Age-Related Effects of Executive Function on Takeover Performance: A Simulator Study in Automated Driving


 The development of highly automated vehicles (HAV) can meet elderly drivers’ mobility needs; however, worse driving performance after a takeover request (TOR) is frequently found, especially regarding non-driving related tasks (NDRTs). This study aims to detect the correlation between takeover performance and underlying cognitive factors comprising a set of higher order cognitive processes including executive functions. Thirty-five young and 35 elderly participants were tested by computerized cognitive tasks and simulated driving tasks to evaluate their executive functions and takeover performance. Performance of n-back tasks, Simon tasks, and task switching were used to generate updating, inhibition, and shifting components of executive functions by principal component analysis. The performance of lane changing after TOR was measured using the standard deviation of the steering wheel angle and minimum time-to-collision (TTC). Differences between age groups and NDRT engagement were assessed by two-way mixed analysis of variance.Older participants had significantly lower executive function ability and were less stable and more conservative when engaged in NDRT. Furthermore, a significant correlation between executive function and lateral driving stability was found. These findings highlight the interaction between age-related differences in executive functions and takeover performance; thus, provide implications for designing driver screening tests or human-machine interfaces.


Introduction
In 2021, the population aged 60 years and older was over 1 billion worldwide, and one in six people will be 60 years or older by 2030 1 . In ageing societies, the proportion of elderly drivers are rapidly increasing.
In Japan, one of the countries with the highest proportion of elderly citizens, 29.1% of the population was over 65 years old in 2021 2 , which will be 37.7% by 2050 3  Driving is an important way for the elderly to maintain their mobility, which is known to strongly affect their independence and quality of life 5,6 . However, elderly drivers are more likely to be involved in crashes at intersections, lane changing, and driving accidents involving multiple vehicles [7][8][9] . Some elderly people even have to give up driving due to the decline in their physical and mental conditions [10][11][12] .
The development of highly automated vehicles (HAV), also known as level 3 automation 13 , is expected to bene t elderly drivers by enhancing their mobility [14][15][16] . During level 3 automated driving, drivers can engage in non-driving related tasks (NDRTs) and can regain control of the vehicle when the system prompts a takeover request (TOR). Elderly drivers tend to engage in various NDRTs during simulated automated driving 17 , including entertainment, working, and dietary activities 18 .

Effects of NDRTs on takeover performance
Engaging in NDRTs is known to affect takeover performance, such as more collisions 19 and more deviation from the center of the lane 20 . The negative effect of NDRT is especially crucial for elderly drivers, considering the decline in their driving ability. Therefore, elderly drivers' behaviors after a TOR could be more cautious or conservative than younger drivers. For example, they prefer to brake more and maintain a longer safe distance 21,22,17 . Moreover, slower reactions and decision making among elderly drivers were found when regaining control after a TOR 23 . Moreover, after a period of manual driving, an increase in reaction time (RT) to a TOR was found for elderly drivers, but not for younger drivers 24 .
Furthermore, Wu et al. 25 revealed age-related differences in the interaction between NDRT and drowsiness and found that performing NDRT did not alleviate the drowsiness of elderly drivers but deteriorated their takeover performance. The distribution of visual attention can also differ between young and elderly drivers. That is, elderly drivers focused more on the road than NDRT compared with younger drivers 26 .
Moreover, they checked their back mirrors less frequently and their xation time became shorter when engaging in NDRT 27 .
Other studies reported no signi cant age effects on driving performance after a TOR 22 or takeover time 28 .
Despite the different study designs, these contradictory results suggest that other factors may underlie age-related differences. Li et al. 29 investigated takeover performance, behavior, and perceptions of elderly people in different age subgroups, suggesting that elderly people should be considered as a heterogeneous group. Beyond chronological age, it is important to explore individual differences in driving performance in response to TOR using explicit factors.

