collection of avian serum
Nine hundred and three individuals of 99 bird species (see Appendix S1 in Supporting Information for full list) were captured for this study. Blood samples were taken within 5 min of capture (to avoid effects of stress) through standard wing venipuncture and collected into un-heparinized microcapillary tubes. Samples were centrifuged in a Zip-spin portable centrifuge and serum was removed and kept on ice for 1–6 h. until it could be frozen at –80 °C.
Ninety-two of our study species were small forest and edge birds, mostly passerine, caught in mist nets in Panama and Michigan. Birds from these species were netted at several locations in and around Gamboa, Panama, in March 2004 and March 2005, and at Kellogg Biological Station near Kalamazoo, Michigan, in June and July 2004, and July 2005. The additional seven species include Leach's storm-petrels (Oceanodroma leucorhoa), savannah sparrows (Passerculus sandwichensis) and tree swallows (Tachycineta bicolor) sampled on Kent Island, New Brunswick, Canada (44°35′ N, 66°46′ W) between 18 and 25 June 2005; Florida scrub-jays (Aphelocoma coerulescens) caught at Archbold Biological Station, Lake Placid, FL throughout 2005, but mostly in January and February; and house sparrows caught in Princeton, NJ from 1 to 5 September 2005. Waved albatross (Diomedea irrorata; caught 8–10 May 2002) and nazca booby (Sula granti; caught 16–23 August 2003) samples were provided by Dave Anderson and Victor Apanius from their studies in the Galápagos Islands, and were sampled following our protocol of taking serum immediately upon capture, centrifuging and freezing. In addition, one blue jay (Cyanocitta cristata), two northern cardinals (Cardinalis cardinalis), two eastern towhees (Pipilo erythrophthalmus) and 11 gray catbirds were caught in Princeton with the house sparrows, though these species are represented in greater number in the sampling from Michigan.
Most species were sampled during the breeding season. For species in Panama, breeding season is generally more diffuse than in temperate species, and March is generally before peak breeding. For species in Michigan, breeding generally tails off by the end of July. Only nazca boobies were sampled entirely outside the breeding season, though most Florida scrub-jays and birds caught in Princeton, NJ were not breeding. Breeding and non-breeding Florida scrub-jays did not significantly differ in levels of any antioxidants (data not shown).
We measured TEAC and uric acid using spectrophotometric methods, and vitamin E and carotenoid levels using High Performance Liquid Chromatography (McGraw & Parker 2006; Cohen et al. 2007), though not all measures were available for all individuals. TEAC reflects levels of circulating micromolecular antioxidants including uric acid, vitamin C, vitamin E and carotenoids, but does not reflect levels of enzymatic antioxidants or other macromolecules with antioxidant properties. See the Methodological Appendix S4 in Supporting Information for details.
TEAC and concentrations of uric acid, vitamin E and all individual carotenoids were log-transformed for normality. TEAC–uric acid residuals were calculated following Cohen et al. (2007). This residual indicates non-uric acid antioxidant capacity. Four carotenoid types were present in enough species to be considered individually in our analyses: lutein, zeaxanthin, β-cryptoxanthin and β-carotene. Additionally, canthaxanthin was used in the intraspecific analysis of northern cardinals, canary xanthophylls in the analysis of cedar waxwings (Bombycilla cedrorum) and α-cryptoxanthin in the analysis of savannah sparrows and tree swallows. Total carotenoid concentrations were calculated, but are not presented here because lutein accounts for most of the variation and thus there is no additional information gained by including this variable. Carotenoid number is the number of types of carotenoids detected in an individual or species.
We tested the dimensionality of relationships among the nine antioxidant variables by first using a principal components analysis (PCA, R v2·5·0, princomp function). We then used a factor analysis (proc factor, SAS, v9·1, SAS Institute, Carey, NC) to confirm these results and subsequently to generate individual-specific scores for a carotenoid factor (CarFac: factor loadings: lutein = 0·76; zeaxanthin = 0·69; β-cryptoxanthin = 0·77; β-carotene = 0·79; carotenoid number = 0·95). Antioxidant measurements for all individuals are provided in the Appendix S2 in Supporting Information.
We calculated Pearson correlation coefficients (SAS, proc corr) at the interspecific level with species average values. Statistically, such correlations should be weighted by sample size, but because sample size is not random with respect to species characteristics (e.g. species with large sample sizes tended to be temperate and omnivorous) we present results from both weighted and unweighted analyses. For each variable, portions of the variance at the individual and species levels and partial correlation coefficients were calculated using a nested anova (SAS, proc nested). Differences among species in each antioxidant variable were tested using a general linear model with random effects (SAS, proc glm, random statement). Pearson correlations were also used for intraspecific analyses on 30 species, but sample sizes were too small for robust PCA or factor analyses. Phylogenetic independent contrasts were run on interspecific analyses as described in the Methodological Appendix S4.
