <1% (3) American Indian or Alaskan Native, and less than 2% (7) unreported.

In the task, a ‘valid patch’ begins with an exploit decision, in which the user collects berries, and ends with an explore decision, in which the user travels to the next patch. Note that any patches where the participant immediately explored was considered a valid decision, too, and thus included in the analysis. Any ‘incomplete patch’ that did not end in the user leaving the patch (explore) was removed. These incomplete patches occurred at the end of session blocks in which the current block timed out after 2 min.

Participants were sent to the training and task via a link in Qualtrics (see https://www.qualtrics.com/uk/?rid=ip&prevsite=en&newsite=uk&geo=BE&geomatch=uk) after completing an informed consent form and several questionnaires. Before starting, participants received detailed written instructions on the task and watched a video demonstration. To familiarize themselves with the task, they also completed a 1 min training (30 sec per block type) before starting the actual experiment. The instructions prior to the training were (emphasis in italics were coloured in original):

• ‘In this experiment, your goal is to collect as many berries as you can in the amount of time you are given.

• First, hover over the centre box until it fills up completely. You may have to wiggle your cursor once for the centre box to recognize your mouse.

• Then, on each trial, you will either choose to stay at the bush you are at or leave the current bush and travel to a fresh bush.

• To pick berries from the current bush, hover your cursor over the bush next to the centre box.

• The number of berries you pick at each attempt will be indicated next to the bush.

• As you keep picking from the same bush repeatedly, the number of berries you pick at each attempt will go down as you will be depleting this bush.

• To travel to the next bush, mouse over the box on the opposite side of the bush.

• You can choose to leave the current bush to travel to a fresh bush at any time. When you do that, you will need to wait while you travel to the next bush.

• However, once you arrive at the new bush, the amount of berries you pick at each harvest will reset to the initial value.

• You will play different blocks in this game. The environment will change between these blocks.

• There is an unlimited number of bushes in the environment, but a finite amount of time. Try to collect as many berries as you can!’

After completing the experiment, participants returned to Qualtrics to finish further questionnaires. Foraging data were stored on a secure AWS server.

During the task, participants collected as many berries as possible in 8 min (4 blocks, 2 min each). The reward function for the task started at 7 plus Gaussian noise, ε, for both long and short travel time patches for all patches:

$$r_0=7+\epsilon ,$$

where ε ∼ N(0, 0.25) and then, for each choice to harvest reward, r_i, decremented by 0.5 plus noise:

$$r_i=r_{i-1}-0.5+\epsilon ,$$

where again ε ∼ N(0, 0.25). At the end of the experiment, participants received 0.3226 cent per berry collected (up to $3 in total) in addition to their base participation compensation of $4.

Participants were randomly assigned to one of two orders of travel times by block: order 1 [block 1: 1 s travel time, block 2: 5 s, block 3: 1 s, block 4: 5 s] or order 2 [block 1: 5 s travel time, block 2: 1 s, block 3: 5 s, block 4: 1 s]. Each block lasted for 2 min. Participants foraged through as many patches as they wished and were able to in those 2 min; patches at the end of each block that did not terminate with a leave decision were removed from analysis. Participants completed an average of 119.95 ± 1.79 trials on average during the task. The maximum number of trials was 179 and the minimum was 25 trials, for a range of 154 trials.

(b) . Survey

Participants completed the Optimal RiskSLIM DSM-5 ASRS Screening Scale [33], consisting of 6 items. The items were:

How often do you have difficulty concentrating on what people are saying to you even when they are speaking to you directly?

How often do you leave your seat in meetings or other situations in which you are expected to remain seated?

How often do you have difficulty unwinding and relaxing when you have time to yourself?

When you're in a conversation, how often do you find yourself finishing the sentences of the people you are talking to before they can finish them themselves?

How often do you put things off until the last minute?

How often do you depend on others to keep your life in order and attend to details?

Response options were never, rarely, sometimes, often, and very often. To score the survey, a complex, linear, proprietary scoring function was received from its authors [33]. We are not at liberty to share the method here, but researchers can obtain it after completing a use agreement with the scale authors (see https://license.tov.med.nyu.edu/product/asrs-dsm-5 for instructions). This is currently free for all academic purposes. If there are issues obtaining the scoring method, researchers could revert to scoring the scale using a simple sum (see electronic supplementary material).