Executive function and elderly drivers' driving behaviors
Cognitive abilities decline with age 30 ; therefore, elderly drivers with worse cognitive abilities are more likely to exhibit dangerous driving behaviors and be involved in crashes [31][32][33] . Driving is a complex task that requires various cognitive abilities. For instance, visual attention and visuospatial cognition are related to overall driving performance 34,35 : perceptual-cognitive capacity is related to speed control and crash risk 36 , simple RT is related to hazard perception 37 , and working memory is related to decision making 38 .
Adrian et al. 39 concluded that functional abilities were more determinant than chronological age in predicting elderly drivers' performance. Studies have shown that age-related decline in various cognitive tasks could be attributed to the aging of the frontal lobe, also known as the frontal aging hypothesis 40,41 .
Furthermore, the frontal lobe has been demonstrated to be correlated with executive function (EF) 42,43 . EF comprises a set of higher-order cognitive processes that coordinate lower-level processes 44 , and can be de ned as processes that control and regulate thought and action 45 . EF may play a crucial role in driving tasks because it requires complex cognition-required task sets to summarize information and supervise actions 46 , such as maintaining continued attention, dealing with irrelevant information, and adapting to different task requirements 47 .
Three main components of EF were distinguished: "Shifting," "Updating," and "Inhibition" 48 . The shifting component refers to shifting between multiple tasks, operations, and mental sets 49 37 , and driving errors in lane position 52 . Researchers also found that inhibition ability, which is related to selective attention, was associated with road test scores 53 , on-road driving performance 35 , and observation errors 52 . Updating ability, which is related to working memory 54 , plays an important role in drivers' decision making 38 .
In comprehensive studies dedicated to the three main components of EF, updating 39,55 , inhibition 56 , and shifting 39 components were found to be signi cantly correlated with driving performance. Walshe et al. 47 reviewed various subprocesses of EF and concluded that inhibition and updating played important roles in car crashes. Moreover, EF components resulting from principal component analysis (PCA) 55 and con rmatory factor analysis 56 showed clear correlations between these latent factors and driving performance. The use of underlying, structured components yielded from the results of cognitive tests rather than pure task performance can alleviate the complexity of EF research brought on by the so-called "task impurity problem." Executive tasks often involve other cognitive tasks; therefore, EF test results may provide additional individual differences unrelated to EF 48 . Exploring the correlation between driving performance and EF components as latent variables can avoid these uncertainties and produce more valid results 56 . Therefore, we expected that the use of latent components generated from EF task performance could effectively establish a correlation with takeover performance.
Overall, several studies have focused on age-related differences in takeover performance, as well as the importance of abilities related to EF in driving behaviors. However, research on the relationship between age-related differences in EF components and takeover performance in automated driving is rare. To ll this gap, we designed a series of computerized cognitive tests and simulated driving tasks to evaluate drivers' EF and takeover performance. EF components were extracted by PCA and their correlations with takeover performance were investigated. The purpose of this exploratory study is two-fold: rst, to examine whether age-related differences exist in EF and takeover performance in automated driving, and second, to explore the relationship between speci c EF components and takeover performance.

Participants
Seventy participants living near Tsukuba, Ibaraki, Japan, were recruited for this study; 35 of them were elderly drivers above 65 years (M = 72.8 years, SD = 3.6 years), including 17 males and 18 females.
Furthermore, 35 were younger drivers below 35 years (M = 27.8 years, SD = 4.4 years), including 24 males and 11 females. The sex ratio did not differ signi cantly between groups (chi-squared = 2.12, df = 1, p = 0.145). All participants provided informed consent complying with the research protocol approved by the Institutional Review Board of the National Institute of Advanced Industrial Science and Technology (AIST). All participants were licensed drivers and had at least one year's driving experience, 90% of them drove more than 1,000 kilometers per year (34% =1,000-5,000 km per year; 58% = more than 5,000 km), and drove frequently in daily life (30% drove 3-4 days per week, and 59% drove daily).

Procedure
Before the experiment, the experimenter welcomed and introduced the participants to the experiment purpose and contents. Participants were then introduced to three cognitive tasks. Prior to each task, participants were given detailed task instructions and su cient training. After completing the cognitive tasks, participants were given a 10-minute break. Hereafter, they were introduced to the simulator-based driving tasks. Participants were provided with two driving practice sessions: a manual driving session with simple maneuvers (e.g., lane changing), and an automated driving session with TOR scenarios similar to that used in the formal experiment. After familiarizing themselves with the simulator operation and driving task requirements, participants completed the experimental driving session.

Materials
The driving task was conducted on a set of desktop driving simulators at the AIST. The driving simulator apparatus included a set of Logitech steering wheel and pedals as input devices, a Panasonic screen (87 mm × 154 mm) with an approximate 45-degree eld-of-view for visual presentation, and a set of Mitsubishi Driving Simulator software packages for monitoring and recording.

Tasks for executive functions
All participants were asked to complete three computerized tasks on a laptop, measuring the EF components. For all tasks, participants observed visual stimuli shown on a laptop screen running Psychopy software and responded only to target stimuli by pressing previously speci ed buttons on the keyboard. Each task lasted 12-15 minutes, with a 5-minute break between tasks. Illustrations of each task are shown in Fig. 1.

