Species sensitivities to artificial light at night A phylogenetically controlled multilevel meta-analysis on melatonin suppression

The rapid urbanization of our world has led to a surge in artificial lighting at night (ALAN), with profound effects on wildlife. Previous research on wildlife's melatonin, a crucial mechanistic indicator and mediator, has yielded inconclusive evidence due to a lack of comparative analysis. We compiled and analysed an evidence base including 127 experiments with 437 observations across 31 wild vertebrates using phylogenetically controlled multilevel meta-analytic models. The evidence comes mainly from the effects of white light on melatonin suppression in birds and mammals. We show a 36% average decrease in melatonin secretion in response to ALAN across a diverse range of species. This effect was observed for central and peripheral melatonin, diurnal and nocturnal species, and captive and free-living populations. We also reveal intensity, wavelength, and timing-dependent patterns of ALAN effects. Exposure to ALAN led to a 23% rise in inter-individual variability in melatonin suppression, with important implications for natural selection in wild vertebrates, as some individuals may display higher tolerance to ALAN. The cross-species evidence has strong implications for conservation of wild populations that are subject to natural selection of ALAN. We recommend measures to mitigate harmful impacts of ALAN, such as using ‘smart’ lighting systems to tune the spectra to less harmful compositions.


I N T RODUC T ION
The rapid pace of urbanization has led to a proliferation of artificial lighting at night (ALAN), a typical form of light pollution.ALAN significantly alters indoor and outdoor light environments and affects a vast majority of the human population globally, with nearly 80% of the world's population and 99% of Americans and Europeans exposed to it (Falchi et al., 2016).The ubiquitous exposure to ALAN poses a range of challenges, particularly to human health and the health of wildlife and ecosystems, which have yet to be fully appreciated by the scientific community (Cho et al., 2015;Gaston, 2019).Indeed, several studies have documented the negative effects of ALAN on human health, pointing to an increased risk of cancer (Walker et al., 2020), disease (Münzel et al., 2021), obesity (Rybnikova et al., 2016), and sleep problems (Aulsebrook et al., 2018).Similarly, researchers have also summarized the detrimental effects of ALAN on wildlife's fitness (Dominoni et al., 2016;Russart & Nelson, 2018) and ecosystem functions (Gaston et al., 2013).However, most of these evidence syntheses have not employed quantitative methods, such as modern meta-analytic techniques, and instead have utilized qualitative approaches to provide a narrative synthesis of the current evidence on ALAN.For example, the use of such qualitative approaches is limited in quantifying the magnitude of the effects of ALAN and the drivers of heterogeneous effects.
Evidence syntheses using meta-analytic techniques increase the sample size and reduce the effects of individual study bias, allowing for a more robust and quantitative understanding of the effects of ALAN (Gurevitch et al., 2018).Recently, Sanders et al. (2021) conducted a quantitative synthesis that highlighted the wide-ranging effects of ALAN exposures on organismal physiology (e.g., hormone levels) and life history traits (e.g., number of offspring).This synthesis provided valuable insights into the general trends of ALAN's biological and ecological effects and served as a pivotal evidence base for a more in-depth synthesis to unravel mechanistic pathways through which ALAN exerts its harmful effects.One possible underlying mechanism is the disruption of endogenous circadian rhythms (e.g., biological clock), including the suppression of melatonin secretion-the main output of circadian clocks in vertebrates (Haim & Zubidat, 2015;Jones et al., 2015;Lee et al., 2020).There is increasing evidence that ALAN-induced circadian disruption and suppression of melatonin secretion are directly linked to adverse health effects (Bass, 2012;LeGates et al., 2014;Panda, 2016;Sancar & Van Gelder, 2021).Therefore, meta-analytically testing hypotheses focused on melatonin can provide mechanistic insights into the harmful impacts of ALAN.
More importantly, previous studies that have examined various aspects of the effects of ALAN have primarily focused on the response magnitude to ALAN (mean differences of a trait between ALAN-exposed animals and 'dark at night' control group of animals).Few, if any, studies have statistically tested whether there is a difference in the response variability among individuals or animals between ALAN exposures (i.e., whether the impact of ALAN exposure is consistent across all individuals).Response variability, also known as inter-individual variability, refers to the phenotypic variation observed across individuals (Nakagawa & Schielzeth, 2012;Yang et al., 2023).Changes in the phenotypic variance are an important mechanism by which populations can adapt to environmental stressors.This is because natural selection acts on individual differences in traits, and an increase in variability can provide a substrate for natural selection to act upon (Bradshaw, 1965).
Therefore, an increase in inter-individual variability can enhance the probability that at least some individuals can fare better under environmental stressors (Bruijning et al., 2020;Philippi & Seger, 1989).In human studies, there is growing evidence of inter-individual differences in the extent of melatonin suppression caused by artificial light exposures (Hébert et al., 2002;Higuchi et al., 2008;Santhi et al., 2012).The biological implications and translational values of such inter-individual differences for certain populations, such as night shift workers, have been recognized recently (Chellappa, 2021).However, there is currently limited knowledge about whether ALAN exposure has consistent (low response variability) or inconsistent (high response variability) effects on the suppression of melatonin in wild populations of animals.
We compiled quantitative evidence on the effects of ALAN on melatonin secretion in wild vertebrates.Further, we exploited the development of modern metaanalytic techniques to provide quantitative insights into the effects of ALAN on melatonin suppression using a state-of-the-art statistical approach, the phylogenetically controlled multilevel meta-analytic model (Cinar et al., 2022;Yang et al., 2023).More specifically, we predict that ALAN exposures generally suppress melatonin in wild vertebrates and quantify the magnitude of the impacts of ALAN on melatonin secretion.We predict that ALAN exposures can increase the inter-individual variability of melatonin levels in wild vertebrate populations because environmental stressors are expected to increase phenotypic variation in a wild vertebrate population (Thompson et al., 2021).We also examine whether changes in magnitude and variability of melatonin levels are a function of predictor variables, such as melatonin measurement characteristics, species attributes, and specifications of artificial light used at night.

