An examination of sustained attention during complex multitasking scenarios

Jonathan C Rann; Amit Almor

doi:10.1186/s41235-025-00674-x

. 2025 Oct 7;10:67. doi: 10.1186/s41235-025-00674-x

An examination of sustained attention during complex multitasking scenarios

Jonathan C Rann ^1,^2,^✉, Amit Almor ^1,^2,³

PMCID: PMC12504179 PMID: 41057555

Abstract

We report results from two experiments that examined the time course of vigilance decrements during a demanding multitasking scenario. Specifically, we implemented a novel paradigm in two experiments in which a total of 123 participants performed a go-no-go target detection continuous performance test (CPT) task simultaneously with a driving-based tracking task. Growth curve analyses of the temporal trajectories of performance of both tasks revealed vigilance decrement effects that varied across CPT and tracking measures, and between different target presentation rate conditions. Our findings highlight the importance of executive function, arousal, and motivation in such dual-task performance and support a multifaceted approach combining elements from the cognitive overload, cognitive underload, and opportunity-cost models of vigilance decrements. Insights from this work can inform the design and development of complex operator–system interfaces and thus increase safety and effectiveness for operators during mission-critical situations.

Keywords: Sustained attention, Vigilance decrements, Multitasking

Significance statement

Many real-world occupations, in both the civilian and military worlds, require operators to maintain adequate levels of alertness to detect and respond to critical signals over long periods. However, this can be difficult due to the well-documented vigilance decrements, which are reflected in the gradual deterioration of several aspects of performance. For example, in tasks that require responding to some events (targets) while avoiding others (non-targets), target hits typically decrease and reaction time typically increases as operators spend more time on the task. Researchers often rely on target detection tasks known as continuous performance tests (CPTs) to investigate the source of these decrements in highly controlled laboratory settings. However, these powerful tools are seldom used to measure performance during demanding real-world multitasking scenarios. In this article, we report the results from two experiments in which participants simultaneously performed a driving-based target tracking task with a go-no-go target detection CPT task for approximately 12 min. Results revealed changes in the onset of vigilance decrements across measures, especially when CPT target presentation rates were faster than when they were slower. Together, these findings support the cognitive overload, cognitive underload, and opportunity-cost models of vigilance decrements and demonstrate that executive function, arousal, and motivation are important psychological constructs involved in sustained attention and multitasking performance. Insights from this work can inform the design and development of complex operator–system interfaces and thus increase safety and effectiveness for operators during mission-critical situations.

Introduction

Sustained attention, generally defined as the ability to continuously engage in relevant activities (Esterman & Rothlein, 2019; Fortenbaugh et al., 2017; Sarter et al., 2001), is a fundamental aspect of many tasks that require operators to detect and respond to critical signals over long periods (e.g., military surveillance, drone operation, long-distance truck driving) (Cardoso et al., 2019; Dorrian et al., 2007; Ghylin et al., 2007; Hancock & Hart, 2002; Heikoop et al., 2017; Künzell et al., 2018; Masoudian & Razavi, 2019; Reinerman-Jones et al., 2016). While performing such tasks, operators must exert varying levels of control for meeting task requirements while also maintaining adequate arousal in order to handle other, unexpected events (Ballard, 1996; Fernandez-Duque & Posner, 2001; Hancock, 1989; Klösch et al., 2022; Langner & Eickhoff, 2013; van Schie et al., 2021). This has been shown to be difficult, as vigilance decrements typically occur in which several measures of performance, such as target hit rate (hits) and reaction time (RT), deteriorate as time-on-task increases (Dillard et al., 2019; Mackworth, 1948; Parasuraman et al., 1987).

Much of the research on vigilance decrements has relied on sustained attention paradigms to investigate their underlying mechanisms under highly controlled experimental settings (for review, Riccio et al., 2002). These tools allow researchers to systematically explore how specific task features can influence operator performance during demanding contexts (Fisk & Scneider, 1981; Parasuraman & Mouloua, 1987; Roca et al., 2011), and how systematically modulating arousal can shape these effects (Buckley et al., 2016; Conners, 2014; He & McCarley, 2011; Luna et al., 2021a, 2021b, 2022a, 2022b; McBride et al., 2007). Understanding vigilance decrements is particularly important for real-life scenarios in which operators perform multiple simultaneous tasks as a function of their job requirements (Chen & Joyner, 2009; Chérif et al., 2018; Karpinsky et al., 2018; St. John et al., 1997; Strayer et al., 2003, 2011, 2019; Tillman et al., 2017). However, few studies have used sustained attention paradigms to closely examine multitasking performance in settings that simulate the complexity of real-world operations. To address this gap, the present study integrates a sustained attention task that yields multiple performance measures with a continuous driving-based tracking task to examine vigilance decrements in a controlled dual-task scenario that mirrors real-world demands.

Sustained attention and vigilance

Sustained attention is supported by the dynamic coordination of the multiple cognitive systems involved in goal-directed information processing (Fortenbaugh et al., 2017; Mirsky et al., 1991; Posner & Petersen, 1990; Sarter et al., 2001). Key among these are selective attention, which involves focusing cognitive resources on a particular stimulus or task goal (Johnston & Dark, 1986; Kahneman, 1973; Navon & Gopher, 1979; Treisman, 1969), and divided attention, which requires managing resources across multiple stimuli or task goals (Allport et al., 1972; Black et al., 2022; Janczyk & Kunde, 2020; Koch et al., 2018; Poljac et al., 2018; Wickens, 2008). These systems are further supported by higher-order mechanisms, most notably working memory and executive function, which store, maintain, and coordinate the processing of task-relevant information (Allport et al., 1972; Baddeley & Hitch, 1994; Black et al., 2022; Lara et al., 2014; Poole & Kane, 2009; Kieras et al., 2000; Kieras & Meyer, 1997; Koch et al., 2018; Poljac et al., 2018; Rubinstein et al., 2001; Sternberg, 1969; Wickens, 2002).

Although sustained attention is often treated as a distinct construct, it nonetheless overlaps conceptually and functionally with the notion of vigilance (Head, 1923; Klösch et al., 2022; MacLean et al., 2009; Oken et al., 2006; van Schie et al., 2021). Vigilance has been described as a state of readiness to detect subtle or infrequent changes in the environment (Mackworth, 1968) and involves physiological components such as alertness, defined as a sensitivity to incoming stimuli, and arousal, defined as the overall degree of alertness across states like sleep and wakefulness (Posner, 2008). In comparison with sustained attention, which reflects focused, goal-directed readiness for specific stimuli or tasks, vigilance encompasses a broader, more diffuse readiness to detect any relevant environmental change (van Schie et al., 2021). In this sense, maintaining vigilance may thus be necessary for sustained attention, but the opposite is not true. While these distinctions are conceptually meaningful, the current work does not aim to disentangle them. Rather, we focus on how performance changes over time—particularly in relation to performance declines observed in real-world settings.