Updating
This component was assessed using a classic version of the sequential letter n-back task 57,58,55,59 .
In the 0-back condition, the target was the rst letter of sequence in the block; in the 1-back condition, the target was any letter identical to the letter immediately preceding it; in the 2-back condition, the target was any letter identical to the letter appearing two trials before. Letters in uppercase appeared in random order for 500 ms with a 2,000 ms interval before the onset of the next trial.
Participants completed 12 trials in each block and 12 blocks in total. Before each block, the task type of this block was shown at the center of the screen. Each block and trial were presented in a randomized order.

Inhibition
This component was assessed using a visual version of the Simon task paradigm 60,61 .
Each block began with a xation cross at the center of the screen. In each trial, a red or green circle appeared on the left or right side of the cross and remained visible for 500 ms. Participants were instructed to press the left button when they saw a green circle, and the right button when they saw a red circle. After the offset of the stimulus, there was a 2,000 ms interval before the onset of the next trial.
Participants completed 12 trials in each block and 12 blocks in total. Six of the blocks were congruent conditions, consisting of congruent trials in which the stimuli were presented on the same side as the associated response button. The other six blocks were mixed conditions, in which each block consisted of half congruent and half incongruent trials (stimuli were presented on the opposite side as the associated button). Each block and trial were presented in a randomized order.

Shifting
This component was assessed using a variant digit version of the task switching paradigm 62-64 .
This paradigm included two digital-related tasks. In each trial, a numeric digit (from 1 to 9, excluding 5) was presented at the center of the screen and was surrounded by a solid or a dashed square. The task rules were determined as follows: if the square was solid-lined, participants needed to determine whether the digit was greater or less than 5; if the square was dash-lined, they needed to determine whether the digit was odd or even. Participants were instructed to press two buttons on the keyboard according to different responses. Digits were presented for 1,500 ms in each trial with a 2,000 ms interval before the onset of the next trial.
Participants completed 12 trials in each block and 16 blocks in total. Eight blocks were task homogeneous conditions, in which only one type of task would appear, and the other eight blocks were task heterogeneous conditions, in which two types were mixed and each type of task would take turns to appear twice. The choice for alternating runs was believed to increase participants' working memory demands when they tried to keep track of the tasks; thus, was a more sensitive way to detect group differences 63,65 . Each block and trial were presented in a randomized order.

Driving experiment
A 2 (age) × 2 (NDRT) between-and within-subjects mixed factor experiment was designed. The betweensubjects factor was age (young and elderly drivers), and the within-subject factors were NDRT engagement (no additional task and task engaged).

Driving task scenario
Participants encountered a variant version of the Lane Changing Task 66 after TOR in a simulated level 3 automated vehicle. Similar scenarios in the automated driving version are commonly used to evaluate drivers' takeover performance when engaging with NDRT 17,67,24 . In this study, we chose a typical scenario: a lane-blocking stationary truck was suddenly revealed by the lane changing of the front car (see Fig. 2).
The participants in the self-driving car (vehicle 1) were in automated driving most of the time, until warned by an audio warning message "Transition to manual driving" in Japanese as the TOR and were then asked to take over full control (both steering wheel and pedals). Participants had to change lanes to avoid collision and pass the stationary truck (vehicle 2), which was revealed by the lane changing of a leading car (vehicle 3). The speed set for automated driving and all moving vehicles were 25 m/s. The TOR occurred at 100 m in front of the stationary truck. Moving cars (every 50 m per vehicle) were set in the right lane next to the participant's current lane, so the participants also had to pay attention to avoid collisions. The scenario ended when the participant drove 200 m after the stationary truck, also noted by an audio notice of "Automated driving starts." Based on previous research on the time budget or lead time for takeover, the common choices were in the range between 4.5 and 7.5 s 17,24,68 . In this study, we set time budget to 4 s to make the scenario riskier and put the participants under more pressure.

Non-driving related task
The NDRT during autonomous driving was chosen as a sequential-letter 0-back task. For every 2 s, the voice of a letter was randomly played by the speaker, and participants were required to press a button on the steering wheel if they heard the target letter. The NDRTs continued for approximately 40 trials until the TOR occurred, and ended when the TOR audio was played.
The auditory, but not visual, NDRT was designed to avoid additional visual distraction in the tra c environment. Based on our pilot study, the current version of the auditory 0-back task of moderate di culty was chosen to keep participants (especially elderly participants) involved in the NDRT during automated driving.

