How specific blue and yellow light changes affect the human internal body clock

To determine the dim-light melatonin onset (DLMO), we used all available evening data points and applied the ‘Hockeystick’ algorithm⁸⁵. We chose a default area of interest upper border or threshold of 5 and selected ‘Hockeystick Time’. In rare cases, the threshold had to be adapted for some individuals. C.B., C.C. and M.S. independently performed the DLMO analyses with CC and MS being blinded to the light exposure conditions. Any divergences were resolved in a final discussion with Dr Mirjam Münch, a colleague at the Centre for Chronobiology, who is very experienced with applying the ‘Hockeystick’ algorithm. In two cases (1× background, 1× blue-dim condition), it was not possible to determine a DLMO on at least one evening as no clear increase in melatonin concentrations could be identified. The inspection of the data showed that the assumptions for mixed linear models were met (for more details, see Supplementary Methods).

The inspection of the data showed that rank transformation of the data resulted in an increased compatibility with the assumptions for mixed linear models. Thus, we additionally report results for rank-transformed data (for details, see Supplementary Methods). Analyses of melatonin values yielded strong evidence against a condition difference during the light exposure (BF₁₀ = 0.09) suggesting that the data were approximately 11 times more likely to occur under H0. This result was largely confirmed using analyses based on rank-transformed data, which yielded moderate evidence against H1 (BF_{10 rank-based} = 0.14). For an overview of the condition mean (intercept) and deviations from the intercept sampled from the posterior distribution, please see Supplementary Table 8. There was moderate (inconclusive) evidence against a condition × time interaction (BF₁₀ = 0.13; BF_{10 rank-based} = 0.13). Please see Fig. 3b for an illustration of the results and Supplementary Fig. 1 for an illustration of the melatonin concentrations on evenings 1 and 2.

Also here, the inspection of the data showed that rank transformation resulted in an increased compatibility with mixed linear model assumptions (for details, see Supplementary Methods). Thus, we additionally report results for rank-transformed data. Analyses yielded anecdotal and thus inconclusive evidence against a condition difference regarding the subjectively reported sleepiness on the KSS during the light exposure. Under the H0, the data were approximately 1.6 times more likely than under the H1 (BF₁₀ = 0.62; BF_{10 rank-based} = 0.23). For the condition mean (intercept) and deviations from the intercept sampled from the posterior distribution, please see Supplementary Table 9. Furthermore, there was moderate (inconclusive) evidence against a time × condition interaction (BF₁₀ = 0.27; BF_{10 rank-based} = 0.14). For a graphical illustration, see Fig. 3d, and for an illustration of the KSS values on evenings 1 and 2 in each light condition, please see Supplementary Fig. 6 (note this was not part of the analysis plan, wherefore we refrain from a statistical analysis).

We assessed objective sleepiness, that is, the alpha (8–12 Hz) to theta (4–7 Hz) ratio during a 3-min Karolinska Drowsiness Test (KDT) with eyes open at the beginning, after 30 min, and at the end of the 1-h light exposure at parieto-occipital electrodes. The inspection of the data indicated that rank transformation resulted in an increased compatibility with the model assumptions. Thus, we additionally report results for rank-transformed data (for details, see Supplementary Information). There was moderate (inconclusive) evidence against a condition effect with the data being approximately 3.5 times more likely under H0 than under H1 (BF₁₀ = 0.28; BF_{10 rank-based} = 0.09). For an overview of the condition mean (intercept) and deviations from the intercept sampled from the posterior distribution, please see Supplementary Table 10. Analyses additionally yielded strong evidence against a time × condition interaction (BF₁₀ = 0.06; BF_{10 rank-based} = 0.08).

Visual comfort was calculated as the average rating from the responses to the questions about how pleasant the lighting was generally, how participants perceived the level of brightness, how glaring the light source was, and how pleasant participants rated the colour temperature (that is, cold vs. warm). The inspection of the data showed that the assumptions for mixed linear models were met (for more details, see Supplementary Methods). There was moderate (inconclusive) evidence against a difference between the conditions regarding visual comfort experienced during the light exposure. The data were approximately six times more likely under the H0 than under the H1 (BF₁₀ = 0.16). Supplementary Table 11 provides an overview of the condition mean (intercept) and deviations from the intercept sampled from the posterior distribution. For the condition × time interaction, there was moderate evidence in favour of H1 (BF₁₀ = 7.68). For a graphical illustration, see Supplementary Fig. 2.