Vigilance decrements

Research consistently shows that performance on sustained attention tasks declines over time (e.g., Dember et al., 1992; Dillard et al., 2019; Ditchburn, 1943; Scerbo, 1998; Teichner, 1974; Warm et al., 2018; Wyatt & Langdon, 1932). These performance declines, commonly referred to as ‘vigilance decrements,’ were first systematically studied by Norman Mackworth during World War II, when he observed declining detection accuracy in radar operators over prolonged periods of monitoring (Baca, 2017; Lichstein et al., 2000; Mackworth, 1948). Building on Mackworth’s foundational work, subsequent research investigated the task-specific factors that influence the rate and severity of these performance declines (Lanzetta et al., 1987; Parasuraman, 1979). Notably, Parasuraman and Davies (1977) developed a ‘vigilance taxonomy’ that links task characteristics to performance outcomes and posits that the likelihood, onset speed, and magnitude of vigilance decrements are shaped by task demands (Hancock et al., 2016; Matthews & Davies, 1988; Robinson & Brewer, 2023; See et al., 1995; Unsworth & Robison, 2020).

These insights have given rise to several theoretical accounts that attempt to explain the underlying mechanisms of vigilance decrements. These accounts are commonly classified into two main types. Resource depletion (or cognitive overload) theories argue that sustained performance tasks are stressful and demanding and that decrements are related to progressive declines in working memory capacity as time-on-task increases (Caggiano & Parasuraman, 2004; Fisk & Scerbo, 1987; Gartenberg et al., 2018; Helton & Warm, 2008; Matthews et al., 1993; Wiener et al., 1984). In contrast, cognitive underload theories emphasize that sustained attention tasks are monotonous and boring (Cummings et al., 2016; Greenlee et al., 2018; McBain, 1970; Scerbo et al., 1992) and that decrements are related to progressive declines in arousal leading to withdrawal of attention from the primary task. A third perspective, the opportunity-cost account (Kurzban et al., 2013), proposes that vigilance decrements occur when individuals reallocate cognitive resources based on cost–benefit tradeoffs, adjusting their response strategies to the perceived utility of continuing to sustain attention (Gyles et al., 2023; Thomson et al., 2016).

Together, these theories underscore the multifaceted nature of vigilance decrements, indicating that performance declines may stem from a complex interplay of cognitive demands, arousal fluctuations, and motivational considerations. While these accounts offer distinct explanatory frameworks, they are not mutually exclusive; rather, multiple mechanisms may operate in parallel within a given task context. This overlap often makes it difficult to isolate specific contributing factors in real-world scenarios, where different mechanisms can produce similar behavioral outcomes. Task context remains critical for interpretation, particularly in distinguishing between cognitive overload and underload, whereas opportunity-cost effects are often most apparent in dual-task settings in which performance deteriorates unevenly across tasks as attention is reallocated based on perceived utility.

Challenges of measuring sustained attention

Although theoretical models offer valuable explanations for vigilance decrements, accurately measuring sustained attention remains a major challenge. To address this, researchers often rely on continuous performance tests (CPTs), which follow a blocked successive discrimination task paradigm in which participants are required to either respond, or withhold response, when detecting target stimuli among distractors (Ord et al., 2021; Riccio et al., 2002; Robertson et al., 1997; Rosvold et al., 1956; Smid et al., 2006). CPTs yield multiple performance metrics, including reaction time (RT), omission and commission errors, and detection accuracy (e.g., hits and correct rejections or CRs) (Borgaro et al., 2003; Edwards et al., 2007; Riccio et al., 2001; Roebuck et al., 2016). CPTs also support signal detection theory (SDT) analyses, including sensitivity (d′), which reflects the ability to discriminate between targets among non-targets, and response bias (β), which reflects a preference for speed (liberal bias) or accuracy (conservative bias) when responding to targets (Azizi et al., 2018; Green & Swets, 1966).

To extract meaningful insights from CPT data, these performance metrics are often averaged across the entire task to assess overall attention and analyzed across successive trial blocks to capture temporal dynamics in sustained attention (Bubnik et al., 2015; Cornblatt et al., 1989; Esterman & Rothlein, 2019). However, even averaging performance within a block can obscure important within-block fluctuations (discussed later), especially when target frequency or task pacing change during the block. Moreover, even knowing that accuracy changes within a block does not reveal why it changes, or which mechanisms (e.g., cognitive overload or waning arousal) drive the change.

To address limitations of block-wise averaging, the Conners’ Continuous Performance Test (CCPT; Conners, 2014) alternates short and long interstimulus intervals (ISIs) within blocks throughout the task. This design allows for disentangling different reasons for performance declines, as shorter ISIs likely impose higher demands on executive function and working memory, whereas longer ISIs likely impose higher demands on maintaining alertness over time (Conners et al., 2003; Egeland & Kovalik-Gran, 2010). By comparing performance declines between short and long ISIs over time (i.e., across successive blocks), researchers can pinpoint the likely reasons for these declines.

Beyond task design, additional challenges arise in how sustained attention is statistically modeled. Many studies rely on linear analysis across discrete time points (e.g., ANOVA, regression), which, as briefly mentioned earlier, can obscure important non-linear changes in performance trends (Winter & Wieling, 2016). These limitations are especially critical in applied settings where extended task exposure may lead to learning-based practice effects that can mask, delay, or even counteract the occurrence of vigilance decrements during the early stages of task engagement, particularly in tasks demanding higher-order executive function (Brown, 2008; Fisk & Schneider, 1981; Norman & Shallice, 1986).

Figure 1 illustrates how practice and vigilance decrement effects can shape performance trajectories. In minimally demanding tasks, performance may remain stable across blocks (Fig. 1a). In contrast, demanding multitasking scenarios can produce linear trends marked by either progressive declines due to vigilance decrements (Fig. 1b—solid line) or improvements due to practice effects (Fig. 1b—dashed line). Performance may also be characterized by curvilinear patterns, which progressively improve in the first few blocks due to early practice effects and then either degrade toward the end of the session due to late-onset vigilance decrements (Fig. 1c—solid line) or flatten due to task attenuation (Fig. 1c—dashed line). Alternatively, demanding multitasking scenarios can lead to opposite patterns in which performance progressively degrades in the first few blocks due to overloading executive resources by having to repeatedly task-switch. Performance can then either improve toward the last few blocks of the session due to late-onset practice effects (Fig. 1d—solid line), or flatten due to task attenuation (Fig. 1d—dashed line).

Fig. 1 — Schematic graph of possible performance trajectories. a Flat performance; b Linear performance indicating either progressive vigilance decrements (solid line) or practice effects (dashed line); c, d Curvilinear performance indicating (c) early practice effects and then later vigilance decrements (solid line) or flat performance (dashed line), or (d) early vigilance decrements and later practice effects (solid line), or flat performance (dashed line). Note that ‘better’ performance according to a hypothetical measure is illustrated by higher (more positive) values. In reality, in different CPT measures, better performance can be indicated by either higher (e.g., target hits) or lower (e.g., RTs and Omissions) values

Given the theoretical accounts of vigilance decrements that we discussed, we can anticipate performance patterns aligned with those illustrated in Fig. 1. As noted earlier, disentangling overload from underload is notoriously difficult because both can produce performance declines over time. However, overload is more likely to emerge during short ISIs, whereas underload is more likely to emerge when stimulation is lower during long ISI blocks. Specifically, cognitive overload models predict steady linear declines (Fig. 1b—solid line) or late curvilinear declines (Fig. 1c—solid) when task demands are high (e.g., for short ISIs). In contrast, cognitive underload accounts predict the same patterns when task demands are low (e.g., for long ISIs). Finally, opportunity-cost accounts predict divergent trends such that improvements in one task may be coupled with declines in the other, or shifts in response bias as attention is redeployed to maximize utility.