M AT ER I A L S A N D M EHOT D S
After conducting a pilot test, but prior to main data collection and modelling, we registered our hypotheses and study protocol (https:// osf.io/ nt543/ ).Given the diverse species included in our dataset and the prevalent data dependencies, we employed phylogenetic multilevel meta-analytic models as our primary statistical techniques (Cinar et al., 2022;Nakagawa & Santos, 2012).Our meta-analysis adheres to the PRISMA statement for systematic reviews and meta-analyses, along with its extension, ensuring transparent reporting of our study (Moher et al., 2009;O'Dea et al., 2021).A complete reporting checklist can be found in Data S1.The data and code used to reproduce our results are available on the GitHub repository https:// github.com/ Yefen g0920/ ALAN_ MEL, enabling the computational reproducibility of our study.

Data collection
We conducted literature searches in four leading academic databases to ensure the comprehensiveness of our searches: Web of Science, Scopus, PubMed, and EBSCOhost.We first used a benchmarking procedure to develop a finely tuned Boolean search string for the four main databases.Briefly, we manually selected 14 benchmark articles and defined the percentage of these articles captured by search string as sensitivity (Table S1).The search string was then optimized over three rounds of improvement until it achieved 100% sensitivity (i.e., all benchmark articles were captured; refer to Table S2 for the optimization process).We also collected grey literature using the same selection criteria, from the BASE database (search strings can be found at Table S1).The final search strategy for each academic database is outlined in Table S3, yielding 2844 bibliometric records on 11 March 2022.After removing duplicates through string matching algorithms in R package synthesisr (Westgate, 2019) and Rayyan's 'Detect duplicates' function (https:// www.rayyan.ai/ ), we obtained1,552 unique bibliometric records for screening.Two researchers (YY and ML) conducted systematic step-by-step screening of these unique studies in Rayyan, as shown in the PRISMA flowchart in Figure S1 (note we use the MeRIT system as per (Nakagawa et al., 2023)).
We formulated a priori eligibility criteria using the population-exposure-comparator-outcome (PECO) framework (Eriksen & Frandsen, 2018;Morgan et al., 2018).Specifically: (a) The eligible studies should be conducted on wild or semi-wild populations.(b) The study authors investigated the effects of artificial light exposure at night on wild vertebrates.(c) A control group was used that was exposed to darkness at night.(d) The study reported melatonin levels in animal tissues or tissue products.The screening and exclusion criteria were validated by randomly selecting 100 studies from the retrieved records during the pilot round (https:// osf.io/ nt543/ ).The title, keywords, abstract, full-text screening eligibility criteria, and decision trees are detailed in Figures S2 and  S3.After title, keywords, and abstract screening, 91 studies remained (with a 2.5% conflict rate between ML and YY, resolved via discussion).After the full-text screening, 38 eligible studies were included for data extraction (with a 12% conflict rate between ML and YY, resolved via discussion).The screening process is summarized in the PRISMA diagram (Figure S1), and the included studies are listed in Data S2.

Measure the magnitude and variability of melatonin response to ALAN
To quantify changes in the response magnitude of melatonin concentrations after exposure to ALAN, we used the response ratio (RR) as the effect size measure (mean ratio).To quantify changes in the response variability of melatonin, we used the coefficient of variation ratio (CVR) as the effect size measure.Both RR and CVR were logarithmically transformed to approximate normality (Hedges et al., 1999;Nakagawa et al., 2015).We defined the two effect size measures, along with their sampling variances, as follows (Hedges et al., 1999;Nakagawa et al., 2015): where x ALAN and x Dark are the (sample) mean of mela- tonin for the ALAN and dark-at-night groups, respectively; n ALAN and n Dark are the corresponding sample sizes, s ALAN and s Dark are the (sample) standard devia- tions (SDs).We used lnRR instead of Hedge's d because the former is more powerful and easier to interpret (parallel results for Hedge's d are reported in .Rmd file in GitHub repository https:// github.com/ Yefen g0920/ ALAN_ MEL; Yang et al., 2022).For example, a negative lnRR indicates that melatonin is suppressed by exposure to ALAN and represents the suppression percentage of melatonin, after back-transformation (e lnRR − 1).A positive lnCVR indicates that the variability of melatonin is increased with exposure to ALAN, which implies increased individual differences in response to ALAN exposure.Since a strong mean and variance correlation in our dataset might lead to an indirect impact of mean on variability (r > 0.9; see Figure S4), we chose lnCVR as our effect size, which is independent of any changes in mean, instead of performing direct comparisons of SDs (Nakagawa et al., 2015;Senior et al., 2020;Senn, 2016).