The inspection of the data suggested that the assumptions for mixed linear models were largely met, and rank transformation of the data did not improve compatibility (for more details, see Supplementary Methods). Analyses revealed extreme evidence in favour of a condition difference during the light exposure with the data being approximately 153 times more likely under the H1 than the H0 (BF₁₀ = 153.33). More precisely, there was conclusive evidence in favour of faster median RTs in the background (mean_background 384.4 ± 35.5 ms) than the yellow bright (mean_{yellow -bright}395.5 ± 32.6 ms; BF₁₀ = 29.06) or the blue dim condition (mean_{blue- dim}399.8 ± 39.1 ms; BF₁₀ = 260.1). There was only anecdotal (inconclusive) evidence against a difference between the yellow-bright and the blue-dim conditions (BF₁₀ = 0.39). Supplementary Table 12 provides an overview of the condition mean (intercept) and deviations from the intercept sampled from the posterior distribution. There was anecdotal evidence against a condition × time interaction with the data being approximately two times more likely under the H0 compared with H1 (BF₁₀ = 0.47). For a graphical illustration, see Fig. 3c and Supplementary Figs. 3 and 4 additionally provides an illustration of the PVT results on evenings 1 and 2 (note this was not included in the analysis plan, wherefore we refrain from statistical analyses).

The inspection of the data suggested that the assumptions for mixed linear models were met and transformation of the data did not increase compatibility with the assumptions (for more details, see Supplementary Methods). For the 10% fastest RTs during the light exposure, there was only anecdotal and thus inconclusive evidence in favour of a condition difference with the data being approximately 1.8 times more likely under H1 compared with H0 (BF₁₀ = 1.77). Supplementary Table 13 provides an overview of the condition mean (intercept) and deviations from the intercept sampled from the posterior distribution. Regarding the condition × time interaction, there was anecdotal evidence against H1 (BF₁₀ = 0.42). Supplementary Fig. 3 provides an illustration of the results. Supplementary Fig. 4 additionally provides an illustration of the PVT results on evenings 1 and 2 (note this was not included in the analysis plan, wherefore we refrain from statistical analyses).

Rank transformation of the data improved especially normality of the residuals and homoscedasticity. Thus, we additionally report analyses based on rank-transformed data. For the slowest 10% RTs during the light exposure, there was conclusive evidence in favour of H1, that is, a condition difference (BF₁₀ = 60.9; BF_{10 rank-based} = 119.8). As for the median RTs, there was conclusive evidence in favour of RTs being faster in the background (mean_background 481.6 ± 67.8 ms) compared with the yellow-bright (mean_{yellow- bright}511.0 ± 81.6 ms; BF₁₀ = 22.6; BF_{10 rank-based} = 40.86) and the blue-dim conditions (mean_{blue dim}516.9 ± 87.4 ms; BF₁₀ = 46.15; BF_{10 rank-based} = 67.31). However, there was moderate (inconclusive) evidence against a condition difference between the blue-dim and the yellow-bright conditions (BF₁₀ = 0.26; BF_{10 rank-based} = 0.23). Supplementary Table 14 provides an overview of the condition mean (intercept) and deviations from the intercept sampled from the posterior distribution. There was moderate to strong evidence against a condition × time interaction (BF₁₀ = 0.09; BF_{10 rank-based} = 0.1). For a graphical illustration, see Supplementary Fig. 3. Supplementary Figure 4 additionally provides an illustration of the PVT results on evenings 1 and 2 (note this was not included in the analysis plan, wherefore we refrain from statistical analyses).

The inspection of the data indicated that the assumptions for mixed linear models were met, and rank transformation did not increase compatibility with the assumptions (for more details, see Supplementary Methods). Analyses yielded inconclusive evidence against a condition difference regarding the onset latency to 10 min of continuous sleep. The data were approximately theee times more likely given the H0 than the H1 (BF₁₀ = 0.31). The mean latency to 10 min of continuous sleep was 17.0 ± 42.2 min in the background, 9.4 ± 13.2 min in the yellow-bright, and 7.5 ± 4.3 min in the blue-dim condition. An overview of the condition means (intercept) and standard deviations sampled from the posterior distribution is provided in Supplementary Table 15. Supplementary Fig. 5 provides an illustration of the results. For a comprehensive overview of the sleep data for each light condition and visit, see Supplementary Table 16.