In summary, sustained attention paradigms such as CPTs have been instrumental in advancing our understanding of vigilance decrements under controlled conditions. These tools have supported key theoretical accounts, including the cognitive overload and underload models. However, most prior work has focused on single-task environments and has not examined how these mechanisms unfold in more ecologically valid, multitasking scenarios that require continuous and concurrent engagement. Moreover, little attention has been given to how nonlinear performance trajectories may reveal different sources of decline. As a result, traditional-CPT findings may not fully generalize to real-world operational settings that demand dynamic coordination across tasks over extended periods (Chaytor & Schmitter-Edgecombe, 2003; Deniaud et al., 2015; Faria et al., 2023; Halperin et al., 1991; Sayer, 2005; Stojmenova & Sodnik, 2018; Thorpe et al., 2019).

The present study

In the present study, we aimed to: (1) develop a clear method for systematically assessing performance changes across time in two simultaneous tasks that resemble many real-world scenarios and (2) evaluate whether the observed patterns are more consistent with cognitive overload, cognitive underload, opportunity costs, or a combination of these. To achieve these aims, we conducted two experiments which implemented a novel dual-task paradigm that we developed that integrated a go-no-go target detection task (Conners, 2014) with a smooth pursuit tracking task adapted from the continuous tracking and reaction (ConTRe; Mahr et al., 2012) paradigm in the OpenDS driving simulator environment (Math et al., 2013). We also analyzed time-related changes in performance using growth curve modeling (Baayen et al., 2008; Bates et al., 2014).

This dual-task framework offers a realistic testbed for understanding how sustained attention is managed in settings that require ongoing monitoring and rapid decision making, conditions that closely reflect many real-world scenarios (Abich IV et al., 2017; Chérif et al., 2018; Lee & Taatgen, 2019).

Experiment 1

Our first question for this study was whether and how performance would change over time for each of the measures and indices associated with the CPT and tracking tasks. Previous studies have demonstrated that performance trajectories may vary in both shape and direction over time due to different factors, such as learning-based practice effects (Alexander et al., 2003; Beglinger et al., 2005) and vigilance decrements (Haubert et al., 2018; Lee et al., 2010; Pattyn et al., 2008). However, no previous study has specifically examined how these temporal trajectories unfold in a combined task paradigm, where a discrete CPT is performed alongside a continuous tracking task in an arrangement that introduces competing cognitive demands and may influence how vigilance-related changes manifest and are measured over time.

Our second question for this study was whether the temporal trajectories of any of our measures or indices mentioned above would differ when target presentation rates are faster compared to when they are slower. The Conners CPT (Conners et al., 2014) represents a well-established method in vigilance research, wherein interstimulus intervals (ISIs) are varied within task blocks to modulate attentional demands (Basner & Dinges, 2011; Conners et al., 2003; Egeland & Kovalik-Gran, 2010; MacLean et al., 2009). This design is based on the premise that longer ISIs place greater demands on sustained readiness to respond, thereby increasing the likelihood of detecting attentional lapses and enabling a more fine-grained assessment of performance over time. Despite its widespread use, this approach has not been applied to examine time-based performance patterns in dual-task settings where a CPT is paired with a continuous tracking task.

Methods

Participants

Following the approval of the University of South Carolina IRB board, a total of 63 native English-speaking participants (age: M = 20.19, SD = 1.48) from the University of South Carolina Department of Psychology undergraduate participant pool took part in the experiment. Participants were compensated with extra credit for their time. We conducted a power analysis prior to testing in R (R Core Team, 2012) using the wp.rmanova function from the WebPower library (Zhang et al., 2018) for a medium effect-size (η²) =.50, alpha (α) =.05, and power (β) =.80. The analysis indicated a sample size of 58 participants, which we exceeded by 5 participants to allow for potential data loss. Of the 63 participants, 45 were female (age: M = 20.2, SD = 1.32) and 18 were male (age: M = 20.2, SD = 1.86). Recruitment criteria for this study specified that participants had to be native speakers of English and must have also held a valid driver’s license. There were no other inclusion or exclusion criteria for selecting participants.

Design

The design of this study was influenced by previous studies that examined task performance using combined multitasking paradigms (Buckley et al., 2016; Castro et al., 2019; Luna et al., 2022a, 2022b; Rann & Almor, 2022). Unlike studies that compared single- and dual-task performance using counterbalanced conditions (e.g., Chiew & Brazer, 2013; Moran et al., 2020), our primary aim was to examine within-session changes during dual-tasking, where the need for switching between the tasks may either exacerbate or mitigate vigilance decrements. To support this focus, we utilized a fixed-order design in which participants completed a single-task CPT session first, followed by a dual-task session. This approach ensured that participants were well-practiced in the demanding CPT so as to increase power for detecting vigilance decrements in both tasks during the second, dual-task, session (Carter et al., 2013; Cornblatt et al., 1989; Hope et al., 1998; Lemay et al., 2004).

Accordingly, we report the results for the two sessions separately. While we present the analyses from both sessions, we caution that we cannot derive any strong conclusion from performance differences between the sessions as the source of these differences can be either the order of the tasks, or the presence vs. absence of the secondary tracking task. Nevertheless, we note that measures that show similar result patterns in both sessions are likely not sensitive to either order or dual- vs. single-task differences.

Tasks

The primary task was a go-no-go target detection task that was based on the Conner’s (2014) CPT and was coded and implemented using E-prime 3.0 stimulus presentation software (Psychology Software Tools, Pittsburgh, PA). It required participants to continuously monitor a computer screen as black letter stimuli (font: Arial; size: 200px) were presented against a white background, and to respond by pressing a foot pedal when they detected target letters (‘A’–‘W’, ‘Y’, ‘Z’) and withhold response when they detected the critical non-target letter ‘X’ (Fig. 2). The foot pedal was a Linemaster T-91-S Treadlite II Foot Switch attached to a Psychology Software Tools 200 A Serial Response Box that interfaced with the computer that ran the CPT.

Fig. 2 — The CPT task. Participants are required to press the foot pedal when they detect a target letter and withhold response when a non-target letter is detected. The blank screen between stimuli lasted for either 1000 ms (short ISI), 2000 ms (middle ISI), or 4000 ms (long ISI)

Each stimulus presentation (whether target or non-target), as well as the subsequent time window including the ISI and up until the next stimulus presentation, was considered a trial. Each session consisted of 270 trials presented over the course of approximately 12 min with trials appearing at the predetermined rate regardless of whether the participant responded. Each session was divided into six blocks, each including 45 trials. Each block consisted of three sub-block conditions, each with a different ISI (1000 ms, 2000 ms, or 4000 ms). The order of the three sub-blocks was randomly determined for each block for each participant. Target and non-target stimuli were randomly selected with an 80–20% ratio of targets to non-targets, each presented for 250 ms. Performance data for this task was sampled per trial and stored in E-prime output data files. The primary CPT task was administered as the only task in the first session and then together with the secondary driving task in the second session.

The secondary driving-based tracking task was the continuous tracking and reaction (ConTRe) task (Mahr et al., 2012) implemented in the OpenDS driving simulator (Math et al., 2013). It required participants to continuously track a yellow target moving across the simulator screen with a blue cursor that they controlled using a steering wheel (a Microsoft SideWinder Precision Racing Wheel) which interfaced with the computer that ran the simulator (Fig. 3).