Data items
We followed our registered data extraction protocol for each eligible study (https:// osf.io/ nt543/ ), using a codebook that we validated using nine representative papers during a pilot stage (Data S3).A complete list of extracted variables can be found in Data S4.For each of the 38 eligible studies, the descriptive statistics of melatonin levels (i.e., x ALAN , x Dark , s ALAN , s Dark , n ALAN , and n Dark ) were retrieved either directly from the text, ta- bles, and supplementary data or digitized from figures using WebPlotDigitizer v4.6 software (https:// autom eris.io/ WebPl otDig itizer/ ).We imputed SDs for 5.9% of missing data (s ALAN and s Dark ) by leveraging the strong mean-variance correlation of studies with complete information (Lajeunesse, 2016).Our analysis confirmed that these imputed SDs did not significantly impact the model estimates (see an .Rmd file in GitHub repository https:// github.com/ Yefen g0920/ ALAN_ MEL).We also collected information on several informative predictor variables (i.e., moderators) relevant to our registered hypotheses, including characteristics of the melatonin measurement (i.e., the timing of release [daytime vs. night-time], location [central vs. peripheral]), the type of species exposed to ALAN (i.e., main activity timing [diurnal vs. nocturnal], living conditions [captive vs. free-living]), and the specifications of the artificial light sources (i.e., intensity, wavelength, colour, phase, and duration).A full list of moderator variables is presented in Data S4.Additionally, we recorded basic bibliographic data, such as title, publication year, and DOI, as well as methodological quality indicators, including study design (observation vs. experiment), melatonin measurement unit, experimental randomization, and observer blinding during measurements.

Statistical modelling
To analyse effect sizes (lnRR and lnCVR), we utilized a hybrid method known as phylogenetic multi-level meta-analytic models with (sampling) variance-covariance matrix in the framework of robust variance estimation (Cinar et al., 2022;Nakagawa & Santos, 2012;Yang et al., 2023).By using this approach, we aimed to obtain a robust estimate of the relationship between effect sizes and ALAN exposure, as described in the workflow for conducting a meta-analysis of animal models (Yang et al., 2023).This approach accounted for the three forms of statistical dependence present in our dataset: (1) multiple lnRR and lnCVR measurements per study and species, (2) phylogenetic relatedness due to shared evolutionary history, and (3) correlation among sampling errors resulting from studies with multiple light intensity and multiple Zeitgeber time measurements.The model specification included species-specific, phylogenetic, study, and observation-level random effects, as well as a sampling variance-covariance matrix (details below).Our approach was partially informed by the information-theoretic approach, which showed that multilevel models provided better fits than traditional random-effects models (Table S4).To guard against the potential model misspecification, we employed clusterrobust inference (also known as, 'robust variance estimation') to examine the significance of model coefficients and construct confidence intervals (CI) (Pustejovsky & Tipton, 2022;Yang et al., 2023), with 'CR1' as smallsample adjustment for the p-value.All models were fitted with the metafor package (Viechtbauer, 2010).We used restricted maximum likelihood (REML) as the estimator (but see Model selection and multi-model inference) and the Quasi-Newton method to maximize the likelihood function over model parameters.All models achieved convergence under default settings and the model parameters were confirmed to be identifiable by checking the profile likelihood.

Intercept-only model: Examine the overall effect of ALAN on response magnitude and variability
To estimate the overall response magnitude and variability of melatonin due to ALAN, we constructed an intercept-only model with lnRR and lnCVR as the response variable, respectively: where Y [j] = lnRR or lnCVR, depending on the hypotheses tested; 0 = intercept, quantifying the effect of ALAN on response magnitude when Y [j] = lnRR, while the effect of ALAN on the variability of melatonin when Y [j] = lnCVR; u spp[k] = species-specific random effect in k-th species, ac- counting for correlation in lnRR within the same species; u phy[k] = phylogenetic random effect, correcting for correla- tion in lnRR due to evolutionary history; u study [i] = random effect at study level for i-th study, accounting for multiple lnRR per study; u obs[j] = random effect at effect size level for j-th effect size, capturing observation-level or residual variance; e [j] = sampling variance-covariance effect, accounting for the precision of lnRR and correlation between sampling errors (Nakagawa & Santos, 2012;Yang et al., 2023).The term e [j] follows a multivariate normal distribution with mean 0 and variance-covariance V, (1) 14610248, 2024, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/ele.14387,Wiley Online Library on [27/02/2024].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License where the diagonal elements represent the sampling variance Var(lnRR) and the off-diagonal elements capture sampling covariance with a constant correlation of = 0.5 (Noble et al., 2017) (note that robust variance estimation was employed to guard against the potential mis-specified V (Yang et al., 2023); see above).All random effects follow a normal distribution with mean 0 and variance component (N(0, 2 I)), where 2 denotes either 2 spp , 2 phy , 2 study , or 2 obs estimated from the model via REML, I = identity matrix.Note that 2 phy assumes dependence among species, and thus we replaced I by phylogenetic correlation matrix (see legend of Figure 2 for details of the phylogenetic tree configuration).
We used a pluralistic approach to quantify the heterogeneity in lnRR and lnCVR (Yang et al., 2023).We first calculated the typical heterogeneity measure I 2 total , which indicates the proportion of variance due to total variance excluding sampling error variance.We decomposed I 2 total at different random-effects levels, including between-species I 2 spp , phylogenetic I 2 phy (phylogenetic signal), between-study I 2 study , and observation-level I 2 obs (residual heterogeneity; Nakagawa & Santos, 2012;Yang et al., 2023).Secondly, we calculated CV B to measure the size of total heterogeneity relative to the estimated overall effect ( 0 ) (Takkouche et al., 1999): In addition, we estimated 95% prediction intervals (PI) to quantify the range in which a new lnRR or lnCVR would fall, assuming no sampling error variance exists.