The ADHD self-report symptom checklist from Ustun et al. [33] used a machine learning algorithm, the risk-calibrated super-sparse linear integer model (RiskSLIM), to generate an optimal scoring algorithm for the ADHD Self-Report Screening Scale survey based on the updated DSM-5 criteria for ADHD. The scale was validated by comparison to a managed care sample of ADHD diagnosed adults (n = 218), a household sample from the National Comorbidity Survey Replication consisting of a face-to-face survey and follow-up (n = 119) and an NYU-Langone sample (n = 300). At the recommended scoring threshold (≥14) for a positive screen, the ASRS has good sensitivity (91.4%), very good specificity (96.0%), an area under curve (AUC) of 0.94, with 67.3% of screen-positive participants meeting DSM-V criteria for ADHD among a combined cohort from the National Comorbidity Survey-Replication and a managed care sample, along with good sensitivity (91.9%), moderate specificity (74.0%), an AUC of 0.83 and a proportion of 82.8% of screen-positive participants who met DSM-V criteria for ADHD among a clinical sample from NYU-Langone Medical Center [33]. The scale has been further validated by other groups [40–44].

(c) . Preregistration

Our design, hypotheses and analyses were preregistered: see https://aspredicted.org/blind.php?x=KLK_XLL.

We made a few justified changes and additions to the pre-registered analysis approach. We analysed total reward per patch and reward rates in both conditions (short versus long travel time) rather than total reward over the entire experiment, because foraging theory predicts differences between patch types. We calculated optimal behaviour and deviations from optimal behaviour in a way that differed from the preregistered version because we considered the approach by Cowie [45] to be more accurate than the pre-registered proposal. Specifically, Cowie's method presents a first principles way to calculate optimal behaviour. In the binomial regression analysis including ASRS, we included ‘cumulative reward in patch’ as an additional predictor because we considered it relevant to the model. We did not include trait and state anxiety or positive and negative emotions as control variables. The categorization of participants into ADHD-positive and ADHD-negative was not preregistered. However, since the ASRS scale uses a complex linear scoring algorithm to classify participants as positive or negative, it made sense to use this qualitative and clinically meaningful distinction between participants in our analysis. ‘Give up time’ (pre-registration) is referred to as ‘time in patch’ or ‘patch residence time’ in the current manuscript, as we considered these to be more accurate terms. The analysis of the link between ASRS with creativity and mind wandering will be reported separately for brevity here. Finally, exclusion criteria for participants were slightly improved. For example, in addition to the pre-registered criteria, we also excluded participants who commented that they did not understand the task. Moreover, trials belonging to incomplete patches at the end of each session were excluded before applying the ≥25 trials criterion.

(d) . Data analysis

All data were analysed in MATLAB (Mathworks, Natick, MA) using custom software. For tests involving comparisons across 1 s and 5 s travel times, 1 subject was rejected who never completed a 5 s travel time patch, leaving n = 456.

For patch residence time comparisons, we averaged the median patch residence time across subjects, to control for outliers. We used within-subject paired t-tests to compare the average patch residence time for 1 s and 5 s travel time patches. Time in patch was the amount of time from the start of the first trial in a patch to the decision to depart a patch.

For reward-related comparisons, we used both cumulative reward (which previous work has shown to be the computationally relevant variable for decisions to harvest rewards in patch [46]) and reward rate (the classic currency for foraging choices in optimal foraging theory [1]). Cumulative reward was defined as the running sum of berries collected within a patch, and reward rate was defined by dividing cumulative reward by the cumulative time in patch. We used within-subject paired t-tests between travel times to compare cumulative rewards and reward rates. We also examined differences in cumulative reward across all blocks in the entire session for ADHD screen-positive and screen-negative participants.

To assess the effects of patch number in session, we first segregated total time in patch and reward rate on patch leave trials by travel time block (1 s or 5 s). We then separately regressed time in patch and reward rate upon departing patches against patch number in session using ordinary least squares (OLS) regressions. Patch number in session was defined as the patch number in the concatenated blocks segregated by travel time.

