Sample size was determined primarily with feasibility in mind. Only four participants could partake at the same time. Given that the experiment could only start in the last week of April 2021, we had a short window of opportunity before the late sunset would make it too troublesome for participants to come to the test site after dark. Based on this and considering possible days with poor weather conditions, we expected to run about 40 participants at best. A sensitivity analysis conducted in G*Power (Faul et al., 2009) for a within-subject ANOVA with 4 repeated measures (i.e., on each of 4 streets; see Design) and assuming a .50 correlation between repeated measures showed that with 40 participants we had 90% power to detect a small to moderate effect size of f = .21 at α = .05.

Unfortunately, technical issues forced us to have three rather than four participants simultaneously per session. As a result, only a total of n = 29 people participated in the experiment. An adjusted sensitivity analysis to this sample size showed that we had 90% power to detect a small to moderate effect size of f = .25 at α = .05. Participants were recruited from the J. F. Schouten database of Eindhoven University of Technology (TU/e). Out of the 29 participants, 13 (44.8%) identified as female and 16 (55.2%) as male. The average age of the participants was M = 23.1 (SD = 3.3; range 19 to 30). Of the 29 participants, 12 (41.4%) and 10 (34.5%) indicated to be “very familiar” and “somewhat familiar” with the TU/e campus respectively. Of the remaining participants, 3 (10.3%) indicated to be “a bit familiar” with the campus and 4 (13.8%) to be “not familiar at all”. Participants received €10 as compensation for their participation.

We conducted a one-factor (4 streets), within-subject, quasi-experimental field study with response time (RTs) and accuracy in an ATCT as objective markers of alertness, and self-reports of sleepiness, arousal, safety and lighting appraisal as subjective outcomes. The experiment took place on 4 pre-selected streets that differed, amongst other things, in the uniformity of lighting (see Figure 2, Table 1 and Table 2). To reduce the time needed to move from one street to the other, and to keep the experiment manageable, we did randomize the starting street, but we did not fully randomize the order of the streets. Instead, Street 2 always followed Street 1, Street 3 followed Street 2, Street 4 followed Street 3, and Street 1 followed Street 4. During each 5 min walk participants engaged in two simultaneous tasks. We selected a dual task approach to increase perceptual load and reduce the chance that the slower and/or inaccurate responses were caused by boredom (Fotios et al., 2015). The primary task was walking along the street, following coloured ribbons that marked the path. An ATCT, which was adapted from the earlier study by Burtt (1916) was the secondary task. Reaction times and accuracy to the ATCT were recorded. At the end of each walk participants completed a short survey.

The experiment took place on the TU/e campus during six days between 22^nd of April and 5^th of May 2021, at least 15 minutes after sundown. Each night consisted of one or two experimental sessions with maximum three participants each, starting between 21:00 and 22:45 hrs. Weather conditions were rather similar between days, except for May 5 when it rained. The average temperature was 9⁰ C. (range 3⁰ C. to 15⁰ C.).

After a short introduction to the experimental procedure, each participant was guided from the starting point to their respective first street. Here, they received a survey and pen attached to a clipboard worn around the neck. They completed the first page of the questionnaire (age, gender, and familiarity with the test site) and were then asked to wear a backpack carrying a laptop to which were attached a two-button response box and a pair of headphones (see Figure 3).

In the first session, participants practiced 10 trials of the ATCT. They were instructed via a pre-recorded voice message to press as fast and as accurately as possible the leftmost green button in response to a high pitch sound, and the right most red button in response to a low pitch sound. During these trials they would receive feedback whether they had responded correctly, incorrectly or not fast enough. No feedback was given during the experimental sessions. When the participant started walking, the experimenters positioned themselves out of the participants’ view. For five minutes the participant walked the street back and forth while following the coloured ribbons and performing the ATCT. After five minutes, the participant was asked, using a pre-recorded message, to walk to the middle of the street and to complete the remaining pages of the questionnaire. These consisted of the Stanford Sleepiness Scale (SSS; Shahid et al., 2012), the Activation Deactivation Adjective Checklist (ADACL; Thayer, 1986), three items for perceived environmental safety (PES; van Rijswijk et al., 2016) and six items for perceived lighting quality (PLQ; van Rijswijk & Haans, 2018).

When they completed the survey, they were escorted to the next street by another experimenter. After having walked all four streets, an experimenter escorted the participant back to the meeting location, where they were thanked and reimbursed for their participation.

