Data collection
We collected data for this study from the literature during the last 15 years of our work on avian growth [14, 17, 38, 39], incubation [37, 40, 41], and nest predation [42, 43]. Our data collection protocols include systematic searches of handbooks and journals, database searches, and searching reference lists of relevant articles.
From the original studies, we extracted information on the location of the study. We used this information to obtain the latitude and longitude where the study was conducted using Google Earth. We then extracted primary data on the growth of body mass, wing length, tarsus length, and tail length when the young were in the nest, i.e. up to fledging. This data came from tables and figures in the original articles. If possible, we also noted the number of nests and/or nestlings weighed and measured. We also extracted fledging age, body mass at hatching, and body mass, wing length, tarsus length, and tail length at the day of fledging. Nestling mass must have been weighed within 10% of fledging age to be included as body mass at hatching. For example, in species staying for 20 days in the nest, the hatching mass must have been weighed between days 0 and 2. We thus obtained data on growth in 456 populations of 336 passerine species. Average number of populations per species was 1.4 (median = 1, range = 1–7). Of course, not all source studies provided data on all traits and thus sample sizes differed across analyses. The numbers of studies providing data on particular characteristics can be obtained from our data set (Additional File 3), while sample sizes for particular analyses can be obtained from supplementary tables (Additional File 1).
We then used ornithological handbooks (see Table S1 in Additional File 1) to find information on relevant predictors and covariates at the species level. These included adult values of body mass, wing length, tarsus length, and tail length. If these data were given separately for males and females, we used their arithmetic average, because the sex of nestlings is almost never known, which precludes sex-specific analyses of growth. We further obtained data on clutch size (number of eggs in a complete clutch), nest height (in meters), the number of care givers during the nestling period (female-only, breeding pair, cooperation of more than two individuals), and the substrate of food collection (ground, vegetation, and air). For justification of these predictors and covariates, see Introduction and for associated predictions, see Table 1. Lastly, we searched primary literature for data on nest predation rates in our sample of species. We succeeded in finding nest predation information as the percentage of depredated nests (with the minimum sample size of 10 nests) in 187 species. We then converted these percentages to daily nest predation rates (i.e., probability of nest depredation per 1 day [43]). Please note that nest predation data came from different populations than growth data. This might have introduced noise into our analyses.
Data processing
We calculated three characteristics of growth rate in nestling passerines at the population level. First, we calculated the ratio of the population-specific trait value at fledging with adult values. If available, we used population-specific adult values, otherwise we used species-specific adult values. We call these ratios relative fledging mass, relative fledging wing length, relative fledging tarsus length, and relative fledging tail length, respectively. They are measures of the relative development of the young when leaving the nest. Relative fledging mass and relative fledging wing length are good predictors of post-fledging flight performance [19] and survival of fledglings in passerines [8, 21].
Second, we estimated two indices of growth rate in the nest. We estimated peak growth rates using the parameter K of a sigmoid growth function, and we did this for body mass, wing length, and tarsus length (tail length was not fitted due to problems with convergence in too many populations). We estimated peak growth rate also when body mass of nestlings was cut at 70% of adult mass, which might improve estimates of growth rates [14]. We could not use the 70% cut-off for other traits due to lack of convergence in most cases. The parameter K is independent of overall body mass and thus is a convenient, scale-free measure of peak growth rate. There are several sigmoid functions used to estimate growth rates in animals (Fig. S1 in Additional File 2), but the logistic function is overwhelmingly used in studies of growth in small birds [14,15,16, 23]. It is a three-parameter model where the nonlinear fitting procedure estimates peak growth rate (K), asymptote of the sigmoid curve (A), and the age of inflection (ti, the point in time where the growth trajectory changes from accelerating to decelerating [14]).
However, the main caveat of relying on a single three-parameter model is that the model may not be flexible enough to return accurate parameter values [44]. We thus also fitted the four-parameter Richards growth function that besides the three above-mentioned parameters estimates also the shape parameter d, which flexibly places the inflection point on the trait (i.e., vertical) axis between 0 and the estimated asymptotic value A [45]. On the contrary, the logistic function fixes the position of d at 50% of A [14]. We thus used both the traditionally used logistic function and more flexible Richards function as recommended by a recent modelling study of passerine growth [46]. We used Unified versions of both sigmoid functions (U-logistic and U-Richards), which ensures comparability of K estimates across different growth functions [46, 47]. One of the potential problems identified with fitting sigmoid growth functions to growth data can be poor estimation of the upper asymptote A, which might bias estimates of K [14, 48]. We thus present evidence that our estimates of A were close to adult trait values across our sample of species (mean correlation coefficient was r = 0.86, n = 8 values; Figs. S2 and S3 in Additional File 2). We also show that fitting of growth curves with the asymptote being estimated vs. fixed at adult value give highly positively correlated estimates of peak growth rates (mean correlation coefficient was r = 0.74, n = 6 values; Table S2 in Additional File 1).