ADHD self-report scale (ASRS) scores were correlated with both time in patch and reward rates. To determine the influence of ASRS on time in patch, we used a mixed-effects linear regression, with random effects of the side of the screen of the choice target, patch number in session, and participant identity and fixed effects of travel time, cumulative reward and ASRS score, with all interactions, and fixed effects of nuisance covariates age, income and gender, with all interactions. In Wilkinson notation,

$$\begin{array}{rl}& \mathrm{Time}\hspace{0.17em}\mathrm{in}\hspace{0.17em}\mathrm{patch}\sim \mathrm{CumulativeReward}\times \mathrm{TravelTime}\\ & \times \mathrm{ASRSScore}+\mathrm{Age}\times \mathrm{Gender}\times \mathrm{Income}\\ & +(\mathrm{Side}+\mathrm{PatchNumberinSession}|\mathrm{ParticipantNumber}),\end{array}$$

which models patch number in session as a random slope and participant identity as a random intercept. To determine the influence of ASRS on decisions to leave patches, we used a mixed-effects binomial regression, with random effects of the side of the screen of the choice target, patch number in session, and participant identity and fixed effects of time in patch, travel time, cumulative reward and ASRS score, with all interactions, and fixed effects of nuisance covariates age, income and gender, with all interactions. In Wilkinson notation,

$$\begin{array}{rl}& \mathrm{ProbabilityChooseLeave}\widetilde{\mathrm{TimeInPatch}}\\ & \times \mathrm{CumulativeReward}\times \mathrm{TravelTime}\times \mathrm{ASRSScore}\\ & +\mathrm{Age}\times \mathrm{Gender}\times \mathrm{Income}\\ & +(\mathrm{Side}+\mathrm{PatchNumberinSession}|\mathrm{ParticipantNumber}),\end{array}$$

which again models patch number in session as a random slope and participant identity as a random intercept. Finally, we used Student's t-tests to compare ADHD screen-positive with screen-negative time in patch, and reward rate distributions for the two groups were compared using a Kolmogorov-Smirnov test.

(e) . Optimal behaviour

First, an exponential curve was fit to the cumulative intake curve defined by the gain function (fit function in MATLAB). The gain function g(t) for time in patch t is well-described by an exponential curve, which can be used to solve from first principles the optimal leave time [cf. 45]:

$$g={\beta }_1\ast (1-{\mathrm{e}}^{-{\beta }_2\ast t}).$$ 2.1

While this continuous and differentiable function is at odds with the discrete real series of rewards, describing the series of rewards as an exponential permits a first principles solution to the optimal patch stay time. To find the optimal leave times, the first derivative of g(t) is set equal to the gain function divided by the average travel time and time in patch:

$${\displaystyle \frac{\mathrm{d}g}{\mathrm{d}t}={\beta }_1\ast {\beta }_2\ast {\mathrm{e}}^{-{\beta }_2\ast t}={\displaystyle \frac{g}{tt+\hat{t}},}}$$ 2.2

for travel time tt and optimal leave time $\hat{t}$. Substituting (2.1) in for g and rearranging terms yields

$$e^{{\beta }_2\hat{t}}-{\beta }_2\hat{t}={\beta }_2\ast tt+1.$$ 2.3

We next used the McLaurin expansion for e, eliminated constants, dropped the fourth-order and higher terms, subtracted out the first-order term and rearranged to arrive at

$$3\ast {\hat{t}}^2+{\beta }_2\ast {\hat{t}}^3-{\displaystyle \frac{6\ast tt}{{\beta }_2}=0.}$$ 2.4

We used the best-fit β₂ from our exponential fit (2.1) to the series of average returns. Then, using the average (tt = 3 sec), short (tt = 1 sec) and long (tt = 5 sec) travel times in turn, we determined the optimal travel time $\hat{t}$ by solving (2.4) with numerical methods (fmincon in MATLAB subject to nonlinear constraints determined by the McLaurin expansion).

3. Results

(a) . Foraging task and behaviour

Adults (n = 457) performed an online foraging task (figure 1). Participants were recruited via Prolific and were a representative US sample (47% female, average age = 46y). Participants were instructed to collect as many berries as possible in 8 min (4 blocks, 2 min each). Participants first initiated a trial by mousing over an intertrial interval (ITI) box. After initiating a trial, participants chose between mousing over an image of a bush to stay and collect rewards at a given patch, in which payoffs declined with each successive choice, or mousing over an empty-sided box to travel to a new patch, which took time. Each participant was tested in two conditions that differed only in the time it took to ‘travel’ to the next bush (either 1 sec or 5 sec). Travel time was indicated by the visual distance between bushes (see figure 1, inset). The order of conditions was counterbalanced across participants. After deciding to depart a patch, bushes moved vertically down the screen, simulating travel, and the participants faced a new set of decisions between staying in a patch and departing, with the choice options—the bush and the empty box switching their sides on the screen.