Overall, participants gave the correct response in 5154 (i.e., 97.0%) of the 5311 trials. Based on a series of Shapiro-Wilk tests — one for each street — with ω ≥ .97 as the criterion for sufficient normality, the proportion of errors was not found to be normally distributed in any of the streets with ω ≤ .88. Using the Skillings-Mack test, we found no statistically significant differences between the streets on PE, with SM = 1.0 and p = .800 (see Appendix, Figure A3A; for descriptive statistics, see Table 3). There were no outliers (i.e., with standardized scores of |Z| ≥ 4; Shiffler, 1988). Graphical summary of the results can be seen in Figure 4.

Reaction times ranged from 203 to 1996 ms. RTs did not meet our normality criterion, with ω ≤ .84 (for descriptive statistics, see Table 3). Therefore, we used the inverse of the reaction times (1/RT) or speed in the remainder of the analyses. After this transformation, the Shapiro-Wilk tests showed sufficient normality for all streets with ω ≥ .98. To test for differences in speed, we used a Mixed Linear Model with speed as dependent variable, street as fixed, and person as random factor. Residuals were found to be normally distributed with ω = .99. We found a statistically significant effect of street with Wald χ²(3) = 23.9 and p < .001. Post-hoc paired comparisons revealed that people responded, on average, faster in Street 4 (473 ms) as compared to Streets 1 (490 ms), 2 (500 ms), and 3 (492 ms), with Z ≥ 2.93 and p ≤ .003 (see Appendix, Figure A3B). None of the other paired comparisons yielded a significant difference, with |Z| ≤ 1.80 and p ≥ .072. Removal of 27 outliers (i.e., with standardized scores of |Z| > 3; Shiffler, 1988) did not alter the interpretation of these findings. As Street 4 was the most non-uniform of the four streets, this finding is consistent with the findings of Burtt’s experiments on ATCT performance (1916), and thus in support of Hypothesis 1.

With ω ≤ .95, CV_rt scores did not meet our normality criterion in any of the streets, and hence the Skillings-Mack test was used to test for differences between streets on sustained attention. No statistically significant differences were found with SM = 0.4 and p (with ties) = .929 (see Appendix, Figure A3C; for descriptive statistics, see Table 3). There were no outliers (i.e., with standardized scores of |Z| ≥ 4).

Responses to the sleepiness item were found to have sufficient normality for all streets, with ω ≥ .98. There were no outliers (i.e., with |Z| > 3). A repeated measures ANOVA with sleepiness as dependent variable and street as factor, showed no differences in sleepiness between the four streets (see Appendix, Figure A4A), with F(3,84) = 0.38, and p = .750 (Huynh-Feldt corrected). Graphical summary of the results can be seen in Figure 4.

Energetic arousal was not found to be normally distributed in one street, with ω ≥ .95. There were no outliers (i.e., with either |Z| > 3 or |Z| ≥ 4) depending on whether normality was met (Shiffler, 1988). Using the Skillings-Mack test, we found no differences between streets in self-reported energetic arousal (see Appendix, Figure A4B), with SM = 1.7 and p (with ties) = .580.

Tense arousal was not found to be normally distributed in any of the streets, with ω ≤ .83. There were no outliers (i.e., with |Z| ≥ 4). Using the Skillings-Mack test, we found no differences between streets in self-reported tense arousal (see Appendix, Figure A4C), with SM = 2.7 and p (with ties) = .205. Descriptive statistics per street are reported in Table 3.

PES was not found to have sufficient normality in three of the streets, with ω ≥ .94. There were no outliers (i.e., with either |Z| > 3 or |Z| ≥ 4 depending on whether normality was met; Shiffler, 1988). Using the Skillings-Mack test, we found streets to differ in PES to a statistically significant extent, with SM = 6.3 and p (with ties) = .033 (see Appendix, Figure A5A). Post-hoc paired comparisons revealed that Street 3 was perceived as safer (M = 4.02) as compared to Streets 1 (M = 3.70), 2 (M = 3.70) and 4 (M = 3.75), with SM ≥ 2.3 and p ≤ .040 (in support of Hypothesis 2). No other statistically significant differences in PES were found between the other streets, with SM ≤ 0.6 and p ≥ .444. Graphical summary of the results can be seen in Figure 4.

PLQ was not found to have sufficient normality in three of the streets, with ω ≥ .94. There were no outliers (i.e., with |Z| ≥ 4). Using the Skillings-Mack test, we found streets to differ in PLQ, with SM = 31.3 and p < .001. Post-hoc paired comparison revealed that Street 3 had better PLQ evaluations (M = 3.79) as compared to Streets 1 (M = 3.07), 2 (M = 2.89) and 4 (M = 2.77), with SM ≥ 15.2 and p < .001 (see Appendix, Figure A4B). No other statistically significant differences in PLQ were found between the other streets, with SM ≤ 0.9 and p ≥ .216. Descriptive statistics per street are reported in Table 3.