Predictor model: Moderating factors for response magnitude and variability
To assess whether lnRR or lnCVR was influenced by predefined moderator variables (predictor variables as described in Data items), we employed uni-moderator meta-regressions to model lnRR and lnCVR (predictor models): where = a vector of slopes representing the effect of ALAN on lnRR or lnCVR for each level of the moderator variables that are relevant to our hypotheses; X = a de- sign matrix of continuous or dummy-coded categorical moderator variables (Schielzeth, 2010;Yang et al., 2023); the other notations are as defined previously.We used an F-distribution-based Q M test (so-called omnibus test) to determine if a moderator variable significantly influences the size of lnRR or lnCVR.Null hypothesis significance test based on t-distribution was used to contrast the difference in slopes between different levels of a categorical moderator variable ( difference ; equivalent to contrast coding).We used multcomp package to perform this test (Hothorn et al., 2016).The marginal R 2 was used to quantify the proportion of heterogeneity explained by each moderator (Nakagawa & Schielzeth, 2013) and was calculated using the orchaRd package (Nakagawa et al., 2021).It makes sense to focus on the subset of nighttime melatonin for moderator analyses (e.g., avoiding meaningless heterogeneity), as daytime and nighttime melatonin showed opposing responses to ALAN exposure (see Results).Results for daytime melatonin are reported in the .Rmd file in GitHub repository https:// github.com/ Yefen g0920/ ALAN_ MEL for those interested in this aspect.

Model selection and multi-model inference
To test the robustness of results obtained from predictor models (moderator effects), we conducted model selection and multi-model inference (Yang et al., 2023).This was done by fitting 2 7 = 128 meta-regression models with all possible combinations of the tested moderator variables and assessing their AICc values to select the best models whose AICc values were <2 units larger than the model with the lowest AICc (440 in this case; Table S29).We also examined the importance of the moderator variables by considering all 128 models' Akaike weights (but see (Galipaud et al., 2017)).We tested the significance of slopes corresponding to different moderator variables across all 128 meta-regression models, considering their Akaike weights.The whole process was automated using the glmulti package with a custom function (Calcagno & de Mazancourt, 2010;Viechtbauer, 2010).All 128 models had the same random-effects structure as the predictor model described above and were fitted with maximum likelihood rather than REML estimation (but see (Cinar et al., 2021)).We did not examine the interaction between moderators because of no relevant registered hypotheses.

Publication bias tests
To assess the robustness of the results obtained from the intercept-only model (overall effects), we tested whether publication bias occurred in our dataset.A visual examination was conducted using a contourenhanced funnel plot with the residuals of the models including important moderator variables selected from model selection against the precision of effect sizes (Nakagawa & Santos, 2012).We used extended Egger's regression to test the symmetry of the funnel plot, where the adjusted standard error was added as a moderator (Nakagawa et al., 2022;Yang et al., 2023).In addition, we examined whether the decline effect existed in our dataset (i.e., time-lag bias) (Costello & Fox, 2022).Note that we only conducted a publication (2) bias test for the effect on the mean melatonin levels (lnRR) as variance difference (lnCVR) is not expected to be the driver of publication bias (Senior et al., 2016;Yang et al., 2022).

Sensitivity analysis and critical appraisal
We examined the robustness of the model results for lnRR and lnCVR, using leave-one-out cross-validation, where we removed one species from the dataset at a time and re-estimated the model estimates by fitting an intercept-only model to the remaining data.This was repeated for each species to evaluate the sensitivity of the model estimates to the species included in the dataset.To assess the quality of the included studies (i.e., risk of bias), we conducted a quantitative critical appraisal using a meta-regression model with qualityrelated variables as moderators, such as study design type, experimental assignment randomization, and observer blinding.The principle was to test if these methodological factors have any systematic effect on the magnitude of melatonin response to ALAN exposure.We did not perform the typical (qualitative) critical appraisal of the included studies because methodologically heterogeneous ecological and environmental studies are generally not amenable to the established risk-of-bias assessment tools.

Dataset characteristics
The dataset we compiled included 437 observations (effect sizes in this case) from 127 experiments in 38 independent studies (Figure S1).These studies were conducted on 31 different wild vertebrate species, covering a wide diversity of species (i.e., birds, mammals, reptiles, fish, and amphibians).As shown in Figure 1, the dataset was highly heterogeneous in terms of the population studied (e.g., intrinsic attributes: sex, species), the exposures to ALAN (e.g., specifications of ALAN: colour, phase), and the outcomes measured (e.g., location of melatonin measurement).
The widths of the bars in Figure 1 suggest that the sample sizes were unbalanced, with some biological and methodological variables having more effect sizes than others.Further, many of the included studies did not clearly state whether they adopted practices that guided against potential risks of bias, such as assignment randomization and observer blinding, and the specifications of the ALAN used were poorly reported (Figure S5).Most studies have focused on the effects of white light on melatonin suppression in birds and mammals, leaving a significant gap in the study of other light colours and taxonomic effects (Figure S6).