Executive function components
The measures for the working memory updating component were RTs for each correct response and accuracy in each condition in the n-back task. For the inhibition component, RTs for each correct response and accuracy in both congruent and mixed blocks were recorded, and the Simon effect 69 was calculated based on the difference in RTs between congruent and incongruent trials in mixed blocks in the Simon task. For the shifting component, RTs for each correct response, accuracy in task homogeneous and task heterogeneous conditions, and switch cost 70 were calculated by the difference in RTs between trials after switching and non-switching trials in task heterogeneous conditions in the taskswitching task.

Driving performance
Driving performance data were recorded at a rate of 60 Hz. Raw data included position, speed, and acceleration data of the vehicle, as well as driver input data (e.g., steering wheel and pedal operations).
According to similar studies 24,25 , we chose two proven valid measures to evaluate the quality of takeover performance: sdSteer: standard deviation of the steering wheel angle, calculated by the standard deviation of the steering wheel angle after TOR before the end of manual driving. This indicated the ability to regain control of the vehicle and smooth lane change. The simple effect test demonstrated that for RTs, in the 0-back (t = 1.19, df = 63.46, p = 0.240) and 1-back (t = 1.68, df = 67.67, p = 0.097) tasks, there were no signi cant differences between age groups, while the elderly group had signi cantly longer RTs (t = 3.55, df = 64.81, p < 0.001). For response accuracy, there was no signi cant difference between age groups in the 0-back (t = -0.27, df = 67.48, p = 0.785) task, but there were signi cant losses in accuracy for the elderly group in the 1-back (t = -2.81, df = 50.87, p = 0.007) and 2-back (t = -6.36, df = 58.67, p < 0.001) tasks.

Simon task performance
We found main effects between age groups in which younger participants performed faster in RTs (F(1,68) = 15.71, p < 0.001) than the elderly; however, no signi cant difference in accuracy (F(1,68) = 1.27, p = 0.264) between age groups was found. We also detected a signi cant increase in RTs (F(1,68)

Task switching performance
We found main effects between age groups where younger participants had faster RTs (F(1,68) = 28.73, p < 0.001) and higher accuracy (F(1,68) = 10.56, p = 0.002) than the elderly. The main effect of task switching was also detected; in switch trials, participants showed longer RTs (F(1,68) = 369.99, p < 0.001) and less accuracy (F(1,68) = 28.27, p < 0.001). Furthermore, there were two-way interactions in the RTs between the age groups and switch conditions (F(1,68) = 11.87, p < 0.001). This can be explained by a signi cantly larger difference in RTs between switch and non-switch trials (the switch cost, also shown in Table 1) found in the elderly group than in the younger group (t = 3.45, df = 53.8, p = 0.001). The descriptive statistics for EF tasks are presented in Table 1.

Principal component analysis for executive function components
The PCA results of the selected measures of performance in EF tasks (normalized data) are shown in Table 2. The measures comprised RTs and accuracy results for the tasks. Accuracy in the 0-back condition was not selected because the differences between individuals and groups were not signi cant.
The measures for RTs in the Simon task and task switching was chosen as the Simon effect and switch cost in each task. The loadings for each component are presented in Table 2. The results showed that only PC1-PC4 had eigenvalues greater than one, yet PC1-PC5 explained over 80% (82.30%) of the variability. Therefore, a ve-component solution was used for further analyses. The interpretation of the main composition of each principal component and the EF components is shown in Table 3.

Correlations between executive function components and driving performance
Correlations between EF components (results from the PCA of EF measures) and driving performance (sdSteer and TTC) were analyzed for each condition (with/without engagement in NDRT) using Pearson's correlation coe cients (see Fig. 4).
A signi cant correlation was found between sdSteer and accuracy of the updating component (PC2, r = 0.26, p = 0.029) without NDRT engagement, and a marginally signi cant correlation for sdSteer and cost of shifting component (PC5, r = -0.22, p = 0.075) was found. Moreover, no signi cant correlation was found between TTC and EF components.
When engaging in NDRT, sdSteer was signi cantly correlated with the cost of the shifting component (PC5, r = -0.29, p = 0.015); no signi cant correlation was found between TTC and EF components.

Discussion
This study investigated age-related differences in the effects of NDRT on takeover performance and examined how individual differences in speci c components of EF could in uence takeover performance during automated driving.