We estimated all pairwise correlations using the within-subject method by Bland and Altman (1995). For this purpose, we calculated an average speed on the ATCT for each participant and each street. We found a moderate negative correlation between energetic arousal and sleepiness of r = -.58 (see Table 4). As expected, higher energetic arousal was associated with less sleepiness. Tense arousal also had a negative association with sleepiness albeit to a lesser extent with r = -.25. As expected, PLQ was positively correlated with PES with a moderate correlation of r = .53. In line with our expectations, we found a different association between PES on one hand, and energetic and tense arousal on the other hand. Whereas, as expected, tense arousal was associated with lower ratings of PES, with r = -.43, energetic arousal was found to be positively related to PES, with r = .34. Correspondingly, the two dimensions of arousal showed a dissimilar association with PLQ, with energetic arousal positively correlated, with r = .29, and tense arousal negatively, with r = -.27. We found a positive correlation between PE and ATCT speed (1/RT) which suggests the presence of a speed-accuracy trade-off. However, we found only very small and statistically non-significant correlations between ATCT speed and any of the self-report measures. Exclusion of outliers did not affect the interpretation of these correlations. Graphical summary of the results can be seen in Figure 4.

We managed to partially replicate Burtt’s (1916) findings about the positive effect of non-uniform lighting on pedestrian’s alertness, indicated by their reaction speed. People responded, on average, faster on Street 4, which had the lowest uniformity. However, there were no differences in the accuracy of responses across the four streets. Similar to the results of Burtt’s experiment, we also found no clear relation between lighting uniformity and CV_rt. Regarding our second hypothesis, differences on self-reports were indeed in line with previous studies showing positive effects of illuminance levels and uniformity on perceived lighting quality (PLQ) and perceived environmental safety (PES) (Boyce, 2014; Bullough et al., 2020; Fotios et al., 2019; Narendran et al., 2016; Peña-García et al., 2015; Portnov et al., 2020). Street 3, which had the highest horizontal and vertical illuminance values and highest uniformity, was perceived significantly safer and PLQ was rated significantly higher than the other streets. However, Street 4, which had the lowest uniformity, was not rated lower in perceived safety than the other streets. Hence, we did not manage to confirm our second hypothesis.

We succeeded in finding some insights regarding our proposed research framework (Jedon et al., 2023). The correlation between energetic arousal and Stanford Sleepiness Scale (SSS) is in line with our expectations as the energetic dimension is presumed to be linked with the sleep-wake cycle (Thayer, 1978; Thayer, 1986). However, SSS and energetic arousal did not vary between streets. We found no correlations between the ATCT and any of the other metrics. The lack of any correlation between RTs and either SSS or energetic dimension of arousal can be considered as a confirmation of the literature about sleepiness and alertness being two different constructs and captured through different metrics (Matthias et al., 2010; Shapiro et al., 2006). Alternatively, although RT tasks are commonly used as a measure of alertness (e.g., PVT, Go / No-Go) they might not be sensitive enough, or can be confounded by effort invested. This needs to be further investigated, for example by having more extreme conditions (e.g., very dark, very bright), a more demanding dual task (e.g., scanning the environment rather than walking), or the addition of effort measures. Using other metrics, namely the PE and CV, showed, although not always significant, slightly higher and more promising correlations with the self-reports and might be therefore of further interest in future research.

We found, as expected (Boyce, 2014; Bullough et al., 2020; Fotios et al., 2019; Narendran et al., 2016; Peña-García et al., 2015; Portnov et al., 2020), a correlation between PES and PLQ. Next, we found correlations of both dimensions of arousal with PES and PLQ but in different directions. The energetic dimension of arousal was positively correlated with both PES and PLQ. The tense arousal seems to have a reverse impact; the higher the tense arousal the lower were the ratings of PES and PLQ. Arousal is often measured only in its intensity. Our findings suggest that Thayer’s (1978) categorization of energetic and tense arousal can bring more insights in research on urban lighting and pedestrian safety perceptions.

The current study reintroduced attention towards concepts of alertness and arousal in the domain of urban lighting via a replication of a largely overlooked study (Burtt, 1916). It tested the sensitivity of relevant metrics in a quasi-field experiment and managed to find effects of urban lighting conditions, potentially uniformity driven, on both perceived safety and alertness. Moreover, promising correlates of perceived environmental safety — energetic and tense arousal — showed opposing relationships with this outcome variable. The study combined both subjective and objective measures and exposed participants to real settings, instead of using proxies (e.g., photographs, videos).

