We followed the PRISMA protocol for collecting data from the published literature (Moher et al., 2009; Nakagawa and Poulin, 2012). We first performed literature searches using the online search engines Scopus (http://www.scopus.com/) and Web of Science (http://wokinfo.com/) with the following keyword string: “nest^* AND (^*urban^* OR anthropogenic OR rural) AND predat^*.” We then screened the titles and abstracts of the papers found by the search, and decided whether the following criteria were met:

The abstract indicated that study was conducted on bird nests, either by monitoring natural bird populations, or by performing experimental studies on artificial nests baited with real or fake eggs.

One of us (EV) screened all the abstracts found by Scopus, while another (GS) screened all the abstracts found by searching the Web of Science (WoS), and coded whether the papers met the selection criteria or not. Papers that were found by both search engines were scored by both observers and used for testing between-observer repeatability (see Table S1). We also performed backward searches, i.e., screened reference lists of relevant papers and reviews, and visited author websites, to find additional eligible papers that might have been missed during the systematic database searches. We did not limit our search to English-language papers, but also included papers written in Spanish, French, German, Russian, or Polish.

Papers considered to meet the selection criteria by at least one observer were taken forward for full-text screening. Each paper was screened by a single person. During full-text screening we excluded the papers that:

Did not define an urbanization gradient (i.e., because all studied sites had the same level of urbanization, or because the information was not provided);

In the latter two cases, we contacted the original authors provided that their e-mail addresses were available. We got feedback from 19 out of 28 contacted authors. For four studies, we obtained raw data from the authors from which we managed to extract effect sizes ourselves; we also included these papers in the analysis.

Studies varied widely in methods used to quantify the levels of urbanization of their study sites. To make studies comparable, two observers (EV and GS) independently scored the level of urbanization of each study on a five-level urbanization scale (using a modified version of the scale proposed by Marzluff et al., 2001a):

Wildland/natural area: Interior of a large (>200 ha) forest or meadow, with little anthropogenic effects.

Information on numbers of sites scored by the two observers, and on between-observer repeatability, is provided in Table S2. Sites that got different scores from the two observers were assigned one of the five scores via subsequent discussion. In most studies (N = 40; 78%), the order of the study sites based on these urbanization scores matched the order indicated by the original authors. However, for some papers (N = 11), sites that represented different urbanization levels according to the authors were given equal scores by us.

We used correlation coefficients between nest survival (i.e., the probability of nests not failing) and urbanization as our effect size. We calculated these correlation coefficients from the numbers of surviving and failed nests at each study site unless already provided in the paper. The numbers of surviving and failed nests were obtained in several ways:

The exact numbers of surviving and failed nests in an experiment were directly reported (N = 22) or given to us by the original authors (N = 4);

There was a single paper that reported only a correlation coefficient (Lumpkin et al., 2012), which we used as our effect size. For all other papers we coded each nest as failed (0) or survived (1), ranked urbanization around the nest according to the five levels on an ordinal scale from the least to the most urbanized based on the urbanization scores assigned by us, and ran a Spearman rank correlation between these two variables for each study. We used the Fisher's Z-transformed correlation (Zr) value as our effect size (see Section Statistical Analyses below for more explanation for this choice). We excluded studies where only test statistics from regression models other than correlation tests were available for which we did not obtain data from the authors (N = 18, Table S3). We excluded these studies because the conversion of multivariate regression estimates to correlation coefficients often does not accurately approximate correlation coefficients when calculated from raw data (Aloe, 2015). Sampling error variance of each effect size (Fisher's Zr) was calculated as the reciprocal of the total number of nests minus three.

Some papers (N = 16 natural nest studies and N = 20 artificial nest studies) distinguished between predation and other forms of nest failure (e.g., abandonment, weather destroying the nest, vandalism by humans). Since we were primarily interested in predation, whenever it was possible, we omitted nests that failed due to causes other than predation, and used only nests that survived and nests that failed due to predation. For papers that did not distinguish between sources of nest failure (N = 15 studies, all on natural nests), we simply used numbers of survived and failed nests to calculate effect sizes, regardless of the actual source of nest failure. Similarly, partially predated nests (i.e., where some eggs or offspring disappeared, but at least one of them survived) were considered as predated in some studies (N = 19) and as survived in others (N = 28), while some studies were conducted on nests with only one egg (N = 4). To test whether these two methodological differences affected our effect sizes, we applied the moderator “source of mortality” as a binary factor to describe whether our effect size was calculated from predation-only mortality or from total mortality, and the moderator “partial predation,” also as a binary factor, to code whether failed nests included the partially predated nests or not.