Executive function performance
The results of EF tasks showed that younger participants had better performance in both reaction speed and accuracy in most conditions of all tasks, indicating that they had better working memory updating, inhibition control, and mental shifting abilities. The age-related differences were consistent with those of previous studies on EF 71,72 . Results of accuracy failed to show signi cant differences between age groups in some conditions, such as the 0-back, 1-back, and congruent Simon task. This may have been because the task could have been too easy for participants in both age groups to detect differences in results. Surprisingly, we did not nd a signi cant difference in the Simon effect, which was larger in the elderly group than in the young group in similar studies 61 . However, in our study, the individual difference in the Simon effect was large within the elderly group. Considering that the accuracy for Simon tasks was relatively high, even for the elderly group, we believe that our Simon task was not di cult enough to detect differences in inhibition control, but mainly re ected components about choice RT.
Moreover, the PCA results of the cognitive tasks data yielded a construct of three main EF components: working memory updating, inhibition control, and mental shifting. This result was consistent with previous research 55,39,56 , suggesting that these three EF components are separable, while the tasks could be correlated with one another 48 . Furthermore, our construct clearly distinguished between reaction speed (PC1, PC4) and accuracy (PC2, PC3) related components in both updating and inhibition; this could contribute to a detailed analysis of the role speed and accuracy played in the correlation between each EF component and driving performance. We chose the components that could describe over 80% of the cumulative percent variation. Although the eigenvalues of shifting cost components (PC5) were less than one, we still found signi cant correlations with driving performance, suggesting that it is a reasonable and important component.

Driving performance
In the analysis of differences in driving performance between age groups and engagement in NDRT, we found signi cantly less stable lateral maneuvering in elderly drivers (p = 0.002) engaging in NDRT (p = 0.016). These ndings are consistent with previous studies on the in uence of age 18, 25 and NDRT 17,20 on drivers' takeover performance. Moreover, the NDRT used in most previous studies required visual attention, while the NDRT in this study (i.e., n-back task) mainly required auditory attention, indicating that pure cognitive workload could signi cantly in uence takeover performance 19,73 .
In the analysis of the TTC at the moment of lane change, we found no signi cant difference between age groups; however, a longer TTC was detected when engaging in NDRT, although the effect was only marginally signi cant (p = 0.089). The results of the number of participants who stepped on the acceleration pedal after the TOR showed that younger drivers (15 out of 35) tended to accelerate when engaged in NDRT compared to elderly drivers (6 out of 35). This suggests that the higher TTC in our study may be explained by the conservative behaviors of elderly drivers dealing with a TOR when engaged in NDRT. Similarly, previous studies found that elderly drivers in similar scenarios tend to exercise more caution and drive slower 22,17 , or apply more frequent and stronger brake to maintain a longer TTC 21 after transitioning to manual driving.

Correlations between executive function components and driving performance
The main nding in this study was that better ability to update and shift components was found to be signi cantly correlated with more stable lateral driving performance after takeover. The shifting component was found to be important for lateral control in both situations, regardless of engagement in NDRT. Similar results were also found in previous studies 39 , wherein the shifting ability measured by the score of the plus-minus task was signi cantly correlated with on-road driving performance. Moreover, a better shifting ability is considered to re ect drivers' performance in switching between the NDRT and driving tasks in urgent situations. A marginally signi cant correlation between shifting components and driving stability appeared even without engaging in NDRT, indicating that shifting ability may play a more important role in a takeover task, compared with other driving tasks mentioned in previous studies, such as manual driving lane changing 55 . This could be partly explained by the complexity of a takeover task: it requires the driver to shift more frequently and urgently between different task sets. The results also showed that engaging in NDRT strengthened the correlation between shifting ability and driving performance. Furthermore, previous studies have detected a relationship between driving behaviors and shifting ability in a trail making test 74,39,34 , wherein the task may contain both cognitive shifting (involving visual attention) and task shifting components. In this study, both the task for evaluating shifting ability (variant digit version of task switching) and the NDRT (audio 0-back task) did not require a shift in visual attention. Thus, even without visual attention shifting, mental shifting ability in uenced drivers' performance in responding to the TOR.
The updating component was also signi cantly correlated with a more stable driving performance, which is consistent with previous research 39, 55 . This signi cant correlation with the working memory updating accuracy was found in situations without NDRT. For a takeover driving task, the content of working memory must be updated and irrelevant information deleted; therefore, poor updating retains irrelevant information and reduces the processing of relevant information 39 . Thus, when encountered with a situation requiring takeover, drivers with poor updating ability could not process the incoming information (such as speed and position status for both self-driving and other vehicles) very well, which led to less stable driving performance. Another important nding was that a signi cant correlation with the accuracy component of working memory updating was found. In our study, the design of the driving task left a relatively su cient time budget for the drivers to take over. This showed that the accuracy of updating (ability to update correct information) compared with the speed of updating (ability to update swiftly), may be more important for lateral stability in a takeover task.
Previous research has demonstrated the contribution of inhibition components to driving performance 56 .
In addition, selective attention, which is related to inhibition ability, was found to be linked to driving performance 35,52 . However, this study did not detect a signi cant correlation between inhibition components and driving performance under any condition; this nding was still consistent with a previous similar study in which various EF components were involved 39,55 . There may be two possible reasons for this inconsistent nding. First, our design of an easy Simon task may have led to poor discrimination of inhibition ability. That is, we found no signi cant difference in the Simon effect between age groups; thus, it may be di cult to detect su cient individual differences in inhibition ability by using the Simon task performance. Second, in our experiment scenarios, the priority of NDRT was lower than taking over the vehicle; thus, the NDRT might have been considered less crucial than the driving task by the participants. That is, as our measures for the inhibition component focused on active suppression for a dominant response, takeover performance could be less relevant to this measure if the NDRT was not prepotent over the takeover task. In such case, the variance of takeover performance may be more related to individual differences in shifting rather than inhibition ability.
We detected a signi cant difference in TTC between age groups but found no signi cant correlation between EF components and TTC. Although the TTC re ects the safety margin in responding to the TOR, it could also be in uenced by drivers' driving style and strategy in urgent situations rather than EF. That is, other non-EF-related factors, such as personality and attitudes 75 , may affect the driving behaviors of elderly drivers.