We assessed whether correlations among levels of antioxidants were heterogeneous across species using multilevel random effects models. We chose a subset of six correlations to focus on as representative of the larger set of 45: TEAC–UA, lutein–zeaxanthin, TEAC–vitamin E, TEAC–carotenoid factor, vitamin E–carotenoid factor and UA–zeaxanthin. Because multilevel models use a regression framework with an independent and a dependent variable whereas correlations assess a symmetrical relationship, we transformed all antioxidant variables into standard normal random variables by subtracting the species-specific mean and dividing by the species-specific standard deviation. Regression intercepts thus become zero and slopes become equal to the correlation coefficient, obviating questions about how to assign dependent and independent variables. An example of the models used is as follows:
- TEACsn = (ρ1 + ρ2s)*UAsn + ɛs(eqn 1)
- UAsn = (ρ1 + ρ2s)*TEACsn + ɛs(eqn 2)
- ρ2s ~ normal (0, σ2)(eqn 3)
- ɛs ~ normal (0, 1 − (ρ1 + ρ2s)2) (eqn 4)
where TEACsn and UAsn are standard normal TEAC and UA, respectively, ρ1 is the average correlation across species, ρ2s is the species-specific deviation from ρ1 for species s, ɛs is the species-specific random error, and σ2 is the variance of the species-specific deviations from ρ1. Equations (1) and (2) are equivalent – both yield identical estimates of ρ1, each ρ2s, and σ2. Larger σ2 indicate greater heterogeneity across species.
Before running the models, we culled the data set separately for each correlation to include only species with three or more individuals without missing data for either antioxidant in the correlation. Data preparation was done in R v. 2·6·0; models were run using Monte Carlo Markov Chain simulations and Gibbs sampling in WinBUGS v. 1·4·3. For each correlation, the model ran 210 000 iterations and discarded the first 10 000 (burn-in). We tabulated 95% credible intervals based on the posterior probability distributions for the parameters ρ1, each ρ2s, and σ2. For readers who are not familiar with multilevel models and Bayesian modelling approaches, we have provided an additional introduction and our WinBUGS and R code in Appendix S4 in Supporting Information.
There are many factors at both the individual and species levels that are potentially related to antioxidant levels, including individual quality, breeding status, life-history strategy and diet. We explore all of these factors in other publications (Cohen 2007; Cohen, Hau & Wikelski 2008a; Cohen et al. 2008b, AAC, unpublished manuscripts), and cannot control simultaneously for all of them here. However, in this study such control would actually limit our ability to characterize the nature of the variation within antioxidant systems (as opposed to pinpoint its causes). In fact, adequate characterization of the variation is a prerequisite for robust analyses of factors that may determine the variation. This study is thus limited more by the factors that were controlled for in our sampling than by the factors that were not: we are unable to make generalizations to wintering birds or to unsampled taxa about the variation we detect. We explore sex differences within 16 species in Appendix S5 in Supporting Information.
Some of our analyses use multiple non-independent variables to assess similar questions – for example, as TEAC and uric acid are mechanistically dependent and highly correlated, if TEAC correlates with lutein, it is likely that uric acid will correlate with lutein as well. When multiple tests are performed, it is common to use a correction method such as a sequential Bonferroni adjustment; however, there is considerable debate as to whether this is generally appropriate (Perneger 1998; Moran 2003). In particular, when the tests are not fully independent, there is no widely accepted methodology for accounting for multiple testing issues. We believe, in this early stage of our observational research, that it is best to present raw P-values and acknowledge that multiple tests were performed in our interpretation of them rather than rely on an arbitrary standard such as α = 0·05. Most importantly, our conclusions do not hinge on the significance of individual tests but rather on the broad patterns seen across multiple tests, or from Bayesian estimates of heterogeneity in parameters. Multiple testing and false discovery rates are often considered irrelevant in a Bayesian framework such as this (Jeffreys & Berger 1992). In particular, as we are only trying to assess ‘significance’ (in this context, a conclusion that a parameter is substantially different from zero) for a small subset of the parameters estimated in the models, multiple testing is of little concern. As will be seen, some patterns are clear and consistent, some are clearly absent, and others are ambiguous and noted as such.