Data inspection suggested that the assumptions for mixed linear models were met. Rank-transforming the data did partially result in an increased compatibility with the assumptions, wherefore we report results from analyses based on rank-transformed data (for more details, see Supplementary Methods). Note that during two nights one participant had a very short rapid eye movement (REM) sleep latency (that is, <30 min), which would have resulted in a very short non-REM (NREM) part. We thus decided to combine the NREM parts of the detected first and second sleep cycles into one. Please also note that in one participant we were unable to analyse the data in one condition (yellow-bright) as it was too noisy. Analyses of SWA (0.5–4.5 Hz) at frontal electrodes F3, Fz and F4 during the NREM part of the first sleep cycle yielded conclusive evidence against a condition difference (BF₁₀ = 0.03; BF_{10 rank-based} = 0.03). Thus, the data were approximately 33 times more likely under the H0 than under H1. Evidence for a time × condition difference remained inconclusive (anecdotal evidence against H0; BF₁₀ = 0.42; BF_{10 rank-based} = 0.43). For a graphical illustration, see Fig. 3e. Supplementary Table 17 provides an overview of the condition mean (intercept) and deviations from the intercept sampled from the posterior distribution.

The inspection of the data showed that rank transformation of the data slightly increased compatibility with the assumptions for mixed linear models. We therefore additionally report results for rank-transformed data (for details, see Supplementary Information). There was strong evidence against a condition difference in melatonin concentrations before the beginning of the light exposure (BF₁₀ = 0.03; BF_{10 rank-based} = 0.03) suggesting that the data were approximately 33 times more likely under the H0 than under the H1. Supplementary Table 18 provides an overview of the condition mean (intercept) and standard deviations sampled from the posterior distribution. For the condition × time interaction, there was extreme evidence against H1 (BF₁₀ = 0.003; BF_{10 rank-based} = 0.003). For a graphical illustration, see Fig. 3b.

The inspection of the data suggested that the assumptions for mixed linear models were largely met, and rank transformation of the data did not improve compatibility (for details, see Supplementary Methods). Before beginning of the light exposure, analyses yielded strong evidence in favour of a condition difference (BF₁₀ = 17.42). Follow-up analyses indicated evidence in favour of faster RTs in the background (mean_background 359.1 ± 36.7 ms) compared with the yellow-bright (mean_{yellow- bright}365.2 ± 42.5 ms) or the blue dim condition (mean_{blue- dim}365.8 ± 37.9 ms; BF₁₀ = 19.8). There was only moderate inconclusive evidence in favour of a difference between the yellow-bright and the blue-dim conditions (BF₁₀ = 5.49). For an overview of the condition mean (intercept) and standard deviations sampled from the posterior distribution please see Supplementary Table 19. Regarding a condition × time interaction, there was extreme evidence against H1 with the data being approximately 222 times more likely under the H0 compared with H1 (BF₁₀ = 0.004). Figure 3c provides a graphical illustration of the results.

The inspection of the data showed that the assumptions for mixed linear models were met (for details, see Supplementary Information). For the subjective sleepiness before the beginning of the light exposure, there was extreme evidence in favour of H1, that is, a condition difference with the data being >100 times more likely under the H1 compared with H0 (BF₁₀ = 397,569,714). More specifically, participants felt more tired in the background (mean_{KSS background}6.19 ± 1.6) compared with the yellow-bright condition (mean_{KSS yellow- bright}5.56 ± 1.7; BF₁₀ = 548,934; extreme evidence in favour of H1) and to the blue-dim condition (mean_{KSS blue- dim}5.47 ± 1.5; BF₁₀ = 321,238,205; extreme evidence in favour of H1). There was moderate (inconclusive) evidence against a condition difference between the yellow-bright and the blue-dim conditions (BF₁₀ = 0.16; moderate evidence against H1). Supplementary Table 20 provides an overview of the condition means (intercept) for the subjective sleepiness scores and standard deviations sampled from the posterior distribution. Analyses further yielded very strong evidence against a condition × time interaction with the data being approximately 86 times more likely under the H0 than under H1 (BF₁₀ = 0.01). For a graphical illustration, see Fig. 3d.