Fig. 3 — ConTRe task. Participants are required to track the moving yellow target using the blue cursor they control using a steering wheel

The yellow target was placed approximately 20 simulated ft in front of the participants’ view on the simulator screen and moved horizontally (i.e., left-to-right, right-to-left) across the screen at a constant lateral speed of 1 simulated meter per second. The yellow target’s direction of movement (left-to-right, right-to-left) changed at random times. Participants only had control of the lateral movement of the blue cursor. Performance data for the tracking task was sampled approximately every 19 ms and stored in a MySQL database.

To allow for accurate temporal synchronization of the tracking and the CPT data, OpenDS was programmed to emit a sound that started the e-Prime CPT task in the second experiment session through a voice key.

Apparatus

The CPT ran on a Dell Precision Tower 7810 computer and was displayed on a Dell 17-inch monitor at a 1280 × 1040px resolution and a refresh rate of 60 Hz. The driving-based tracking task ran on a Dell XP 435t/9000 computer and was displayed on a Dell 27″ full HD 1920 × 1080 flat-panel monitor with a refresh rate of 60 Hz. This tracking task computer was used to collect data for approximately half the participants in E1 and was replaced by a Dell OptiPlex 790 computer after malfunctioning. Both computers ran the Windows 10 Pro operating system.

The monitors for both tasks were placed approximately two feet directly in front of the participants. The monitor for the CPT task was placed behind and beneath the larger monitor that displayed the driving-based tracking task. For the CPT, the horizontal viewing angle was approximately 31° and the vertical visual angle was approximately 25°; for the tracking task, the horizontal viewing angle was approximately 52° and the vertical visual angle was approximately 40°. During the first experiment session (i.e., Session 1), the larger monitor was turned off so that participants only attended to the CPT displayed on the smaller monitor (Fig. 4a). During the second experiment session (i.e., Session 2), both tasks were simultaneously presented thus requiring participants to split their attention between both monitors as they performed the tasks (Fig. 4b).

Fig. 4 — Setup—a Session 1, in which the larger top monitor was turned off and the CPT was displayed on the lower monitor; b Session 2, in which both monitors displayed their associated tasks. The red cross was not visible to participants during the experiment and is included here as to illustrate participants’ approximate line of sight

The steering wheel used for the tracking task was mounted to the edge of the desk in front of participants. The foot pedal was placed directly in front of participants on the floor at the approximate position their right foot would naturally rest. Participants’ line of sight was located approximately in front of the bottom of the monitor displaying the tracking task, at a point where the CPT and tracking task stimuli were approximately equidistant (red cross, Fig. 4b).

Procedure

Before the experiment, participants confirmed valid driver’s licenses and silenced their phones to prevent distractions. After providing informed consent, participants were seated in a testing room and instructed on how to perform the CPT task. They then completed a 120-s practice session to familiarize themselves with the task, including exposure to all three ISI conditions. The first experimental session followed, with the researcher monitoring compliance from outside the room via auditory cues.

Following Session 1, participants received instructions for the dual-task condition and were told to give equal priority to the CPT and tracking tasks. A second 120-s practice session was conducted to acclimate them to the combined task environment. Participants then completed the second experimental session, again with the researcher monitoring task engagement through sound cues from both the pedal and steering wheel.

Measures

The independent measures used for analysis included experiment block (1, 2, 3, 4, 5, 6) and ISI condition (1000 ms, 2000 ms, 4000 ms). We collected and/or calculated eight dependent measures in total. From the CPT task we collected five measures: target hit rate (hits), omissions, commissions, correct rejections (CRs), and reaction times (RTs). However, in this CPT variant, omissions and commissions are codependent with hits and CRs, respectively, such that the performance in one measure (e.g., hits) over time is oppositely reflected in the performance of its complementary measure (e.g., omissions). Given this redundancy, we do not analyze omissions here and only utilize commissions for calculating the sensitivity (d’) and response bias (β) SDT measures. From the driving-based tracking task we collected tracking distance. To relate tracking performance more directly to cognitive demands, we segmented tracking data around critical CPT events in which participants were presented with visual stimuli (targets and non-targets). This allowed us to directly relate fluctuations in tracking performance to concurrent cognitive demands imposed by the CPT.

Table 1 describes these measures as they are used in this experiment.

Table 1.

Description of measures used in experiment 1

Task	Measure/index	Definition
CPT	Hits (Number of target hits)	Correctly pressing foot pedal when a target is presented
	Omissions (Number of target misses)	Failing to press the foot pedal when a target is presented
	Commissions (Number of false alarms)	Pressing the foot pedal when a non-target is presented
	CRs (Number of Correct Rejections)	Correctly not pressing the foot pedal when a non-target is presented
	RTs (Speed of response to targets)	The time between the onset of a target and the correct pressing of the foot pedal
	d’ (Sensitivity)	SDT-based sensitivity (d′) for accurately discriminating between targets and non-targets
	β (Response bias)	SDT-based response bias (β) regarding speed-vs-accuracy tradeoffs
Tracking	Distance (Average distance from target)	Average distance between target and cursor

Open in a new tab

Note that we do not analyze omissions in this study at all (indicated in gray shading in the table) and only utilize commissions to calculate SDT measures, since both measures are fully codependent with hits and CRs, respectively

Data preparation

The CPT data from the practice sessions were discarded, and the CPT data from the first and second experiment sessions were retained for analysis. The data from the driving-based tracking task in the second experiment session, which was stored in a MySQL table, were exported to.csv format for further data processing and analysis in R 4.2.2 (R Core Team, 2012) and R Studio version 1.1.447 (RStudio Team, 2016), which were used for all the analyses reported in this paper.

Prior to statistical analysis, the data files for both the CPT and driving-based tracking tasks were temporally aligned and merged so that tracking performance could be accurately matched with the CPT data. Matching was based on the audio signal (a beep) that OpenDS was programmed to emit to start the CPT task on e-prime through a voice key. Following the alignment, tracking data were filtered to retain only the data from the 500 ms before and after each CPT stimulus presentation. This was done to create matched fixed time windows of tracking data across conditions since the conditions with 4000 ms ISI had twice as much data as the 2000 ms ISI conditions and four times as much data as the 1000 ms ISI conditions (each condition included the same number of CPT targets) (Ballard, 2001). For more fine-grained analysis, the tracking data were further segmented based on whether the corresponding CPT events were targets responded to correctly (hits), non-targets correctly not responded to (CRs), or targets not responded to (commissions), allowing us to assess how tracking performance varied across specific cognitive demands. Because of the low number of omissions, we did not calculate tracking distance averages for trials with these events. Finally, trials with RTs less than 100 ms were removed since they did not likely reflect valid responses (Conners, 2014).

Data analysis

We utilized growth curve analyses (GCAs) to analyze the data in our experiments. GCAs are specialized statistical techniques for summarizing longitudinal data with best-fit lines (or smooth curves) that characterize performance trends within observed time windows (Bollen, 2007; Byrne & Crombie, 2003; Kristjansson et al., 2007). This framework is based on building and comparing mixed-effects models that include, in addition to the usual fixed and random effects, fixed terms representing temporal changes of different orders (Baayen et al., 2008; Barr, 2013; Peugh, 2010).

For our GCA, we used the lme4 (Bates et al., 2014) statistical package. The GCAs consisted of terms for linear (i.e., time¹) and quadratic (i.e., time²) time orders, terms for ISI (1000 ms, 2000 ms, 4000 ms), and a random effect of participants on the intercept. All factors included in the models used sum contrast coding (Schad et al., 2020). We attempted to fit more complex random factor terms to the data, but these models did not converge.