Figure 1.

Virtual patch foraging task. Participants began each trial by moving their mouse cursor over an intertrial interval (ITI) box. After 1 s delay, the ITI box turned green. Participants next decided to stay at a given patch and gather points ('exploit' decision) or to replenish the patch in exchange for a travel time delay ('explore' decision). Travel times were either 1 s or 5 s, in blocks. Inset: screenshots of the two foraging environments, with 1 s travel times on the left and 5 s on the right.

Consistent with predictions from foraging theory's MVT [19], participants abandoned patches in a short travel time (1 sec) environment earlier than they left patches in a long travel time (5 sec) environment (paired t-test on median patch residence time by subject, p < 1 × 10⁻¹⁷, t (d.f. = 455) = −9.1942; figure 2a). Participants earned significantly more rewards on average in long travel time environments (paired t-test, p < 1 × 10⁻⁴⁴, t (d.f. = 455) = −15.7081; figure 2b), but longer residence times in long travel time patches did not yield significantly higher average reward rates (RR; average RR_short = 1.4597 ± 0.0271 points s⁻¹, average RR_long = 1.4765 ± 0.0242 points s⁻¹; paired t-test, p > 0.22, t (d.f. = 455) = −1.2113).

Figure 2.

Human participants' (n = 457) behaviour on the virtual patch foraging task. (a) Mean median total time in patch (s) for the 1 s travel time (left) and 5 s (right). (b) Mean rewards harvested for 1 s travel time (left) and 5 s (right).

Average reward rates were significantly less than the theoretically optimal reward rates in both short travel time environments (RR_{short,optimal} = 2.6118 points s⁻¹; one-sample t-test, p < 1 × 10⁻¹⁶⁹, t (df = 456) = −45.0179) and long travel time environments (RR_long,optimal = 2.2256, p < 1 × 10⁻¹¹⁹, t (d.f. = 455) = −32.4675). This suboptimal behaviour resulted from participants staying for too long in each patch (time in patch TiP; short patches: TiP_short = 19.5430 ± 0.1360 s, TiP_{short,optimal} = 6.0603 s; long patches: TiP_long = 25.5998 ± 8039 s, TiP_long,optimal = 12.9049 s), as observed in prior studies in other animals including nonhuman primates [22,47–50].

In addition to these travel time and reward rate effects, exploratory data analysis uncovered an effect of patch number in session. Participants performed up to four blocks of patches, starting with a block of 1 s patches, then a block of 5 s patches, then 1 s and then 5 s again, or the reverse order (5, 1, 5, 1; see §2a). Some participants showed an effect of patch number in session on patch residence time and reward rate. Segregating by travel time, some subjects showed an effect of patch number in session for patch residence time (1 s patches: OLS, 113/457 participants significant (p < 0.05) slope, 74 negative; 5 s patches: OLS, 85/457 participants significant slope, 53 negative; 168 / 457 unique participants showed some effect). Two example participants are shown in figure 3a; the left participant (no. 4) did not show an effect of patch number in session (1 s patches: OLS, β_slope = 0.1741 ± 0.1490 s/patch number, p > 0.25; 5 s patches: OLS, β_slope = 0.1860 ± 0.4653 s/patch number, p > 0.69), whereas the participant on the right (no. 270) did (1 s patches: OLS, β_slope = −0.8646 ± 0.1550 s/patch number, p < 0.001; 5 s patches: OLS, β_slope = −1.5769 ± 0.2684 s/patch number, p < 0.001). For reward rate, some subjects showed an effect of patch number in session (1 s patches: OLS, 131/457 participants significant (p < 0.05) slope, 126 positive; 5 s patches: OLS, 96/457 participants significant slope, 84 positive; 187/457 unique participants showed some effect). These same two participants illustrate these effects on reward rate (figure 3b), with the participant on the left (no. 4) showing no effect of patch number in session on reward rate (1 s patches: OLS, β_slope = 0.0114 ± 0.0291 (berries s⁻¹)/patch number, p > 0.7; 5 s patches: OLS, β_slope = 0.0129 ± 0.0163 (berries s⁻¹)/patch number, p > 0.45), unlike the participant on the right (no. 270) who did (1 s patches: OLS, β_slope = 0.0554 ± 0.0113 (berries s⁻¹)/patch number, p < 0.001; 5 s patches: OLS, β_slope = 0.1269 ± 0.0174 (berries s⁻¹)/patch number, p < 0.001). Of those participants who showed an effect of patch number in session, patch residence time tended to decrease and reward rates increase across sessions; in short, participants tended to perform better as they progressed through the task.