One major limitation of the present study was that we, in contrast to Burtt (1916), were not able to directly manipulate uniformity and illuminance. The streets differed in other characteristics as well, such as the presence of greenery, construction works or surrounding buildings. Hence, we cannot make any causal claims with respect to the effect of lighting on alertness and PES. In addition, and as in any field experiment, there were also momentary situational events that may have affected the results. A few examples include fireworks, passers-by (pedestrians, bicycles and cars) or the presence of animals. Furthermore, headlights from cars and uncontrolled lighting from surrounding buildings could have affected the lighting conditions at the time. These spontaneous events could not be captured in our daily measurements of lighting conditions as they were done after finishing the experimental sessions. The occurrence of these real-life events, however, also adds to the ecological validity of this study as all of these effects could possibly occur in daily life. Future studies should directly manipulate the light conditions within otherwise similar environments or otherwise attempt to measure and control for other (not necessarily light related) street characteristics (e.g., by measuring prospect, concealment, entrapment; van Rijswijk et al., 2016; van Rijswijk & Haans, 2018). Alternatively, one could employ day–night time comparisons (Fotios et al., 2019).

The sample of participants was rather homogenous in the manner of their familiarity with the site, as most of the participants were students from the university and knew the campus. This may explain why the PES ratings of all streets were relatively high and with little variation as familiarity with the environment is found to have a strong impact on its perceived safety (Crosby & Hermens, 2019; Day et al., 2003; Traunmueller et al., 2016). Hence, the streets might not have been perceived as unsafe at all because the participants were familiar with the environment. If the streets would be unfamiliar to the participants, they might consider them as more unsafe and have different requirements from the surrounding lighting, in particular on its uniformity (Bullough et al., 2020; Fotios et al., 2019; Kimura et al., 2014; Narendran et al., 2016; Nasar & Bokharaei, 2017; Portnov et al., 2020). Future studies should try to use participants that are not familiar with the environment. Another solution would be to manipulate the environmental factors in a way to make the setting more unfamiliar and/or make them seem more unsafe (e.g., manipulate the factors of concealment and entrapment).

Another possible confounding variable that we did not take into account, was the walking speed of participants. Donker et al. (2011) found that lower illuminance levels lead to increased walking speed, but Pedersen and Johansson (2018) found the opposite relationship. However, walking could also be understood as a form of exercise. Hence, participants’ speed could have an effect on the energetic dimension of arousal, as it is connected with the voluntary motor activity (Thayer, 1978). Moreover, Donker et al. (2011) proposed increased fear of the participants as an explanation for the faster walking speed under darker conditions. Walking speed could therefore be also an indicator of perceived safety, anxiety and/or tense arousal.

Last, the current field study does not allow us to attribute effects on alertness or safety perceptions to either visual or non-visual pathways. In fact, both may well play a part: for instance, alertness may be enhanced by visually perceived external cues (Matthias et al., 2010) as well as by non-visual pathways (Cajochen, 2007; Lok et al., 2018; Xu & Lang, 2018). As acknowledged earlier, our study does not even allow to conclusively attribute effects to lighting parameters per se. If, in future studies one would like to disentangle the underlying visual and non-visual mechanisms, one may have to make use of metameric light conditions. Such however, was not the goal of the present study. Our main aim was to explore and test, for the first time, methods and measurement instruments for disentangling the relationships between safety perceptions, anxiety, alertness, and arousal.

To uncover what pedestrians’ needs are in respect to urban lighting, we suggest a shift of research perspective from a primary (or even singular) focus on visual performance tasks towards studying also psychophysiological concepts, such as alertness, arousal and anxiety. Our results show that these psychophysiological concepts, mainly energetic and tense arousal, are relevant when studying safety perceptions. However, more controlled experiments are needed to ascertain whether results such as ours were affected by lighting conditions or other environmental characteristics. Likewise, further research is needed to understand how exactly alertness, arousal and anxiety interrelate in the manifestation of the experience of safety.

Our findings about the impact of non-uniform lighting, its positive effect on the speed of responses without a negative effect on perceived safety, could be important for the discussion on minimizing the unfavourable effects of urban lighting. Positive effects of non-uniform urban lighting could be also welcomed by architects and lighting designers as these are more encouraging for creative and artistic lighting (Boyce, 2014; International Association of Lighting Designers, 2012; Jedon et al., 2022). However, due to the lack of experimental control and other confounds typical for field experiments we should not draw firm conclusions and should not attribute any effects found solely to the uniformity of urban lighting. It is impossible to make claims on the causal relationship between lighting characteristics, including uniformity, and RTs, in a quasi-experimental field study. Therefore, better controlled research is needed on uniformity of lighting.