The inspection of the data showed that the assumptions for mixed linear models were rather not met, particularly the residuals were not normally distributed (for details, see Supplementary Information). However, a rank transformation did not solve this issue, wherefore we refrain from reporting analyses based on rank-transformed data. There was extreme evidence in favour of a condition difference between the three light exposure conditions before the light exposure (BF₁₀ = 110.93) with the data being approximately 111 times more likely under the H1 than the H0. The mean perceived brightness (± standard deviation) in the background condition (range 1–5) was 3.3 ± 0.56, in the blue-dim condition 3.4 ± 0.53, and in the yellow-bright condition 3.5 ± 0.55. More specifically, there was extreme evidence in favour of a condition difference between the background and the yellow-bright conditions (BF₁₀ = 752.3), moderate (inconclusive) evidence against a difference between background and the blue-dim conditions (BF₁₀ = 0.31) and moderate (inconclusive) evidence in favour of a difference between the blue-dim and yellow-bright conditions (BF₁₀ = 7.72). For an overview of the condition mean (intercept) and standard deviations sampled from the posterior distribution, please see Supplementary Table 21. For the condition × time interaction, there was anecdotal (inconclusive) evidence against an effect (BF₁₀0.36).

Participants came for an initial adaptation night (8-h sleep opportunity) to the laboratory to screen for potential sleep disorders (see below for the PSG setup) and to make them acquainted with the laboratory setting. Participants showing signs of sleep disorders or presenting with sleep efficiency <70% would have been excluded from the study.

For resting-state and nocturnal PSG recordings (that is, EEG, EOG and EMG) we used ambulatory BrainProducts LiveAmp devices. We planned to record PSG at a sampling rate of 250 Hz (was actually 500 Hz; see ‘Deviations from protocol’ section) from 21 scalp channels mounted in a cap (EasyCap) and 4 EOG channels with FCz as the online reference. In addition, we placed two mastoid electrodes for later re-referencing according to the sleep staging criteria of the American Association of Sleep Medicine, as well as two chin EMG electrodes and two electrocardiogram electrodes⁷⁹. During the adaptation nights, we additionally recorded from two EMG electrodes on the tibialis anterior muscle of one leg to screen for nocturnal leg movements, a respiration belt and a nasal cannula to screen for respiratory problems.

Participants were planned to be asked to rate or respond to various aspects of the light exposure using a six-question, seven-item Likert scale questionnaire¹². This questionnaire was administered in German. The questions were planned to be about the comfort of light (‘Allgemein ist das Licht angenehm’; überhaupt nicht (1) to sehr stark (7)), the perceived brightness (‘Wie empfinden Sie die Helligkeit des Lichtes?’; sehr dunkel (1) to sehr hell (7)), light level preference (‘Ich hätte es lieber …’; deutlich dunkler (1) to deutlich heller (7)), glare (‘Dieses Licht blendet mich’; überhaupt nicht (1) to sehr stark (7)), the perceived colour temperature (‘Wie empfinden Sie die Lichtfarbe?’; sehr kalt (1) to sehr warm (7)) and general wellbeing (‘Wie fühlen Sie sich im Moment?’; unwohl (1) to wohl (7)). For deviations from the planned questionnaire, please see the ‘Deviations from protocol’ section.

Light exposure was planned to be delivered from a vertical front lighting panel (width 220 cm, height 140 cm) that consists of 24 light-emitting diode (LED) panels (RGBW), each containing 144 LEDs (that is, total of 3,456 LEDs) covered by diffusing material^80,81. Two participants were planned to sit next to each other at a distance of 1.5 m from the wall, facing the wall. The LED primaries have the following properties (peak wavelength ± full width at half maximum; CIE 1931 xy chromaticity): blue (462 ± 23 nm; 0.14, 0.06), green (518 ± 32 nm; 0.16, 0.72), red (631 ± 16 nm; 0.70, 0.30) and white (peaks 446 nm and 560 nm; 0.33, 0.34). Due to the coronavirus pandemic, an alternative setup with two separate laboratories had to be used. For details, please see the ‘Deviations from protocol’ section.