For model comparison, we compared four models that increased in complexity:

The Base model only included baseline time terms and the random factors (with no fixed terms representing ISI or its interaction with any of the time terms) (Model 1, Table A). Note that the Base model included Intercept, Linear, and Quadratic terms/predictors but not any of the fixed factors we manipulated. This means that the Base model could capture the temporal characteristics of the data that were constant across ISI conditions.
The Intercept model built upon the base model by adding an interaction of ISI with the intercept (Model 2, Table A).
The Linear model built upon the intercept model by adding the interaction of ISI with the linear (i.e., time¹) time order (Model 3, Table A).
The Quadratic model built upon the linear model by adding the interaction of ISI with the quadratic (i.e., time²) time order (Model 4, Table A).

We then compared each of these models, using maximum likelihood estimates (Long, 2012) to determine the best time order model to use for the analysis. Following Long’s recommendations, we considered a model to provide a better fit than a simpler model using a p <.1 criterion.

In our analysis, we used sum coding, meaning that our factor estimates represent the difference of each level from the grand mean rather than from the third level (Brehm & Alday, 2022). This allows for a more transparent interpretation of model coefficients.

Following Long (2012), we encourage readers to consider the GCA results not only by whether a given measure changes over time, but also by the shape and direction of those changes (e.g., linear vs. curvilinear trends) and how they varied by ISI. We follow these interpretive strategies in the Results and Discussion sections of this and the following experiment, and in the General Discussion, where we synthesize the theoretical implications of the different result patterns across measures, ISIs, and experiments.

Results

Data from seven participants were removed due to hardware malfunction. Data from the remaining 56 participants (age: M = 20.14, SD = 1.51) were used for analysis. While this sample size is two participants short of the 58 suggested by our power analysis, this difference is negligible given the overall large number of participants. Of these 56 participants, 39 were female (age: M = 20.1, SD = 1.33) and 17 were male (age: M = 20.2, SD = 1.89).

Session 1 CPT measures

Target Hits: None of the models provided a better fit of the data than the Base model according to our criterion. We therefore chose the Base model (Table 2a). Inspection of the model coefficients (Table 3a) and visual inspection of the graph (Fig. 5a) show: 1. a significant positive linear trajectory (i.e., improving performance) across all ISI conditions that tapered off in the last two blocks and 2. no significant differences in performance between ISIs.

Table 2.

Model comparison tables for CPT measures in E1 session 1—(a) hits, (b) CRs, and (c) RTs

Model	nPar	AIC	BIC	Loglik	Deviance	Χ²	df	p
(a) Model comparison table for hits
Base	5	6333.4	6358	− 3161.7	6323.4
Intercept	7	6335.9	6370.3	− 3160.9	6321.9	1.5444	2	0.462
Linear	9	6339.5	6383.8	− 3160.8	6321.5	0.3679	2	0.832
Quadratic	11	6342.6	6396.7	− 3160.3	6320.6	0.9403	2	0.6249
(b) Model comparison table for CRs
Base	5	9151.3	9175.9	− 4570.6	9141.3
Intercept	7	9150.2	9184.6	− 4568.1	9136.2	5.0606	2	7.96E− 02
Linear	9	9151.2	9195.5	− 4566.6	9133.2	2.9864	2	0.22465
Quadratic	11	9153.7	9207.8	− 4565.9	9131.7	1.499	2	0.4726
(c) Model comparison table for RTs
Base	5	− 1641.6	− 1617	825.8	− 1651.6
Intercept	7	− 1941.5	− 1907.1	977.8	− 1955.5	303.93	2	< 2e− 16***
Linear	9	− 1943.6	− 1899.4	980.8	− 1961.6	6.1257	2	0.04675*
Quadratic	11	− 1939.8	− 1885.7	980.9	− 1961.8	0.182	2	0.913

Open in a new tab

* indicates p < .05, ** indicates p < .01, *** indicates p < .001

Table 3.

Chosen model coefficients tables for CPT measures in E1 session 1—(a) hits, (b) CRs, and (c) RTs

Fixed Effects	Estimate	Std. Error	df	t value	p
(a) Base model coefficients for hits
(Intercept)	98.7021	0.5492	56	179.725	< 2e−16 ***
time¹	1.0731	0.402	952	2.67	0.00772 **
time²	− 0.4492	0.402	952	− 1.118	0.26404
(b) Intercept model coefficients for CRs
(Intercept)	78.4392	2.05	56	38.264	< 2e−16***
time¹	− 0.1423	1.6297	952	− 0.087	0.93
time²	− 1.6669	1.6297	952	− 1.023	3.07E−01
ISI1	− 1.1574	0.9409	952	− 1.23	0.219
ISI2	− 0.959	0.9409	952	− 1.019	0.308
(c) Linear model coefficients for RTs
(Intercept)	5.9600	0.0166	56.0000	358.2630	< 2e−16***
time¹	0.0271	0.0064	952.0000	4.2540	0.0000***
time²	− 0.0270	0.0064	952.0000	− 4.2500	0.0000***
ISI1	− 0.0613	0.0037	952.0000	− 16.6870	< 2e−16***
ISI2	0.0019	0.0037	952.0000	0.5050	0.6136
time¹:ISI1	− 0.0155	0.0090	952.0000	− 1.7170	0.0863
time¹:ISI2	− 0.0062	0.0090	952.0000	− 0.6900	0.4906

Open in a new tab

* indicates p < .05, ** indicates p < .01, *** indicates p < .001

Fig. 5 — Growth curve analysis graphs for CPT measures in E1 Session 1—a hits, b CRs, and c RTs. Dots represent observed mean performance in each block and ISI combination, and lines represent the chosen model’s predictions

CRs: The Intercept model provided the best fit of the data according to our criterion (Table 2b), χ²(2) = 5.06, p <.08. Inspection of the model coefficients (Table 3b) and visual inspection of the graph (Fig. 5b) show: 1. no overall temporal trends in the data and 2. no significant differences between ISI conditions.

RTs: The linear model provided the best fit of the data according to our criterion (Table 2c), χ²(2) = 6.13, p <.05. Inspection of the model coefficients (Table 3c) and visual inspection of the graph (Fig. 5c) show: 1. an overall positive linear trajectory (i.e., degrading performance) during the 4000 ms ISI; 2. an overall negative quadratic trajectory (i.e., declining then tapering off or slightly improving performance) during the 1000 ms and 2000 ms ISIs; and 3. overall fastest (i.e., best) performance during the 1000 ms ISI and slowest (i.e., worst) performance during the 4000 ms ISI.

Session 1 SDT indices

Sensitivity: The Intercept model provided the best fit of the data according to our criterion (Table 4a), χ²(2) = 6.03, p <.05. Inspection of the model coefficients (Table 5a) and visual inspection of the graph (Fig. 6a) show: 1. no overall temporal trends in the data and 2. marginally worse sensitivity during the 1000 ms and 2000 ms ISI conditions than in the 4000 ms condition.

Table 4.