Figure 3.

Participants' patch residence time and reward rate on leaving a patch as a function of patch number in session, concatenated across blocks for 1 s and 5 s travel times separately. (a) Total time in patch (s) versus patch number in session, for 1 s (green) and 5 s (red) travel times. Participant no. 4 (left) did not show a patch number in session effect, whereas participant no. 270 (right) showed a steep decline in total time in patch as a function of patch number in session for both travel times. (b) Reward rate on patch leave (berries s⁻¹) versus patch number in session for 1 s (green) and 5 s (red) travel times. Participant no. 4 (left) did not show a change in reward rates as a function of patch number in session, whereas participant no. 270 (right) did, increasing their reward rate.

(b) . ADHD Self-Report Scale and foraging

We next examined how each of the task variables was related to participants' ADHD Self-Report Scale scores (ASRS [33]). We first tested for patch-level effects of ASRS by performing a mixed-effects linear regression of time in patch (n = 10 353 patches) against a random effect of patch number in session and participant, fixed effects of travel time, cumulative reward and ASRS scores and all interactions, and a number of nuisance covariates including age, gender and income and all interactions (see §2d). We reasoned that because of differences across subjects regarding patch numbers in session, covariate should be treated as a random effect. All fixed effect covariates were z-scored. We uncovered a main effect of cumulative reward (p ∼ 0, Bonferroni-corrected), although there was no main effect of ASRS scores. We found two-way effects of travel time and cumulative reward and of ASRS score and cumulative reward (p < 5 × 10⁻⁹ or less, Bonferroni-corrected). Finally, we also found a three-way interaction effect of travel time, cumulative reward and ASRS score (p < 0.0005, Bonferroni-corrected). Figure 4 depicts the regression betas for the covariates of interest.

Figure 4.

Regression betas for fixed effects of travel time, cumulative reward, and ASRS scores and all interactions on time in patch (n = 10 353 patches). *: p < 0.0005 Bonferroni-corrected.

We also examined trial-level effects of ASRS, testing our prediction that participants who score highly on the ASRS are more likely to choose to leave patches, all else being equal. We performed a mixed-effects binomial regression of all decisions to stay (= 0) or leave (= 1) a patch (n = 55 526 trials) against a random effect of patch number in session and participant, fixed effects of patch residence time, travel time, cumulative reward in patch and ASRS, including all interactions of the fixed effect covariates, and a number of nuisance covariates and their interactions (see §2d). All fixed effect covariates were z-scored. Consistent with our predictions, all main fixed effects of covariates of interest (p < 0.005 or less, Bonferroni-corrected) as well as all their two-way interactions (p < 5 × 10⁻⁵ or less, Bonferroni-corrected) were significant. The three-way interaction of time in patch, cumulative reward and ASRS score was also significant (p < 5 × 10⁻⁷, Bonferroni-corrected), as was the four-way interaction (p < 5 × 10⁻⁸, Bonferroni-corrected). Figure 5 depicts the betas for the covariates of interest.

Figure 5.

Binomial regression betas for fixed effects of time in patch, travel time, cumulative reward and ASRS scores, and all interactions on stay or leave patch decisions (n = 55 526 trials). *: p < 0.005 Bonferroni-corrected.

In the light of these significant interaction effects, we next compared participants who screened positive on the ASRS scale (n = 206) with those who screened negative (n = 251), the clinically relevant comparison. A positive screen—a score of 14 or more—indicates that the individual exhibits symptoms consistent with an ADHD diagnosis. Consistent with our predictions, ASRS screen-positive participants stayed in patches for significantly shorter durations than screen-negative (Student's t-test, p < 0.005, t (d.f. = 455) = −3.0085; mean median time in patch for positive TiP₊ = 19.3170 ± 0.9057 s; TiP_– = 23.3907 ± 0.9758 s). However, unexpectedly, they also attained significantly higher reward rates than screen-negative participants (Student's t-test, p < 0.00001, t (df = 455) = 4.5466; mean reward rate for positive RR₊ = 1.5896 ± 0.0356 points s^–1; RR_– = 1.3608 ± 0.0350), contrary to our prediction. Though ASRS screen-positive participants stayed less long in patches than screen-negative participants, and earned higher reward rates, they still tended to overharvest patches, staying significantly longer than the optimal leave times predicted by the MVT for both patch types (one-sample t-test against optimal TiP; mean median TiP for 1 s patches = 17.2336 ± 1.0061 s, p < 1 × 10⁻²², t (df = 204) = 11.1059; mean median TiP for 5 s patches = 21.2876 ± 0.9211, p < 1 × 10⁻¹⁶, t (df = 204) = 9.0785). Figure 6 illustrates the consistently shorter patch residence times in both conditions for ADHD screen-positive compared to screen-negative participants. Figure 7 illustrates the right-shifted distribution of ADHD positive reward rates after segregation by positive or negative ASRS screening. Cumulative rewards across sessions were significantly greater for ADHD screen-positive than screen-negative participants (Student's t-test, p < 0.0001, t (df = 455) = 4.1000; mean cumulative reward for screen-positive C₊ = 602.6460 ± 9.5222; screen-negative C₋ = 521.6306 ± 10.0790).