Species-specific response
For response magnitude, 28 out of 31 wild vertebrates (90%) showed a statistically significant response to ALAN exposure (Figure 3).While melatonin concentrations of 22 species were suppressed by ALAN exposure, melatonin concentrations were increased in 6 species.These 6 species with increased melatonin concentrations ranged from amphibians (Rana catesbeiana), reptiles (Anolis carolinensis), birds (Emberiza melanocephala and Perdicula asiatica), to mammals (Sigmodon hispidus and Ictidomys tridecemlineatus).For inter-individual difference in melatonin suppression, 15 species showed a statistically significant increase, while 3 species showed a decrease.Best linear unbiased prediction (BLUP) for each species based on the fitted model where phylogeny and species identities were added as random effects is reported in Figure S7.
Nocturnal melatonin was more sensitive to ALAN than diurnal melatonin ( difference  S8 and S9).However, there were no statistically significant differences in the magnitude and variability between central and peripheral.

Model selection and multi-model inference
Model selection revealed that light colour, phase, and intensity at night, location of melatonin, and living conditions and active times of species were important moderator variables (Figure 7).Multi-model inference confirmed that higher light intensity at night led to greater suppression of melatonin (Table S28).Other moderating effects were not observed as being significant when using multi-model inference.The model with the lowest AICc, including the location of melatonin, light colour, phase, and intensity at night as moderators, had a 21.7% possibility of being the best one among 128 fitted models and could explain 20.6% of the variation in the data.A similar pattern was found for melatonin variability in response to ALAN (Table S29).

Publication bias, sensitivity analysis, and critical appraisal
We found little evidence of publication bias, as indicated by the symmetrical funnel plots, the absence of larger effects in smaller studies, and the lack of temporal decrease in the magnitude of effect sizes (Figures S11 and S12).The robustness of the model estimates was further confirmed through leave-one-out cross-validation, in which the exclusion of a particular species at a time did not greatly alter the results (Figure S13).Experimental studies yielded a statistically significant overall effect and observational studies had a statistically non-significant overall effect, however no statistically significant difference between the two study types was observed ( difference[lnRR] = 0.36, SE = [−0.32,1.03], t 36 = 1.08, p = 0.289; Table S30).The incomplete reporting of other study quality variables, such as assignment randomization and observer blinding in the primary studies, rendered the corresponding quantitative risk of bias assessment infeasible (Figure S5).The model estimates were obtained using uni-moderator meta-regression (predictor models).The remaining details are the same as in Figure 2.

Cross-species-level evidence of melatonin suppression by light pollution
In light of the well-studied impact of ALAN on melatonin suppression in humans and laboratory rodents (Kozaki et al., 2015;Mclntyre et al., 1989;Reiter et al., 2007;Wright & Lack, 2001), this study aimed to quantify the impacts of ALAN on melatonin secretion in wild vertebrates and identify the key predictors.The present study, which utilized a phylogenetic multilevel meta-analytic approach to synthesize data from 31 wild vertebrates, provided the first acrossspecies-level quantitative evidence on the extent of melatonin suppression by ALAN.Exposure to ALAN significantly reduced melatonin concentrations by an average of 36% (Figure 2).The heterogeneity statistic I 2 total was found to be 94.9%, with a standard deviation of nearly two times the size compared to the overall effect (CV B = 1.8), indicating that a large heterogeneity was found in the melatonin suppression in response to ALAN (I 2 total = 94.9%).The species-specific differences only explained 3% of melatonin suppression, indicating suppression was consistent across birds, mammals, reptiles, fish, and amphibians.The phylogenetic signal might play an important role in affecting the effect of ALAN on melatonin suppression, as evolutionary history explained 17.1% of the variance.It means some taxa are more affected than others (Figure 3), especially six species being found with increased melatonin concentrations.These findings indicate that the effect of ALAN on melatonin is widespread across wild vertebrates and could have far-reaching implications for their health and well-being (Haim & Zubidat, 2015), emphasizing the need for effective management of outdoor lighting and mitigation strategies to reduce the negative impacts of ALAN on wildlife (Jones et al., 2015).

Moderators of melatonin suppression by ALAN
The high degree of heterogeneity observed in our data illustrates the necessity of using biologically and methodologically relevant characteristics to explore the moderator variables that modulate the magnitude of ALAN effects on melatonin secretion.Our analyses identified several informative moderator variables, including the timing of the release, location of melatonin,