Model comparison tables for SDT indices in E1 Session 1—(a) sensitivity and (b) response bias

Model	nPar	AIC	BIC	Loglik	Deviance	Χ²	df	p
(a) Model comparison table for sensitivity
Base	5	1705.4	1730	− 847.7	1695.4
Intercept	7	1703.4	1737.8	− 844.69	1689.4	6.0342	2	4.89E−02*
Linear	9	1704.7	1749	− 843.37	1686.7	2.6407	2	0.26705
Quadratic	11	1706.2	1760.3	− 842.11	1684.2	2.5172	2	0.28405
(b) Model comparison table for response bias
Base	5	− 775.57	− 750.99	392.79	− 785.57
Intercept	7	− 773.05	− 738.64	393.53	− 787.05	1.4819	2	0.4767
Linear	9	− 769.43	− 725.19	393.71	− 787.43	0.375	2	0.829
Quadratic	11	− 766.07	− 712	394.04	− 788.07	0.6456	2	0.7241

Open in a new tab

* indicates p < .05, ** indicates p < .01, *** indicates p < .001

Table 5.

Chosen model coefficients tables for SDT indices in E1 session 1—(a) sensitivity and (b) response bias

Fixed effects	Estimate	Std. Error	df	t value	p
(a) Intercept model for sensitivity
(Intercept)	2.34348	0.05242	56	44.706	< 2e−16***
time¹	0.03499	0.04047	952	0.864	0.3876
time²	− 0.05139	0.04047	952	− 1.27	0.2045
ISI1	− 0.03943	0.02337	952	− 1.687	0.0919
ISI2	− 0.01652	0.02337	952	− 0.707	0.4797
(b) Base model for response bias
(Intercept)	− 0.4814	0.0112	56	− 43.0330	< 2e−16***
Time¹	0.0075	0.0121	952	0.6210	0.5350
Time²	− 0.0119	0.0121	952	− 0.9890	0.3230

Open in a new tab

* indicates p < .05, ** indicates p < .01, *** indicates p < .001

Fig. 6 — Growth curve analysis graphs for SDT indices in E1 session 1—a sensitivity and b response bias. Dots represent observed mean performance in each block and ISI combination, and lines represent the chosen model’s predictions

Response Bias: None of the models provided a better fit of the data than the Base model according to our criterion (Table 4b). Inspection of the model coefficients (Table 5b) and visual inspection of the graph (Fig. 6b) show: 1. no overall temporal trends in the data; 2. no significant differences in response bias between ISI conditions; and 3. response bias was indistinguishable in all ISIs and was overall liberal.

Session 2 CPT measures

Target Hits: The Intercept model provided the best fit of the data according to our criterion (Table 6a), χ²(2) = 9.74, p <.01. Inspection of the model coefficients (Table 7a) and visual inspection of the graph (Fig. 7a) show: 1. no overall temporal trends in the data and 2. highest (i.e., best) performance during the 4000 ms ISI condition followed by the 2000 ms ISI condition.

Table 6.

Model comparison tables for CPT measures in E1 session 2—(a) hits, (b) CRs, and (c) RTs

Model	nPar	AIC	BIC	Loglik	Deviance	Χ²	df	p
(a) Model comparison table for hits
Base	5	5816.1	5840.7	− 2903.1	5806.1
Intercept	7	5810.4	5844.8	− 2898.2	5796.4	9.7378	2	0.007682**
Linear	9	5813.9	5858.1	− 2897.9	5795.9	0.5217	2	0.770386
Quadratic	11	5817.3	5871.4	− 2897.7	5795.3	0.5115	2	0.774321
(b) Model comparison table for CRs
Base	5	9324.6	9349.2	− 4657.3	9314.6
Intercept	7	9328.2	9362.6	− 4657.1	9314.2	0.4054	2	8.17E−01
Linear	9	9329.3	9373.5	− 4655.6	9311.3	2.8876	2	0.23603
Quadratic	11	9325.9	9380	− 4652	9303.9	7.3421	2	0.02545*
(c) Model comparison table for RTs
Base	5	− 1422.6	− 1398	716.29	− 1432.6
Intercept	7	− 1692	− 1657.5	852.97	− 1706	273.36	2
Linear	9	− 1692.3	− 1648	855.15	− 1710.3	4.3476	2
Quadratic	11	− 1688.5	− 1634.4	855.25	− 1710.5	0.2049	2

Open in a new tab

* indicates p < .05, ** indicates p < .01, *** indicates p < .001

Table 7.

Chosen model coefficients tables for CPT measures in E1 session 2—(a) hits, (b) CRs, and (c) RTs

Fixed effects	Estimate	Std. error	df	t value	p
(a) Intercept model coefficients for hits
(Intercept)	98.40443	0.24992	56	393.751	< 2e−16***
time¹	0.19565	0.31921	952	0.613	0.5401
time²	− 0.20566	0.31921	952	− 0.644	0.5195
ISI1	− 0.46296	0.1843	952	− 2.512	0.0122*
ISI2	− 0.06614	0.1843	952	− 0.359	0.7198
(b) Quadratic model coefficients for CRs
(Intercept)	71.0979	2.4254	56	29.314	< 2e−16***
time¹	3.4624	1.7623	952	1.965	0.04974*
time²	− 4.7843	1.7623	952	− 2.715	0.00675**
ISI1	0.6283	1.0175	952	0.618	5.37E−01
ISI2	− 0.463	1.0175	952	− 0.455	0.6492
time¹:ISI1	3.4387	2.4923	952	1.38	0.16799
time¹:ISI2	− 3.8892	2.4923	952	− 1.561	0.11897
time²:ISI1	− 0.4763	2.4923	952	− 0.191	0.84849
time²:ISI2	6.0833	2.4923	952	2.441	0.01483*
(c) Intercept model coefficients for RTs
(Intercept)	6.0839	0.0189	60	322.224	< 2e−16***
time¹	0.0468	0.0072	952	6.4820	0.0000***
time²	− 0.0001	0.0072	952	− 0.0140	0.9885
ISI1	− 0.0704	0.0042	952	− 16.8770	< 2e−16***
ISI2	0.0148	0.0042	952	3.5550	0.0004***

Open in a new tab

* indicates p < .05, ** indicates p < .01, *** indicates p < .001

Fig. 7 — Growth curve analysis graphs for CPT measures in E1 Session 2—a hits, b CRs, and c RTs. Dots represent observed mean performance in each block and ISI combination, and lines represent the chosen model’s predictions

CRs: The quadratic model provided the best fit of the data according to our criterion (Table 6b), χ²(2) = 7.34, p <.03. Inspection of the model coefficients (Table 7b) and visual inspection of the graph (Fig. 7b) show: 1. an overall positive linear trajectory (i.e., improving performance); 2. a negative quadratic trajectory (i.e., improving then degrading performance) in the 1000 ms and 4000 ms IS conditions; and 3. a flat trajectory (i.e., constant performance) in the 2000 ms ISI.

RTs: The intercept model provided the best fit of the data according to our criterion (Table 6c), χ²(2) = 273.36, p <.001. Inspection of the model coefficients (Table 7c) and visual inspection of the graph (Fig. 7c) show: 1. positive linear trajectories (i.e., worsening performance) across all ISI conditions and 2. significantly faster (i.e., better) performance in the 1000 ms ISI condition than in the 2000 and 4000 ms ISI conditions.

Session 2 SDT indices

Sensitivity: The quadratic model provided the best fit of the data according to our criterion (Table 8a), χ²(2) = 7.89, p <.02. Inspection of the model coefficients (Table 9a) and visual inspection of the graph (Fig. 8a) show: 1. a negative quadratic trajectory (i.e., increasing then declining sensitivity) in the 1000 ms and the 4000 ms ISI conditions; and 2. a flat trajectory (constant sensitivity) in the 2000 ms ISI.

Table 8.