The wall was be calibrated from a vantage point centred between the two participants, and both irradiance and radiance measurements were planned to be taken. To characterize spectral shifts of the LED wall, we planned to measure the spectrum of each of the RGBW primaries at 16 settings at 8-bit resolution.

Stimuli were either a constant light matched in its photoreceptor activation profile to that of daylight at 6,500 K (D65, used during the first experimental visit; ‘background condition’; Fig. 1f (note that the Stage 1 protocol wrongly stated that this light was used during the adaptation night and the second night of each experimental visit)), or sinusoidally flickering light stimulating the S cones in an balanced but opponent fashion with luminance (blue-dim: planned +37% S-cone contrast, planned −37% L + M contrast from background, Fig. 1g; yellow-bright: planned −37% S-cone contrast; planned +37% L + M cone contrast from background, Fig. 1h), alternating flickering for 30 s to constant background for 30 s. The flicker frequency was 1 Hz so as to provide a continuous tonic signal to the cones that might otherwise adapt to a continuous light biased towards either −S + (L + M) or +S − (L + M) (Fig. 1e). This frequency was also used in the habituation stimulus in a seminal study determining the cardinal directions of colour space¹⁸. To avoid adaptation to flicker, we flickered for 30 s, then held the light constant at background for 30 s, corresponding to the approximate time scale of habituation in the purported S-opponent psychophysical channel¹⁸. All stimulus conditions are summarized in terms of their nominal photoreceptor excitation⁸² in Table 1.

To find the settings on the RGBW channels of the lighting panel, we implemented an optimization minimizing the squared error between desired and actual contrasts using ‘fmincon’ routine as implemented in MATLAB (MathWorks). Critically, our stimuli were designed to produce no differential stimulation of melanopsin-expressing ipRGCs (0%).

During the scheduled day (that is, from wake-up until the end of the protocol), participants were planned to be in dim light (<8 lux) provided by fluorescent lighting.

Analyses of salivary melatonin were planned to be done with enzyme-linked immunoabsorbent assays (ELISAs) in an in-house laboratory. We planned to use ELISA kits (NovoLytiX), which have a minimal detection limit (limit of quantification) of at least 1.6 pg ml⁻¹. Eventually, a RIA was used instead of an ELISA (see ‘Deviations from protocol’ section). The DLMO was determined by fitting evening melatonin profiles by a piecewise linear-parabolic function using the hockey-stick algorithm (planned: v2.4, used version: 2.5; see ‘Deviations from protocol’ section) to calculate the DLMO⁸⁵.

For the PVT data, trials with RTs ≥100 ms were considered valid trials³⁹. We then computed the median RTs as well as RTs for the fastest and slowest deciles of valid trials. KSS ratings were analysed without further pre-processing steps. As for the neuroendocrine responses, statistical evaluation was planned to be with repeated-measures ANOVAs and planned contrasts (note that we implemented an ANOVA-like approach using linear models to circumvent case-wise deletion; see ‘Deviations from protocol’ section).