Further, we calculated average growth rate of body mass, wing length, tarsus length, and tail length. We calculated it as log(trait at fledging)/development time. It is sometimes called “relative growth rate” (in g g− 1 day− 1) and technically it is the derivative of log(trait) over time [49, 50]. Its calculation brings two problems. First, it assumes exponential growth that is rare in animals [51]. Passerine growth in the nest is typically sigmoidal and decelerating on the log-linear scale (Fig. S4 in Additional File 2), and thus instantaneous relative growth rate also decreases with time. Second, log(trait at fledging)/time assumes the same starting trait value at hatching. A better estimate would be log(trait at fledging/trait at hatching)/time, as was argued also for seed size in plants: log(plant size/seed size)/time [52, 53]. Nevertheless, estimates of hatchling size are rare due to difficulties of measuring tiny hatchlings, with the exception of body mass. We thus show that i) estimates of average mass growth rates obtained using the two methods are highly positively correlated (r = 0.81; Fig. S5 in Additional File 2), and ii) results obtained with estimates of average growth rate calculated without hatchling mass are similar to those obtained with hatchling mass (Table S3 in Additional File 1). Despite all the caveats mentioned above, we used average growth rates in addition to peak growth rates, because the latter do not express overall growth achieved in the nest and average growth rates can thus bring additional insights into the evolution of growth patterns (see below). It is also important to realize that average growth rates are the same irrespective of a particular growth trajectory in the nest, and thus their estimation can be more robust than estimation of peak growth rates, the latter coming with their own errors and biases [14, 48].
Statistical analyses
We used phylogeny-based comparative methods to test our hypotheses. We downloaded 500 phylogenetic trees for our species from a publicly available archive at birdtree.org using Hackett constraint, all species, and version 2 (V2) of the archive [54]. We calculated one Bayesian maximum credibility tree using TreeAnnotator [55] and used this tree in our comparative analyses. We analyzed our data at the species level using the mean values of response variables calculated for each species across populations available for that particular species. To obtain mean values across populations, we first calculated point estimates of a given variable (e.g. wing growth rate) for individual populations and then took arithmetic means of these point estimates across all available populations for a given species. We used phylogenetic generalized least squares, PGLS [56], to fit our multiple regression models on species means in the ape [57] and caper packages [58] for R 3.5.1 software. We checked that residuals from these models were normally distributed and homoscedastic and that there were no non-linearities [59].
We selected only populations where the sample size was known and at least five nestlings were measured. We then checked that this procedure increased the quality and reliability of our data by calculating within-species repeatability of population estimates as the intraclass correlation coefficient using the ICC package for R software [60]. Indeed, repeatability of our dependent variables increased on average from 0.36 to 0.57 in peak growth rate K (Fig. S6 in Additional File 2) and from 0.87 to 0.91 in fledging values of tarsus, wing and tail length, body mass, and brood age (Fig. S7 in Additional File 2). Although repeatability of population-level estimates of our dependent variables was moderate to high, remaining within-species variation might still affect parameter estimates. Moreover, phylogenetic uncertainty could also affect parameter estimates. We dealt with both these problems by fitting a subset of our models also on population-level data across several phylogenies using phylogenetic mixed models implemented in the MCMCglmm package [61] as in our previous work [37]. However, results were similar to those obtained using species-level PGLS and we thus report only PGLS results and provide all data needed to replicate our analyses or use other fitting methods.
We modelled three types of response variables: i) fledgling traits, ii) fledging age, and iii) growth rates. Fledgling trait is the value of a given trait (body mass, tarsus, wing, and tail length) at the time of fledging, and we used either absolute trait value or relative trait value (see above). Fledging time is the age of nestlings at the time of fledging (in days). Growth rates are either average growth rates or peak growth rates estimated by parameter K of a sigmoid growth function (either Logistic or Richards). Peak growth rates were modelled either as “absolute” peak growth rate, where we put the adult value of a given trait among predictors, or as “relative” peak growth rate, where we put growth rate of body mass among predictors. Thus, in the first case, it was a classic allometric adjustment, while in the second case we modelled trait growth rate (tarsus and wing length) relative to body mass growth rate.
We scaled all continuous variables (subtracted mean and divided by one standard deviation) to obtain parameter estimates comparable across variables and models. However, parameter estimates for factors (the number of care givers and substrate of food collection) were still not comparable to other estimates [62]. Many variables were skewed and thus we used log10 or square root transformations to improve their distribution. As dependent variables, we used log10(fledging trait), and log10(fledging time), while growth rates remained untransformed. As predictor variables, we used log10(daily nest predation rate + 0.01), square root(nest height), square root(clutch size), and log10(adult trait), while other predictors remained untransformed. These transformations are also noted in tables reporting results in the Additional File 1.
Clutch size was strongly correlated with absolute latitude (r = 0.66, n = 453 at the population level; r = 0.69, n = 333 at the species level for all species; r = 0.68, n = 229 at the species level for species entering analyses; clutch size log10-transformed). To avoid collinearity of predictors, we fit two sets of models. One set with latitude without clutch size and the other with both latitude and clutch size included. When studying geographic effects of latitude, we used three strategies. First, we fitted absolute latitude as a predictor, because many life-history traits change systematically with increasing distance from the equator in birds [15, 16]. Second, we also fitted an interaction of absolute latitude with hemisphere (northern vs. southern) to find out whether the slope of latitudinal effect differed between the hemispheres. This interaction was, however, never statistically significant and thus we omitted it from all models. Third, tropical, southern temperate, and northern temperate birds (delimited by 23.5°N and 23.5°S) often differ in their life histories and behavior [8, 36, 37]. We thus also fitted latitudinal band (northern temperate, tropical, southern temperate) as a predictor in our models that excluded absolute latitude.
We obtained data on nestling growth and development from 456 populations of 336 species of passerines worldwide (dataset available in Additional File 3). However, in 84 populations the number of measured nestlings was not known, while in 77 populations it was lower than five. Thus, we ended up with 295 populations of 231 species where the number of measured nestlings was known and was at least five (Fig. S8 in Additional File 2). The most limiting predictor was nest predation rate, which was available for only 152 out of these 231 species and was not available in the remaining 79 species. Thus, our final sample size that was used in all analyses was 152 species, but it was usually lower due to lacking other variables (mainly growth rates and fledging values of individual traits). Scaling analyses that did not include nest predation were the only exception, because there we were able to use up to 230 species.