Model comparison tables for SDT indices in E1 session 2—(a) sensitivity and (b) response bias

Model	nPar	AIC	BIC	Loglik	Deviance	Χ²	df	p
(a) Model comparison table for sensitivity
Base	5	1877	1901.6	− 933.5	1867
Intercept	7	1880.5	1914.9	− 933.26	1866.5	0.4783	2	0.7873
Linear	9	1882	1926.2	− 931.98	1864	2.5564	2	0.27854
Quadratic	11	1878.1	1932.2	− 928.04	1856.1	7.8915	2	0.01934*
(b) Model comparison table for response bias
Base	5	− 632.1	− 607.52	321.05	− 642.1
Intercept	7	− 637.11	− 602.7	325.55	− 651.11	9.0079	2	0.01107*
Linear	9	− 634.82	− 590.58	326.41	− 652.82	1.7127	2	0.4247
Quadratic	11	− 631.57	− 577.5	326.79	− 653.57	0.7521	2	0.68658

Open in a new tab

* indicates p < .05, ** indicates p < .01, *** indicates p < .001

Table 9.

Chosen model coefficients tables for SDT indices in E1 session 1—(a) sensitivity and (b) response bias

Fixed effects	Estimate	Std. error	df	t value	p
(a) Quadratic model coefficients for sensitivity
(Intercept)	2.153712	0.060536	55.999996	35.577	< 2e−16***
time¹	0.07745	0.043806	952.000001	1.768	0.07738
time²	− 0.122764	0.043806	952.000001	− 2.802	0.00517**
ISI1	− 0.004099	0.025291	952.000001	− 0.162	0.87127
ISI2	− 0.012764	0.025291	952.000001	− 0.505	0.6139
time¹:ISI1	0.090254	0.061951	952.000001	1.457	0.14548
time¹:ISI2	− 0.081461	0.061951	952.000001	− 1.315	0.18885
time²:ISI1	− 0.023175	0.061951	952.000001	− 0.374	0.70842
time²:ISI2	0.161276	0.061951	952.000001	2.603	0.00938**
(b) Intercept model coefficients for response bias
(Intercept)	− 0.484496	0.008131	56	− 59.587	< 2e−16***
time¹	0.0088	0.0132	952.000001	0.6670	0.5050
time²	0.0085	0.0132	952.000001	0.6420	0.5210
ISI1	0.0186	0.0076	952.000001	2.4360	0.0150
ISI2	0.0024	0.0076	952.000001	0.3110	0.7560

Open in a new tab

* indicates p < .05, ** indicates p < .01, *** indicates p < .001

Fig. 8 — Growth curve analysis graphs for SDT indices in E1 Session 2—a sensitivity and b response bias. Dots represent observed mean performance in each block and ISI combination, and lines represent the chosen model’s predictions

Response Bias: The intercept model provided the best fit of the data according to our criterion (Table 8b), χ²(2) = 9.01, p <.02. Inspection of the model coefficients (Table 9b) and visual inspection of the graph (Fig. 8b) show: 1. a flat trajectory (i.e., constant performance) during all ISI conditions and 2. response bias across all ISI conditions was overall liberal, and most liberal during the 4000 ms ISI condition.

Session 2 tracking measure

Average Tracking Distance during Hits: The linear model provided the best fit of the data according to our criterion (Table 10a), χ²(2) = 10.42, p <.01. Inspection of the model coefficients (Table 11a) and visual inspection of the graph (Fig. 9a) show: 1. a positive quadratic trajectory (i.e., rapid improvement in earlier block and slower improvement or deterioration in later blocks) in all ISI conditions; 2) a negative linear trajectory (i.e., improving performance) throughout the session in the 1000 ms ISI condition, but in the 2000 ms and 4000 ms ISIs this trajectory was qualified by worsening performance in later blocks; and 3. overall worse performance during the 1000 ms ISI than during the 2000 ms and 4000 ms ISIs.

Table 10.

Model comparison tables for tracking distance measures in E1 session 2—(a) hits, (b) CRs, and (c) commissions

Model	nPar	AIC	BIC	Loglik	Deviance	Χ²	df	p
(a) Model comparison table for tracking distance during hits
Base	5	− 2002.5	− 1978	1006.3	− 2012.5
Intercept	7	− 2033.8	− 1999.4	1023.9	− 2047.8	35.2418	2	2.23E−08***
Linear	9	− 2040.2	− 1996	1029.1	− 2058.2	10.4216	2	0.005457**
Quadratic	11	− 2036.3	− 1982.2	1029.1	− 2058.3	0.0701	2	0.965542
(b) Model comparison table for tracking distance during CRs
Base	5	− 1226.9	− 1202.5	618.47	− 1236.9
Intercept	7	− 1233.6	− 1199.3	623.78	− 1247.6	10.6232	2	4.93E−03**
Linear	9	− 1230.5	− 1186.4	624.25	− 1248.5	0.9497	2	0.621988
Quadratic	11	− 1228	− 1174.1	624.97	− 1250	1.4404	2	0.486663
(c) Model comparison table for tracking distance during commissions
Base	5	− 677.42	− 654.24	343.71	− 687.42
Intercept	7	− 678.63	− 646.18	346.31	− 692.63	5.21	2	7.39E−02*
Linear	9	− 678.53	− 636.82	348.26	− 696.53	3.9019	2	0.1421
Quadratic	11	− 674.71	− 623.73	348.36	− 696.71	0.1858	2	0.9113

Open in a new tab

* indicates p < .05, ** indicates p < .01, *** indicates p < .001

Table 11.

Chosen model coefficients tables for tracking distance measures in E1 session 2—(a) hits, (b) CRs, and (c) commissions

Fixed effects	Estimate	Std. error	df	t value	P
(a) Linear model coefficients for tracking distance during hits
(Intercept)	2.77E− 01	1.05E− 02	5.60E + 01	26.358	< 2e− 16***
time¹	− 1.81E− 02	6.22E− 03	9.52E + 02	− 2.909	0.00371**
time²	1.56E− 02	6.22E− 03	9.52E + 02	2.509	1.23E− 02*
ISI1	2.01E− 02	3.59E− 03	9.52E + 02	5.603	2.75E− 08***
ISI2	− 3.18E− 03	3.59E− 03	9.52E + 02	− 0.885	3.76E− 01
time¹:ISI1	− 2.72E− 02	8.79E− 03	9.52E + 02	− 3.09	2.06E− 03**
time¹:ISI2	0.020918	0.00879	952.000001	2.38	0.01752*
(b) Intercept model coefficients for tracking distance during CRs
(Intercept)	2.82E− 01	1.05E− 02	5.60E + 01	26.936	< 2e− 16***
time¹	− 3.87E− 03	9.49E− 03	9.33E + 02	− 0.408	0.6835
time²	1.93E− 02	9.48E− 03	9.33E + 02	2.037	0.042*
ISI1	1.37E− 02	5.50E− 03	9.34E + 02	2.483	1.32E−02*
ISI2	3.20E− 03	5.47E− 03	9.33E + 02	0.584	5.59E−01
(c) Intercept model coefficients for tracking distance during commissions
(Intercept)	2.56E− 01	1.04E− 02	5.38E + 01	24.526	< 2e− 16***
time¹	8.74E− 03	1.29E− 02	7.11E + 02	0.675	0.4997
time²	1.80E− 02	1.31E− 02	7.13E + 02	1.376	0.1692
ISI1	1.53E− 02	7.56E− 03	7.13E + 02	2.024	4.33E− 02*
ISI2	− 1.46E− 02	7.54E− 03	7.15E + 02	− 1.932	5.38E− 02

Open in a new tab

* indicates p < .05, ** indicates p < .01, *** indicates p < .001

Fig. 9 — Growth curve analysis graphs for tracking distance measures in E1 session 2—a hits, b CRs, and c commissions. Dots represent observed mean performance in each block and ISI combination, and lines represent the chosen model’s predictions

Average Tracking Distance during CRs: The intercept model provided the best fit of the data according to our criterion (Table 10b), χ²(2) = 10.62, p <.01. Inspection of the model coefficients (Table 11b) and visual inspection of the graph (Fig. 9b) show: 1. positive quadratic trajectories (i.e., improving then declining performance) during all ISI conditions; and 2. overall worse performance during the 1000 ms ISI than during the 2000 ms and 4000 ms ISIs.