F I G U R E 7
Relative importance of moderator variables across 2 7 = 128 candidate models.Relative importance for each moderator variable was calculated as the sum of the Akaike weights of that moderator appearing in all models.The Akaike weights represent the probability that each of the 128 candidate models is the best model, reflecting the posterior model probability given a prior distribution in a Bayesian framework (Cinar et al., 2021).The marginal R 2 , the proportion of variance explained, was estimated from the uni-moderator model with the corresponding moderator variable as the fixed effect.
active time and living conditions of species, and specifications of light sources used at night.Of these moderator variables, the results of model selection suggest that the specifications of light sources were more important than other factors (Figure 7).This highlights the possibility of designing desirable specifications of light sources used at night to minimize the disruption of circadian rhythms by ALAN (Davies & Smyth, 2018;Grubisic et al., 2019;Hölker et al., 2021).Below, we discuss how each of the examined moderator variables modulated the strength of the effect of ALAN on melatonin suppression.
Higher light intensity at night was found to associate with more severe melatonin suppression.Across 31 wild vertebrate species examined, a one-unit increase in lux on a logarithmic scale resulted in 13.1% increase in melatonin suppression, while this dose-response relationship did not exist when W/m 2 was used as the unit of measurement.This finding calls for the use of biologically relevant light measurements that incorporate spectral sensitivity of the species being studied (e.g., photometric measurements), rather than radiometric measurements that only consider the physical properties of light sources (Grubisic et al., 2019;Hölker et al., 2021;Jechow & Hölker, 2019;Tidau et al., 2021), when assessing the impact of ALAN.Importantly, the interdisciplinary nature of ALAN research requires that stakeholders, including biologists, physicists, and engineers, work closely together to design biologically meaningful and easy-to-implement devices and methods to standardize light measurements (Kalinkat et al., 2021).On average, 1 nm decrease in light wavelength on a logarithmic scale resulted in 0.3% increase in melatonin suppression, with blue colour leading to a 59% decrease in melatonin secretion, on average (Figure S8 and Tables S20-S23).These findings are in line with previous human studies, which have also demonstrated intensity-and wavelengthdependent effects on melatonin suppression (Cajochen et al., 2005;Kayumov et al., 2005;Kozaki et al., 2015;Mclntyre et al., 1989).
Notably, our results showed that the exposure duration of ALAN did not exhibit a significantly significant suppression of melatonin secretion (range from 0.5 to 450 days, mean = 38 days, median = 13 days), suggesting that long-term exposures to ALAN did not induce a larger suppression of melatonin secretion.This observation implies that wild vertebrates might have developed adaptive mechanisms to cope with chronic ALAN exposure over time, or other types of phenotypic plasticity, such as changes in physiology or behaviour, might help mitigate ALAN's impact on melatonin release (Gaston et al., 2013;Hopkins et al., 2018;Swaddle et al., 2015).We found that early-night and late-night ALAN resulted in less suppression of melatonin levels compared to midnight counterparts.The phase-dependent response of melatonin secretion provides valuable insights into the underlying mechanisms of how ALAN affects cues for the timings of biological processes (Bedrosian & Nelson, 2017;Gaston et al., 2017), where the early night may be perceived as an extension of 'subjective' dusk and late-night as 'subjective' early dawn.This interpretation is supported by the behavioural changes observed in buntings (Emberiza bruniceps), which extend their daytime activity into the early night and advance the onset of daytime activity into the late night when exposed to ALAN (Kumar et al., 2021).
Furthermore, we found that nocturnal melatonin was more sensitive to ALAN than diurnal melatonin, which is to be expected given that melatonin is secreted less during the day and more at night (Emet et al., 2016).Both the master circadian clock and the peripheral clock are disrupted by ALAN because central and peripheral melatonin levels were found to be equally affected by ALAN.Captive and free-living species also had an equivalent response to ALAN, suggesting that living conditions may not be an important predictor of the impact of ALAN.In contrast, field-caught squirrels had less melatonin suppression compared to laboratory-reared counterparts, suggesting that a stable living condition appear to buffer the effects of ALAN on melatonin suppression (Reiter et al., 1983).The observed discrepancies may stem from the diverse ecological and physiological contexts of the studied species.While Reiter et al. (1983) focused on laboratory-reared squirrels and field-caught counterparts, our study encompassed a broader range of species with varying ecological backgrounds.The impact of animal living conditions on ALAN response could be species-specific, influenced by factors such as natural adaptation to specific light environments, circadian variations, or other ecological factors that differ among captive and free-living species.Additionally, it is also possible that the methodology of the included experiments could have confounding effects on ALAN response.
Interestingly, ALAN had similar effects on diurnal and nocturnal species, which contradicts the earlier prediction that nocturnal vertebrates are more sensitive to ALAN than diurnal counterparts because the former has evolved to adapt to night-time activities (Sanders et al., 2021).Indeed, the diurnal species can be affected by ALAN through interspecific interactions (Gomes et al., 2021;Russ et al., 2015;Sanders et al., 2022).Currently, environmental protection legislation tends to focus on species with special conservation status (Hölker et al., 2021;Schroer et al., 2020), leaving other species vulnerable to stressors such as ALAN.However, our findings suggest that ALAN has the potential to affect a wide range of species, regardless of their diurnal clocks, living conditions, and natural activity patterns.Therefore, it is important to consider the effects of ALAN on a broader range of species in environmental protection legislation rather than focusing only on species with special conservation status.