Figure 6.

Patch residence times for ASRS screen-negative (blue) and screen-positive (red) participants for 1 s travel time (left) and 5 s travel time (right) patches. Points are mean median total times in patch, error bars are ±1 s.e.m.

Figure 7.

Distribution of patch reward rates across all patches for ASRS screen-negative (blue) and screen-positive (red) participants. Large peaks at 0 are from participants departing patches immediately upon entry. Vertical dashed lines are optimal reward rates (see §2e): red: 1 s travel time optimal reward rate; black: global optimal reward rate; green: 5 s travel time optimal reward rate.

We performed two sets of analyses to provide further support for our conclusions. First, we ran the same regressions and approximated the comparisons using OLS with a simpler scoring scale for the ASRS screen: the sum of the scores of the answers to each item on the questionnaire. The results were almost identical to those reported above (see electronic supplementary material). Second, 24 of our participants reported a previous diagnosis of ADHD (of those 24, 18 screened positive using the RiskSLIM scoring sheet [33]). We again ran the same regressions and the analyses as above, comparing those 24 participants with the 428 participants who reported no previous ADHD diagnosis (5 participants who did not answer that question were left out). The results were once again largely confirmatory (see electronic supplementary material), albeit noisier than the outcome of the first set of confirmatory analyses, perhaps due to so few participants receiving previous ADHD diagnoses.

4. Discussion

Foraging poses a fundamental challenge for humans and other organisms [12]. Here, we investigated how humans choose between staying in a depleting resource patch or leaving for a new one. Participants made these decisions under short or long travel times between patches, reflecting rich and poor environments. As predicted by the marginal value theorem (MVT), participants stayed longer in patches when travel times were longer. Consistent with a range of evidence across species [45,47,49,51], including humans [52,53], participants stayed longer in patches than predicted by the MVT, potentially reflecting gathering information about the environment [54], a lack of competing demands due to predation or socializing, or misrepresentation of rewards [55] or time [49]. The influence of travel times on patch residence time resulted in significantly different total rewards collected but not significantly different reward rates between conditions.

We next tested for links between foraging decisions and ADHD Self-Report Scale (ASRS) scores. Correcting for nuisance covariates, we found that interactions between ASRS scores and cumulative reward, travel time and patch number in session predicted earlier patch leaving. A trial-level binomial regression also uncovered significant interactions between ASRS and cumulative reward, patch number in session and time in patch. Comparing ASRS screen-positive with screen-negative participants revealed shorter patch duration times and higher reward rates for screen-positive individuals. We followed-up and supported our findings with two sets of further analyses, performing the same set of analyses using the sum of the scores of the individual ASRS questions and comparing participants who had received a previous ADHD diagnosis with those who had not.

There are several strengths and weaknesses to our design. Our online design enabled gathering ADHD symptom self-reports from many participants during the height of the COVID-19 pandemic. However, our findings relied on self-report without monitoring task engagement. Finally, although the ASRS has been validated [33] and re-validated [40–44] and self-report scales have been found to be informative for detecting ADHD [56–58], clinical validation is still needed to confirm the relationship between foraging behaviour and ADHD.