For studies spanning multiple years, we extracted separate effect sizes for each year, with “study year” as a moderator. However, in some papers only pooled data from multiple years were available. In these cases we used the median value of all study years as a measure of “study year.” We also recorded the following moderators: whether the study reported observational data collected in a natural population (N = 28) vs. experimental data using artificial nests (N = 20); the number of eggs per nest (average number for natural nests, exact number for artificial nests); nest position (ground nest or elevated, e.g., on a tree or building); nest height above ground in meters; nest openness (cup nest—a hemispheric, open nest on the ground or a branch; orb nest—a spherical nest closed on top; or cavity nest—one inside a den or a box); the number of days the nest was exposed to predators (the length of the experiment for artificial nest, the length of the nesting cycle for natural nests); and, in case of natural nests, the study species. We also recorded the year and journal in which the paper was published.

Before conducting analyses, we converted correlation coefficients r (Spearman's r) into a standard normal metric (using Fisher's z-transformation), Zr. This was done because r is bounded between –1 and 1, and thus does not follow a normal distribution (Hedges and Olkin, 1985). All analyses were performed using the transformed values (Zr), but the results were transformed back to correlations (r) for visual presentations. Correlational effect of r = 0.1, 0.3, and 0.5, can be considered as small, medium and large effect sizes, respectively, following Cohen's (1988) tentative benchmarks. Point estimates from statistical models were considered significantly different from zero when their 95% Confidence Intervals (CI) did not overlap zero.

We performed all statistical analyses within R statistical software v3.2.4 (R Core Team, 2016). For meta-analysis and meta-regression we used the metafor package (Viechtbauer, 2010). The phylogenetic tree for the bird species included in the dataset was created based on the global phylogeny of birds (Jetz et al., 2012) and visualized using the ape package (Paradis et al., 2004). We used multilevel meta-analyses, representing a type of linear mixed model (Viechtbauer, 2010; Nakagawa and Santos, 2012), to control for various sources of non-independence in the data. Non-independence can be present when multiple effect sizes are extracted from the same study, multiple effect sizes are available for the same species, and/or due to phylogenetic relationships among species. To take such non-independences into account, we included random intercepts for study identity, species identity (for natural nests), and phylogeny (for natural nests; using phylogenetic meta-analysis, Hadfield and Nakagawa, 2010; Nakagawa and Santos, 2012) in our models.

We first ran a simple meta-analytical model (an intercept-only model without any moderators) on all available data to examine the overall effect, which tested whether the meta-analytic mean was different from zero. This meta-analytic mean represents the overall relationship between urbanization level and nest predation rate over all data. We quantified total heterogeneity in the dataset by computing I² statistic (Higgins et al., 2003; Supplementary Methods S2.1). We also assessed separate meta-analytical models (and quantified heterogeneity) for the two data subsets composed of effect sizes from studies using either artificial or natural nests, respectively. To interpret the difference between the predation rates in differently urbanized habitats, we converted our meta-analytical means from correlation coefficients to odds ratios (OR; Borenstein et al., 2009); both numbers are reported in the Section Results.

We conducted meta-regression analyses to identify variables accounting for heterogeneity across studies. Continuous moderators were standardized prior to the analyses, so that they had a mean of 0 and SD of 1. Nest height above ground was log-transformed due to skewedness. We first quantified whether the relationship between urbanization and predation differed on average between the two major study approaches, i.e., artificial vs. natural nest studies. To do so, we added nest type (artificial vs. natural) as a moderator. Because the data subsets for artificial and natural nests had different combinations of applicable moderators, we then performed separate meta-regression analyses on these two data subsets. We did not construct models with multiple moderators, because the sample sizes of the subsets were inadequate and/or information on key moderators was missing for many data points. For artificial nests we considered the following moderators: nest openness (cup/hole), nest position (elevated/ground/mix), average egg number per nest, study duration, median study year, study publication year, minimum (min) urbanization score of the site gradient (1/2/3/4), maximum (max) urbanization score of the site gradient (3/4/5). For the natural nests, we considered the following moderators: source of mortality (i.e., whether failures due to other causes than predation were excluded from calculating the effect sizes or not: yes/no), nest openness (cup/hole/orb), nest position (elevated/ground), nest height above ground in meters, average egg number per nest, study duration, median study year, study publication year, minimum urbanization score of the site gradient (1/2/4), maximum urbanization score of the site gradient (3/4/5).

To test whether our results were sensitive to the way we scored urbanization, we repeated all analyses with effect sizes calculated from alternative urbanization scores based on the authors' scores.

We also assessed evidence for publication bias (Supplementary Methods S2.2), which can affect conclusions of a meta-analytic studies when published studies are biased toward significant findings (Rothstein et al., 2005).