Individual difference in response to ALAN
On average, exposure to ALAN resulted in a statistically significant 23% increase in response variability for melatonin suppression (Figure 2).Response variability is defined as biologically important individual differences in the physiological response to the same intervention and has important implications for stratified, precision or personalized intervention (Erden, 2015;Localio et al., 2020;Senn, 2018).The individual differences observed in the current study were attributed to increased variability (independent of mean) in melatonin levels rather than random errors (Atkinson et al., 2019;Dankel & Loenneke, 2020;Nakagawa et al., 2015).This likely reflects individual differences in light sensitivity, with some individuals in a population being more resistant to ALAN (non-responders) and others being more susceptible (responders).It is also possible that the variability in behaviour plasticity could lead to the individual differences in melatonin levels in response to ALAN.Similar findings have been observed in human studies (Mclntyre et al., 1989;Santhi et al., 2012;Spitschan & Santhi, 2022), where a >50-fold difference in light sensitivity has been found across individual humans (Phillips et al., 2019).The drivers underlying individual differences in light sensitivity are thought to be linked to individual differences in relevant traits (Chellappa, 2021;Hébert et al., 2002;Meyer et al., 2022), such as age, chronotype, genetic polymorphisms, and retinal differences caused by these factors (e.g., different pupil sizes (Gaddy et al., 1993;Higuchi et al., 2008)), albeit ceiling effects are also a possible cause (i.e., melatonin suppression reaches measured or biological limits) (Yang et al., 2021).Although a fuller understanding of the underlying mechanisms necessitates more fundamental research, our findings underscore the importance of considering individual differences in light sensitivity when designing outdoor lighting (e.g., streetlights; see more in Practical recommendations for policymaking and lighting practice), as it might explain the differential vulnerability of populations to circadian disruption and its impact on health (Phillips et al., 2019).We also found that exposure to white light at night could further increase the response variability of melatonin suppression to 28.9%.The observed increase in response variability of melatonin suppression may be attributed to the broad-spectrum nature of white light, which contains various proportions of wavelengths effective in suppressing melatonin secretion.Short wavelengths, in particular, have been found to induce the strongest melatonin suppression (Figure 6).The CCTs of the white light sources included in the dataset were variable, ranging from 2900 K to 6500 K, with an average value of 3707 K.The wide ranges in CCT reflect differences in the spectral composition of the white light, and these differences may contribute to variations in melatonin suppression responses among individuals.

Limitations and research gaps
The effect of ALAN on melatonin suppression was likely to be free of file-drawer problems and temporally stable, as indicated by our visual checks and statistical tests.Leave-one-out cross-validation suggests that no particular species biased the evidence.The experimental studies did not have a larger magnitude of melatonin suppression than observational studies, albeit the former study type usually allows more strict control of environmental variation (Yang et al., 2022).Non-blinded studies are expected to have larger effect sizes (Holman et al., 2015).However, it was infeasible to determine whether such a practice could exaggerate the evidence in our dataset due to incomplete reporting (no studies explicitly stated the blinding of observers in our dataset).The same applied to experimental randomization, as only two studies explicitly stated randomization among 38 eligible studies.Therefore, the use of reporting guidelines, such as the ARRIVE guidelines (Du Sert et al., 2020), by the authors of primary studies is necessary to allow full use of a critical appraisal tool like SYRCLE (Hooijmans et al., 2014), and to better facilitate reproducibility and high-quality evidence synthesis (Parker et al., 2016;Spitschan & Santhi, 2022;Voelkl et al., 2020).
The poor reporting practices in most of the included studies were supposed to be the major reason for these infeasible analyses, along with inconsistencies, which we observed between uni-moderator and multiple-moderator models (Figure S5 and Table S28).The current reporting practice of light conditions, such as specifications and light intensity, were poor (Figure S5), making it impossible to replicate the published studies and compare the results to with each other (Aulsebrook et al., 2022).Thus, our findings also highlight the need for placing more emphasis on elaborating on the physical parameters of ALAN to enable detailed assessments across taxonomic groups.For example, the effects of the spectral composition of ALAN light sources on melatonin suppression in reptiles and amphibians (Figure S6).Other important light parameters, such as correlated colour temperature (CCT) and flicker (Schroer et al., 2020), are also underrepresented in our data set and deserve investigation.The ecological relevance of ALAN studies is weakened if light intensity is set at unrealistically dark or bright levels.These problems are exacerbated by poor reporting of light measurements, especially since different studies use different units to measure light intensity, and different light meters have different sensitivities (Aulsebrook et al., 2022).Therefore, future studies need to transparently report light measurement methods, especially harmonizing light intensity measurement units to produce more replicable and comparable evidence and promote easier translation into lighting policy and practice (Aulsebrook et al., 2022).We note that animal-centric light units can also be achieved by using sophisticated light measurement methods such as hyper-colorimetric multispectral imaging (Colantonio et al., 2018;Tidau et al., 2021).We also advocate for the adoption of standardized methodologies for measuring melatonin in future ALAN research to facilitate more meaningful comparisons of the evidence regarding ALAN's impacts.Additionally, the sample sizes in our dataset were unbalanced, with the diurnal species having less sample sizes than the nocturnal species (Figure 1).The unbalanced sample sizes might affect the overall generality of the harmful effect of ALAN, highlighting the need for investigating more impacts of ALAN on diurnal species.