In our sample, 206 of 457 participants, or 45%, screened positive for ADHD, approximately 10x higher than the rate of ADHD in the adult population. Several factors may account for this discrepancy. First, ADHD-screen positive participants were not diagnosed with ADHD; only a clinical assessment can confirm such a diagnosis. Second, we conducted our study online using Prolific, which may have elevated rates of positive screens, as was found using an older ADHD screen in a previous mTurk sample [59]. Finally, our data collection occurred during the COVID-19 pandemic, which negatively impacted mental health for broad swaths of the public [60], as confirmed by online self-reports [61]. Notably, the pandemic is linked to a global increase in ADHD symptoms [62]. While difficult to estimate, this context undoubtedly played a role in the frequency of ADHD symptoms reported in our sample.

Higher reward rates in ADHD screen-positive participants may reflect adaptive specialization in the capacity to focus. Attention is known to be regulated in part by dopaminergic receptors and transporters [63,64]. Animal models of ADHD show dopaminergic deficits in the basal ganglia and prefrontal cortex [65], and both the D2 [66] and the D4 [67] dopamine receptors are associated with ADHD. These same receptors are implicated in differences in nutrition and health in comparisons of nomadic and settled populations [27]; specifically, the interaction between lifestyle and allele length for the D4 receptor predicts healthier body mass. Longer alleles for the D4 receptor are a risk factor for ADHD [67]. Taken together, allele length for the D4 receptor is a risk factor for ADHD and correlates with better nutritive outcomes in nomadic populations. Consistent with our reported findings, we speculate that ADHD serves as an adaptive specialization for foraging [68–71], thus explaining its widespread prevalence and continued persistence in the human population. Individuals with ADHD may be more reward-seeking, consistent with the role of dopamine in motivated behaviour [72] as well as heightened impulsivity on delayed discounting tasks [73]. Individuals with ADHD may also be more exploratory as a result of increased noradrenergic drive originating in the locus coeruleus and broadcast to prefrontal cortex [29]. Typical individuals, by contrast, may be driven by a variety of motivations, including gathering more information about the environment through persistent foraging.

Our task focuses on core explore–exploit foraging decisions, a type of decision ubiquitous in the real world. For example, the Aché people of Paraguay hunt in accordance with the predictions of optimal foraging theory [74], as do Nahua mushroom foragers [75], and human shopping behaviour is also consistent with optimal foraging predictions [76–78]. However, there are key differences between naturalistic foraging behaviour and patch leaving tasks like ours. Notably, there are minor metabolic costs to our task, whereas naturalistic foraging involves substantial metabolic costs associated with leaving a patch and searching for a new one. In addition, our task required participants to gather points—an abstract resource that stands in contrast to basic alimentary or other resources like calories. Nonetheless, animals and humans show similar patterns of behaviour that accord with foraging theory, whether foraging for resources in the real world or foraging for food pellets or points on a computer in the laboratory [22,50,55,79].

Our task was designed to understand how participants make patch leaving decisions and whether participants with different cognitive phenotypes exhibit variability in those decisions. These decisions are characterized by tracking resource intakes and deciding to replenish that resource. In the real world, this decision would involve searching for a new resource instead of merely resetting it. Unlike the real world, our task does not require search (cf. the distinction between searching and harvesting in [80]). Nonetheless, by focusing on patch leaving decisions, our task provides fundamental insights on foraging choices without the confounding factors involved in searching for new patches.

ADHD-like cognitive phenotypes confer advantages or disadvantages depending on the environment. Foragers with ADHD-like traits may fare poorly in environments with multiple, learnable depletion rates (like the task used in [79]), by leaving patches before gathering enough information to learn the different patch types. In environments in which exhausted patches are renewed, primates are known to track renewal rates, such as the availability of fruit [81–83]. In these environments, ADHD-type foragers may be likelier to visit resources too early or too late due to noisier estimates of depletion rates, resulting in fewer rewards. However, in competitive environments where foragers must keep track of other foragers, impulsively leaving patches could yield a competitive advantage by enabling learning about competitors and capturing newly renewed resources first.

Many decision contexts differ from foraging ones by presenting agents with multiple options simultaneously rather than single depleting resources. We hypothesize that individuals with ADHD-type traits may struggle more in these contexts, a proposal backed by existing research [84]. For instance, a study by Frank et al. [85] used a reinforcement learning task to assess learning in neurotypical (healthy) individuals, ADHD patients on medication and ADHD patients off medication. The participants had to learn which of two arbitrary symbols was associated with more rewards by learning the expected value of selecting each option. In their task, outcomes were independent of previous choices, there was no depleting resource and choices were between options instead of accepting or rejecting a single offer. They found that patients off medication displayed lower choice accuracy and switched choices more frequently after rewarding feedback than patients on medication or healthy controls. This study illustrates how in an environment with symbolic associations between rewards and options, switching leads to maladaptive information gathering to learn those associations, in contrast to a patchy foraging environment like our task, where switching behaviour may be adaptive.