Following data acquisition, the signal from EEG channels was filtered and re-referenced to a linked mastoids reference. For the analyses of objective sleepiness, we first corrected for eye movements using an independent component analysis, which was followed by manual exclusion of other artefacts (for example, muscle artefacts). Subsequently, artefact-free data were segmented into 2-s time bins and subjected to fast-Fourier transformations (FFT) using the ‘mtmfft’ function as implemented in the Fieldtrip toolbox yielding a frequency resolution of 0.5 Hz. We then extracted absolute power in the theta (4–7 Hz) and alpha (8–12 Hz) frequency range and averaged across parietal and occipital electrodes (that is, P3, Pz, P4, O1, Oz and O2) and segments within each frequency band. We then computed the alpha/theta ratio. For sleep analyses, we had planned to score sleep semi-automatically with an algorithm (The SIESTA Group^86,87). These analyses were planned to be based on EEG data from electrodes F3, F4, C3, C4, O1 and O2 (downsampled to 128 Hz and re-referenced to the contralateral mastoid electrode) along with the signal from the EOG channels that have been placed according to Rechtschaffen and Kales criteria and the chin EMG signal. However, for budgetary reasons, we eventually used the Somnolyzer algorithm as implemented in the Philips Respironics Sleepware G3 software instead of the SIESTA scoring (see ‘Deviations from protocol’ section). Additionally, we computed EEG SWA (that is, delta power density between 0.5 Hz and 4.5 Hz) as an indicator of sleep propensity across the night within each NREM part of a sleep cycle. To this end, we computed EEG SWA for each decile of the NREM part of the first NREM-REM cycle^88,89. For the computation of delta power density, artefact-free data were segmented into 2-s time bins and subjected to FFTs yielding a frequency resolution of 0.5 Hz. Thereafter, FFT results were averaged in the 0.5–4.5 Hz range at frontal electrodes F3, Fz and F4 within each percentile of each NREM cycle. For each NREM cycle, the analyses thus yielded ten measures per participant. The analysis procedure described here is based on a publication by Chellappa and colleagues²⁵.

Only data from participants for whom the light exposure took place during the rising arm of the melatonin curve and after the DLMO were included in the analyses.

Our sample size and recruitment approach were intimately linked to the analytic strategy. We intended to collect data until obtaining sufficient evidential strength for our primary hypothesis (C1) that the circadian phase shift is larger in the yellow-bright flickering condition than in the blue-dim flickering condition and the constant light condition, or until reaching our resource limit (n = 16). Formally, this would have been achieved by calculating sequential Bayes factors³⁵. First, the repeated-measures ANOVA described above would have been run sequentially after data from each new participant was included in the data set. Participants would then have continuously been recruited for the study until sufficient evidence for the full model would have been reached (male, female alternating). We consider a BF of 10 as ‘strong’ evidence, following standard categorizations: 1 < BF < 3—anecdotal evidence, 3 < BF < 10—moderate, 10 < BF < 30 to BF >30—very strong evidence. According to the sequential Bayes factor approach, if this threshold was reached for the primary hypothesis, the study would have been halted, and the evidence strength would have been reported according to the Bayes factor. Otherwise, the study would have been halted if the resource limit is reached (n = 16). A minimum number of four participants (two men, two women) were planned to participate in the study, in case the evidence threshold would have been reached with one, two or three participants already.

The protocol lasted for 32.5 h instead of the 32 h previously described in the text, because we initially did not take into account that the last melatonin sample and round of assessments would take approximately 15–20 min.

Saliva samples were frozen at −28 °C rather than the previously stated −20 °C, because we used a different freezer.

Instead of a 6-min version of the PVT, we used a 10-min version. Specifically, the 10-min version has been associated with increased sensitivity to modulations of both sleep homeostatic and circadian drives and to improvements in alertness after wake-promoting interventions³⁹. A shorter version has only been recommended if the protocol does not allow for the 10-min version^92,93. As the protocol allowed us to use the longer version, we felt using the longer version would improve the validity of our measurements.

Instead of five snacks every 2.5 h with 0.2 of the total basic metabolic rate estimated with the Mifflin–St. Jeor equation⁷⁸, we served six snacks of 0.17 of the total basic metabolic rate every 2.5 h. This was to shorten the time between the last snack and the bedtime to 5.5 instead of 8 h, and thus to prevent participants from becoming too hungry in the evening. Additionally, we allowed participants to use their phones during a ‘social hour’ scheduled 4 h after wake-up as we have previously experienced this to increase adherence to long study protocols. Importantly, participants wore orange-tinted blue-blocking glasses during this time and were instructed to keep display brightness as low as possible.

We recorded data at a sampling rate of 500 Hz instead of the previously planned 250 Hz with 500 Hz being the standard sampling rate in our laboratory. However, as we were interested only in frequencies up to 12 Hz, this change has not affected the analyses or results.