Practical recommendations for policymaking and lighting practice
Turning off artificial lighting at night is the most straightforward strategy to eliminate the negative effects ALAN (Smith, 2009).However, the use of artificial lighting at night is essential for human societies, contributing to our safety and activities.Previous attempts to develop outdoor lighting policies mitigating the impacts of ALAN appear to have been ineffective due to complex considerations of safety, economic, political, and cultural needs, as well as health, environmental, and ecological issues (Barentine, 2020;Hölker et al., 2010Hölker et al., , 2021;;Pérez Vega et al., 2021).Our findings provided the first quantitative and across-specieslevel evidence of ALAN's effect on the magnitude of melatonin suppression.Such findings can inform easily enforceable, inexpensive, wildlife-friendly, and evidence-based outdoor light policies.Among all possible recommendations (Barentine, 2020;Grubisic et al., 2019;Hölker et al., 2010;Pérez Vega et al., 2021), our quantitative evidence strongly supports three.First, we can tune the spectral composition of artificial light sources used at night to minimize the proportion of blue light (Gaston & Sánchez de Miguel, 2022), because our meta-analytic finding confirmed that short wavelengths (blue portion) suppress melatonin levels more than other wavelengths across species.Second, we can adopt a discontinuous lighting strategy (Kumar et al., 2021), such as using motion sensors to turn lights on or higher only when necessary and avoiding artificial lighting during midnight hours.Third, we should reduce the light intensity of all types of artificial light to mitigate the impact on melatonin suppression (Agathokleous, 2023).The light intensity should be minimized as low as possible, while ensuring that the ALAN still serves a safety purpose, as even a very low level of light night can suppress melatonin secretions (Aulsebrook et al., 2022).Implementing these interventions is realistic and readily achievable, given that more and more streetlights are expected to be replaced by Light Emitting Diodes (LEDs), as LEDs are well suited for 'smart' lighting systems (Gaston & Sánchez de Miguel, 2022).We believe that these recommendations are practical and readily implementable.We hope many decision-makers or stakeholders consider adopting our recommendations, minimizing the negative impact of ALAN on wildlife, whose coexistence with humanity is becoming increasingly challenging.

AC K NOW L E DGE M E N T S
YY and JP were supported by the National Natural Science Foundation of China (32102597) and the China Agriculture Research System (CARS-40).SN, ML, and YY were supported by Australian Research Council Discovery Grant (DP210100812).The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.Open access publishing facilitated by University of New South Wales, as part of the Wiley -University of New South Wales agreement via the Council of Australian University Librarians.

CON F L IC T OF I N T E R E ST STAT E M E N T
We declare no conflict of interest.

PE E R R EV I EW
The peer review history for this article is available at https:// www.webof scien ce.com/ api/ gatew ay/ wos/ peer-review/ 10. 1111/ ele.14387 .

F
Visual representation of the characteristics of the primary studies included in the dataset.The vertical bars indicate key categorical variables.The widths of the vertical bars indicate the numbers of effect sizes represented by each level of the categorical variable.The flow lines between vertical bars indicate the connections and overlaps of different levels of the categorical variables.For descriptions of the individual categorical variables, refer to codebook / meta-data (Data S3).F I G U R E 2 The effect of artificial light at night (ALAN) on melatonin suppression in wild vertebrates.(a) Model estimate (white diamond) for the average degree of melatonin suppression, quantified as the difference in means between the ALAN and dark-at-night groups on a logarithmic scale, with negative values indicating melatonin suppression in the ALAN group.(b) Model estimate (white diamond) for the average individual difference in melatonin suppression, quantified as the difference in variances between the ALAN and dark-at-night groups on a logarithmic scale, with positive values indicating increased inter-individual variability of melatonin suppression in the ALAN group.Shorter whiskers represent 95% Confidence Intervals.Longer lines without whiskers represent 95% Prediction Intervals.N obs is the number of effect sizes used in a model.N species is the number of species included in the model.Each circle is an effect size, and its size is scaled by the precision (inverse sampling error variance) of that estimate.were observed in the response of melatonin suppression between diurnal and nocturnal species ( difference[lnRR] = −0.29,CI = [−1.13,0.55], t 36 = −0.70,p = 0.486; difference[lnCVR] = −0.01,SE = [−0.47,0.46], t 36 = −0.04,p = 0.968) or captive and free-living species ( difference[lnRR] = −0.24,CI = [−0.61,0.14], t 36 = −1.28,p = 0.208; difference[lnCVR] = −0.15,CI = [−0.48,0.18], t 36 = −0.95,p = 0.349).

F
The moderating effect of melatonin (Mel) measurements characteristics: measurement timing (Daytime, Nighttime) and location (Peripheral, Central) on: (a) Response magnitude of melatonin suppression, and (b) Response variability of melatonin suppression.

F
The moderating effect of species attributes: species' activity time (Nocturnal, Diurnal) and living conditions (Free-living, Captive) on: (a) response magnitude of melatonin suppression, and (b) response variability of melatonin suppression.The model estimates were obtained using uni-moderator meta-regression (predictor models).The remaining details are the same as in Figure2.14610248, 2024, 2, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/ele.14387,Wiley Online Library on [27/02/2024].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons LicenseF I G U R E 6 The moderating effect of intensity and wavelength of artificial light at night (ALAN).(a) Light intensity and the magnitude of melatonin suppression.(b) Light intensity-dependent and the variability of melatonin suppression.(c) Light wavelength and the magnitude of melatonin suppression.(d) Light wavelength and the variability of melatonin suppression.The remaining details are the same as in Figure 2.