As a second example, a study by Dekkers et al. [86] investigated how ADHD participants and healthy controls made choices between a more certain ‘safe’ option and a ‘risky’ option, each with explicitly verbally communicated expected values. On some trials the safe option had the higher expected value, whereas on others the risky option had a higher expected value. ADHD participants less frequently selected the option with the higher expected value. This decision context involves a two-alternative forced choice where each decision was independent and non-depleting, differing markedly from the patch foraging task. Here, neither switching nor staying was an intrinsically adaptive strategy, and optimal decisions required information gathering and deployment (i.e. in reading and computing the expected value of options). Modern decision contexts often involve non-depleting, independent choices, usually framed in symbolic terms. Such tasks may pose specific challenges for individuals with ADHD-type traits, whose decision-making mechanisms may be better suited for foraging-like environments.

A burgeoning literature reveals links between foraging proficiency and attention to ADHD-like traits. Foraging for cryptic prey types places attentional demands that lead to selection of prey of a particular type [87], consistent with diet selection rules in optimal foraging theory [1]. In addition, a comparison of feature search and conjunctive search revealed that search type augments attentional demands and consequently foraging costs, resulting in differences in give-up times [88], a key foraging measure [89]. Finally, search paths are longer and more variable in individuals with greater ADHD symptoms [90]. Our findings build on this literature, suggesting that heightened ADHD traits may confer certain foraging advantages by lowering exploration thresholds.

The DMN, a distributed neural circuit with high resting-state activity [28,29], shows elevated activity during task performance in individuals with ADHD [30,31], potentially accounting for more frequent errors [29]. Neurons in posterior cingulate cortex (PCC), a central node of the DMN, signal decisions to explore [33] and the information driving these decisions [18]. In addition, electrical microstimulation in PCC promotes exploration of a non-default option [34]. ADHD is associated with noradrenergic activity [28–30], and stimulant medications like reboxetine [91] or atomoxetine [92] target these systems. Notably, stimulation of the locus coeruleus—the source of noradrenaline in the brain—provokes early patch leaving in rats [32], and treatment with reboxetine in humans induces earlier patch departures, higher reward rates at patch leaves and more optimal foraging [50]. The locus coeruleus projects to a number of DMN nodes, including posterior cingulate cortex [93,94]. Together, these findings suggest that noradrenergic activity ‘tunes’ circuits regulating explore–exploit decisions, resulting in advantages for ADHD-type individuals in foraging.

In conclusion, our virtual foraging task reveals that people adaptively follow predictions of optimal foraging theory. They adapted their foraging strategies to the local statistics of patches and stayed longer in long travel time patches compared to shorter travel times, as predicted by optimal foraging theory. Nevertheless, people generally overharvested patches and consequently did not conform to the optimal patch-leaving time specified by foraging theory. In addition, we discovered that participants that screened positive for ADHD more readily abandoned patches and achieved higher reward rates than did participants who screened negative. Given the over-staying displayed by participants overall, those with elevated ASRS scores made exploratory decisions that were more closely aligned with the predictions of optimal foraging theory, and, in this sense, behaved more optimally. The increased foraging proficiency of participants with ADHD-like behaviour observed here suggests that the prevalence and persistence of ADHD in human populations may serve an adaptive function in some environments.

Ethics

All methods were approved by the Internal Review Board of the University of Pennsylvania (IRB #823436).

Data accessibility

All data are available at https://doi.org/10.5281/zenodo.10178228 [95].

Supplementary material is available online [96].

Declaration of AI use

We have not used AI-assisted technologies in creating this article.

Authors' contributions

D.L.B.: data curation, formal analysis, software, writing—original draft, writing—review and editing; V.U.L.: conceptualization, investigation, methodology, project administration, supervision, writing—original draft, writing—review and editing; F.P.: data curation, writing—review and editing; N.A.: investigation, supervision; E.M.B.: writing—original draft, writing—review and editing; A.R.: conceptualization, software; M.L.P.: conceptualization, funding acquisition, project administration, writing—original draft, writing—review and editing.

All authors gave final approval for publication and agreed to be held accountable for the work performed therein.

Conflict of interest declaration

We declare we have no competing interests.

Funding

The authors report no financial conflicts of interest or disclosures.