Due to internal communication issues, the visual comfort questionnaire slightly deviated from the planned version and corresponded to a version used in an earlier study conducted in the laboratory. In more detail, instead of a seven-item Likert scale, a five-item Likert scale was used. Furthermore, the questions slightly deviated from the initially planned ones. Importantly though, the relevant questions to assess visual comfort were still included. Contrary to what had been planned, participants were not asked to rate their light level preference (‘I would prefer it to be lighter/darker’) and their general wellbeing (‘I feel well/unwell’). Instead, participants rated the light according to how wake-promoting versus tiring it was (‘Die Beleuchtung in diesem Zimmer…’; hilft mir, wach zu bleiben (1) to ermüdet mich stark (5); Engl.: ‘The lighting in this room…’; helps me to stay awake (1) to is very tiring (5)) and whether they felt it helped them to concentrate (‘Die Beleuchtung in diesem Zimmer…’; hilft mir, mich besser zu konzentrieren (1) to stört meine Konzentrationsfähigkeit (5); Engl.: ‘The lighting in this room…’; helps me to concentrate (1) to disturbs my ability to concentrate (5)). These two questions were however not (planned to be) included in the calculation of the visual comfort score; the relevant questions were still asked.

Due to the coronavirus pandemic, it was not feasible to have two participants sit in front of the planned LED wall for several hours next to each other. Thus, we used two custom-made displays, which provided a stimulus that was functionally equivalent to the originally planned stimulus, in two separate comparable (sleep) laboratories. Specifically, light exposure delivered through a modified 27-inch display consisting of a total of five sets of primaries, with peak spectral emissions at 430, 480, 500, 550 and 630 nm (refs. ^44,94). The backlights were independently controlled at 8-bit resolution using Digital Multiplex (DMX). For an illustration of the laboratory setup, please see Supplementary Picture 1.

Due to some technical differences in the device primaries between the planned LED wall and the displays, the validated contrast between the ‘blue-dim’ and the ‘yellow-bright’ conditions was ±25% instead of the originally planned nominal contrast of ±37% (blue-dim: +25% S-cone contrast, −25% L + M contrast from background, Fig. 1g; yellow-bright: −25% S-cone contrast; +25% L + M cone contrast from background, Fig. 1h). This validated contrast (±25%), although lower, however still yields a psychophysically substantial and physiologically meaningful contrast on the post-receptoral channel.

Sitting accommodations and furniture were placed in a way to prevent illuminances higher than 8 lux, and, if necessary, participants were instructed to not take positions in the room associated with higher illuminance levels. Nevertheless, for short periods of time, illuminances may have exceeded 8 lux (Supplementary Table). 6

Analyses of salivary melatonin were done with radioimmunoassay (RIA) rather than ELISAs in the NovoLytiX laboratory rather than an in-house laboratory. RIA and ELISAs by NovoLytiX generally reveal comparable results, but the laboratory uses RIA as analyses are less time consuming. The RIA has an analytical sensitivity of 0.2 pg ml⁻¹ and a limit of quantification of 0.9 pg ml⁻¹. For the calculation of the DLMO, we used version 2.5 of the hockey-stick algorithm (planned: v2.4), which was the most recent one available at the time of the data analysis⁸⁵.

Unlike originally planned we did not use the service provided by the SIESTA group for budgetary reasons. Instead, sleep was scored automatically with the Somnolyzer algorithm as implemented in the Philips Respironics Sleepware G3 software v. 4.0.1.0 (refs. ^86,87). Importantly, the Somnolyzer algorithm is based on the algorithm used by the SIESTA group and an automatic scoring software that has been cleared by the U.S. Food and Drug Administration (FDA). Due to the change of the algorithm, analyses also included data from electrodes P3 and P4 in addition to F3, F4, C3, C4, O1 and O2. Furthermore, the data did not have to be downsampled to 128 Hz before analysis. However, the G3 software required us to apply some filtering before the sleep stage analyses. Filtering was according to AASM criteria⁹⁵ with the EEG and EOG signals being filtered between 0.3 Hz and 35 Hz and the EMG between 10 Hz and 100 Hz with a Notch filter at 50 Hz to remove line noise.

Instead of a classic ANOVA design using the function ‘anovaBF’, we implemented an ANOVA-like design using linear models and the function ‘lmBF’. We chose to use linear models instead of the classic ANOVA function as this would have resulted in a case-wise deletion in case of missing data (for example, if data from even just 1 out of 14 PVTs per evening were not available or one DLMO out of the three conditions could not be determined).