Application
Our ndings contribute to a better understanding of drivers with age-related differences in EF ability interacting with a TOR in HAV. Knowledge of the relationship between individual differences in EF and driving behavior can be used to evaluate driving capability. Research on general assessment, such as the trace-route task method, also showed that it could be a valid and reliable tool for evaluating EF and distinguishing drivers with risky driving behaviors 76 . Similarly, our ndings can guide quick screening for elderly drivers' cognitive status for a TOR of HAV, such as an exam for elderly drivers' license updating, or a lite system monitoring the cognitive status of elderly drivers. Moreover, an understanding of this relationship could help human-machine interface (HMI) design to support elderly drivers by considering their EF status. Previous studies also suggested that age-related changes in cognitive functions should be considered in the design and application of automated driving for elderly drivers 77 , and the perceived usability of HMI design is related to cognitive performance 78 . This study's ndings can further provide more information about drivers' behaviors and cognitive status. Thus, these results can bene t HMI design, especially for drivers with age-related declines in EF.

Limitations
While the study revealed the possible correlation between EF components and driving performance after a TOR, there were several limitations that must be discussed.
First, the use of a desktop simulator may have limited our ability to discover valid correlations. A previous study found that the association of EF components with driving could be mediated by computer game skills, and the relations were even stronger when the skilled participants were excluded 55 . Thus, future research should utilize more real-world driving simulators with higher delity to investigate driving behaviors. Second, due to the design of the NDRT, participants answered the NDRT by pressing buttons on the steering wheel, which means that the RTs after the TOR could have been di cult to measure precisely. Since EF may be highly correlated with RTs, further investigation on the correlation between RTrelated measures and EF ability is needed. Third, the task evaluating inhibition control was not effective enough to reveal individual differences and to yield signi cant results for a correlation with driving performance. Therefore, further research is necessary to build credible correlations between driving performance and inhibition ability using more valid cognitive measures.

Conclusion
In this study, we designed and conducted a series of computerized cognitive and simulated driving tasks to evaluate EF and takeover performance. We examined 1) the EF components of updating, inhibition, and shifting yielded by results from cognitive tasks, and 2) takeover performance of lateral maneuvering by steering wheel standard deviations and longitudinal safety margin by the TTC at the moment of lane change. The results demonstrated a signi cantly lower ability of EF among elderly participants and less stable and more conservative driving behavior of elderly drivers when engaged in NDRT. We further observed a signi cant correlation between EF (working memory updating accuracy and shifting cost) and takeover performance. Overall, these ndings provide important insights into elderly drivers' driving behaviors and EF, which can be utilized to expand on the present study's ndings and improve driving safety for elderly.
Declarations Figure 1 Illustration of three cognitive tasks. Participants were given detailed instructions before each task. All instructions were shown in Japanese during actual experiments.

Figure 2
Scenario of takeover request and driving tasks.