Average Tracking Distance during Commissions: The intercept model provided the best fit of the data according to our criterion (Table 10c), χ²(2) = 5.21, p <.08. Inspection of the model coefficients (Table 11c) and visual inspection of the graph (Fig. 9c) show: 1. no overall significant temporal trends in the data and 2. overall worse performance during the 1000 ms ISI than during the 2000 ms and 4000 ms ISIs.

Discussion

With respect to the first part of our first question for this experiment, which was whether performance would change over time for each of the measures and indices associated with the CPT and tracking tasks, we note that some, but not all measures showed different performance trajectories across the session time-blocks. With respect to our second question, which was whether the temporal trajectories of any of our measures or indices would differ when target presentation rates are faster compared to when they are slower, we note that again, some, but not all measures showed different performance trajectories in the different ISI conditions.

Although our primary interest lies in the dual-task Session 2, results from the single-task Session 1 help validate our CPT paradigm. In Session 1, hits showed a consistent practice effects in all ISIs (Fig. 1c—dashed line). CRs were generally worse in the 1000 ms and 2000 ms ISIs than in the 4000 ms ISI but remained stable over time (Fig. 1a). RTs showed both linear and curvilinear performance across ISIs. Specifically, in the 1000 ms and 2000 ms RTs became slightly slower over the next few blocks and then slightly improved in the later blocks (Fig. 1d—solid line). In the 4000 ms ISI, performance became slower across the entire session (Fig. 1b—solid line). As for the SDT indices, sensitivity was worse in the 1000 ms and 2000 ms ISIs than in the 4000 ms ISI but remained stable over time, while response bias remained consistently liberal. Overall, these results suggest that the CPT in Session 1 was sensitive enough to detect some temporal performance changes (e.g., practice effects in hits) and vigilance decrements (progressively slowing RTs in the 4000 ms ISI). This was especially evident in how the ISI manipulation modulated RTs and revealed patterns aligned with underload-related decline. However, the task in Session 1 may not have been sufficiently demanding to elicit effects across all measures, as performance was dominated by extended practice effects. This limitation motivated the design of Session 2, where prior CPT experience and the added cognitive demands of multitasking were expected to produce clearer and more widespread patterns of performance changes.

In Session 2, performance on both CPT and tracking tasks also showed a mix of practice effects and vigilance decrements. Hits were relatively flat across ISIs, with overall worse performance in the 1000 ms ISI, suggesting that this high-demand condition taxed participants throughout. CRs followed curvilinear trajectories in the 1000 ms and 400 ms ISIs, indicating initial practice gains followed by vigilance decrements, while remaining stable in the 2000 ms ISI. RTs were the only CPT measure to show consistent vigilance decrements across all ISIs, with a steady decline in speed from the earliest blocks onward (Fig. 1b—solid line). Regarding the SDT indices, sensitivity mirrored performance in the CRs measure, with curvilinear trajectories in the 1000 ms and 4000 ms ISI, and flat performance in the 2000 ms ISI; and response bias remained constantly liberal during the session.

Tracking distance showed more nuanced patterns. During hits, tracking in the 1000 ms ISI condition progressively improved across the session, though gains slowed in later blocks. In the 2000 ms and 4000 ms ISIs, tracking showed mixed practice and vigilance decrement effects. In contrast, tracking during CRs followed a curvilinear trajectory across all ISIs consistent with early practice effects followed by later vigilance decrements, while tracking during Commissions remained relatively flat, indicating performance stability not strongly affected by either learning or fatigue. Notably, differences in tracking performance across ISIs during hits were not mirrored by corresponding differences in CPT Hit rates.

Overall, vigilance decrements in the CPT were most evident in declining CRs and sensitivity (in the 1000 ms and 4000 ms ISIs), and in RTs (across all ISIs). In the tracking task, vigilance decrements emerged primarily in the 2000 ms and 4000 ms ISIs during both hits and CR trials, though not during Commissions. These patterns suggest that under high cognitive demand, participants prioritized maintaining accurate target detection in the CPT task, particularly in the more time-pressured ISI conditions, even as performance on the concurrent tracking task declined.

Furthermore, the vigilance decrements observed in the 1000 ms and 4000 ms ISIs for CRs and sensitivity suggest that, in these two ISI conditions, despite their general preference for speed over accuracy (as indicated by response bias), participants prioritized responding accurately to true targets at the expense of inhibiting responses to non-targets. This may initially seem to suggest that maintaining a stable rate of target hits in these two ISI conditions similarly taxed executive resources. However, differences in the other measures between these two ISI conditions suggest that the similar temporal patterns in CRs and sensitivity are driven by different underlying mechanisms. In the 1000 ms ISI, as cognitive load increased, participants likely preserved target hit rates at the expense of CRs, leading to reduced sensitivity. This was also reflected in the longer practice effect in tracking distance during hits and in overall faster and less accurate responses in the 1000 ms ISI than in the other ISI conditions. In contrast, in the 4000 ms ISI condition, it is more likely that decreasing arousal may have caused greater difficulty in CRs, which is consistent with the overall higher hit accuracy and better tracking performance in this ISI condition, and with the linear decrease in RTs across blocks in all ISIs. Interestingly, performance appeared most stable in the 2000 ms ISI, suggesting an optimal balance of arousal and task demand avoiding both overload and under engagement (Wiener et al., 1984). To help the reader understand the overall pattern of results, Appendix B includes a summary table of the findings and their theoretical implications.

These findings highlight that vigilance decrements in E1 did not stem from a single source, but rather reflected multiple contributing mechanisms that varied across measures and ISI conditions. Building on this interpretation, Experiment 2 was designed to address two limitations. First, practice effects may have masked vigilance decrements in Session 2, as participants did not complete the tracking task in isolation (i.e., as the single task) beforehand. Second, the inclusion of three ISIs may have reduced statistical power to detect differences between the shortest and longest ISI conditions.

Experiment 2

The purpose of Experiment 2 (E2) was to test whether the results of E1 can be replicated and extended after addressing the two concerns about E1. Specifically, we made two changes to the paradigm in E2. First, we included an additional practice session to better acclimate participants to the tracking task and therefore increase the likelihood of observing vigilance decrements. Second, we removed the 2000 ms ISI condition to increase power. This experiment therefore addressed the same questions: whether and how performance would change across the six blocks for each of the specific measures associated with the CPT and tracking tasks, and how would the temporal trajectories of any of our measures differ when target presentation rates are faster compared to when they are slower? In particular, we asked whether the different patterns of vigilance effects in the 1000 ms and 4000 ms conditions will hold in this more powerful design.