We screened the abstracts of 412 papers, 138 of which were taken forward for full-text screening (Figure S1). We excluded 87 papers based on our inclusion criteria (fully detailed in Table S3). Our final dataset thus comprised 117 effect sizes from 51 papers published between 1985 and 2015 (Table 1). The median number of nests monitored within a publication was 104 (mean = 177.6, SD = 246.5); only in eight cases were effect sizes calculated based on sample sizes of more than 500 nests. Almost half of the effect sizes came from studies that scored mortality in natural nests (58, vs. 59 effect sizes for studies using artificial nests). Effect sizes from natural nest observations represented 32 different species from 21 Families within 6 Orders (with 25 species from 16 families belonging to the Order Passeriformes). Most species were represented in our meta-analysis by only one study; only data for the House wren Troglodytes aedon, European magpie Pica pica and Common blackbird Turdus merula were available from 2, 3, and 3 studies, respectively (Table 1). All continents (except for Antarctica) were represented in the meta-analysis, with most data from North America (46 effect sizes from 23 studies) and Europe (46 effect sizes from 15 studies; Table 1).

Studies using artificial nests differed from those with natural nests in several aspects of their study design (Table S4). First, artificial (vs. natural) nests were, on average, followed for much shorter periods of time (Mean ± SD = 12.2 ± 5.9 vs. 41.2 ± 21.1 days). Second, studies using artificial nests were more often performed in less urbanized locations. Third, artificial nests were located on the ground in about half of the studies (28 out of 59 effect sizes), whereas natural nests were usually located at least 2 m above ground level, with only 6 out of 58 effect sizes coming from studies on ground nests. Fourth, artificial nests were usually open, i.e., cup-shaped (90%; 53 out of 59 effect sizes) and 10% were hole-like (nesting box/cavity). In contrast, natural nests were open in 62% of our data points (36 out of 58 effect sizes), and 34.5% were hole-like. Fifth, failures of artificial nests were always assumed to result from predation (100%), whereas most studies on natural nests reported overall survival rates that did not distinguish between predation and other sources of mortality (66%; 38 out of 58 effect sizes). Finally, all artificial nests were considered as predated when at least one egg/offspring died, while in natural nests usually only complete brood loss was counted as a predation event (97%; 56 out of 58 effect sizes).

We found no overall relationship between nest survival and level of urbanization: the meta-analytic mean for the slope between urbanization gradient and nest survival was not distinguishable from, and was centered around, zero [meta-analytical mean (β) = −0.003; 95% Confidence Intervals (CI) = −0.080 to 0.074; OR = 0.99; Figure 1, Table S5]. Effect sizes were, at the same time, highly heterogeneous (I² = 92.7%). High heterogeneity suggests effects of moderators, and this finding thereby justified our subsequent meta-regression analyses (Higgins and Thompson, 2002).

When meta-analytic means were estimated separately for the studies using artificial vs. natural nests, we found weak evidence for an effect of urbanization on nest survival for both cases. The meta-analytic mean for artificial studies was small and negative, with confidence intervals marginally overlapping with zero, suggesting a decreasing nest survival with increasing urbanization (β = −0.118; 95% CI = −0.238 to 0.006; OR = 0.65; Figure 1; Table S5). In contrast, the meta-analytic mean for natural nest studies was small and positive; again, confidence intervals were marginally overlapping zero, suggesting that nest survival increased (instead of decreased, see above) with urbanization (β = 0.079; 95% CI = −0.007 to 0.165; OR = 1.33). The latter effect was lessened when phylogeny was taken into account (β = 0.034; 95% CI = −0.163 to 0.228, Figure 2, Table S7; we note that phylogenetic effects were not applicable for artificial nest studies, as those did not have associated species). Total heterogeneity was above 90% in all meta-analytic models (Tables S5–S7).

When we included nest type (artificial vs. natural nests) as a moderator in the meta-regression model on the full data set, the difference between average effect sizes for artificial and natural nests was small but statistically significant (difference between β artificial and β natural = 0.195; 95% CI = 0.050 to 0.332; Figure 1; Table S5), implying that the slopes of the urbanization-mortality relationship differ between the two nest types. Moderators considered affected the relationship between urbanization and nest predation neither in artificial nest (Figure 2A; Table S6) nor in natural nest studies (Figure 2B; Table S7). When we used species identity as a moderator, we found high variation among species (Figure S2, Table S8); note, however, that most species were represented with only one study in the meta-analysis, making further interpretations of this latter result difficult.

All analyses using alternative urbanization scores (based on the quantification of urbanization levels by the original study authors) had results qualitatively identical to these from the main analyses (Tables S9–S12), indicating that our results are robust to the way urbanization was scored.

We found no evidence for publication bias. Visual inspection of funnel plots and Egger's regression test revealed no indication of funnel shape asymmetry [Figure 3; t₍₁₁₅₎ = 0.716, p = 0.476]. A trim-and-fill method also implied that there were no missing effect sizes, consistent with the absence of publication bias.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.