Recordings

Recordings, each representing a single vocalizing Hermit Thrush, were collected from the Cornell Lab of Ornithology's Macaulay Library ( macaulaylibrary.org), xeno-canto ( www.xeno-canto.org), Michigan State University's Avian Vocalization Center ( avocet.zoology.msu.edu), and the Ohio State Borror Laboratory of Bioacoustics ( blb.osu.edu), as well as a number of independent recordists. Recording dates ranged from 1951 to 2015. In cases where the exact GPS coordinates or altitude information was not provided for a given recording, that information was obtained by inputting location information reported by the recordist into Google Earth. The recordings varied substantially in quality, including compressed MP3 files from xeno-canto and from some of the independent recordists. Although compression of recordings can influence the measurement of acoustic variables (e.g., Medina-Garcia et al. 2015), this variation in quality did not influence our findings: rerunning the MANCOVA described below with compressed MP3 recordings excluded from the analysis did not alter our results, nor did running the same MANCOVA with recording quality (compression or no compression) treated as a covariate.

Given the lack of evidence for song-type sharing in Hermit Thrush (Jones 2005, Roach et al. 2012), we were able to use visual inspection of song spectrograms to examine recordings made in close proximity to each other in order to eliminate any duplicate recordings of individuals. The resulting 101 recordings of unique individuals were classified into major geographic group (Northern, Western Lowland, and Western Mountain) and subspecies based on recording location using the detailed geography-based range descriptions of the subspecies making up the groups provided by Dellinger et al. (2012).

Spectrographic analysis of song structure was conducted using RavenPro 1.4 ( www.birds.cornell.edu/raven) (Hann window size 1000 FFT). For each song type within a recording, we measured 8 acoustic variables associated with frequency and duration: frequency at the start (FI1) and end (FI2) of the introductory note, the absolute change in frequency from the start to the end of the introductory note (FIΔ), the minimum (FPMin) and peak energy (FPPeak) frequencies of the post-introductory portion, the bandwidth of the post-introductory portion (difference between maximum and minimum frequency; FPBand), and the durations of the introductory note (TI) and the post-introductory portion (TP) (Figure 1).

In the interest of comparing structure among individuals, and to avoid issues related to pseudoreplication, we calculated each variable's mean across all song types within a recording. However, it was also important to account for the fact that there is substantial variation within individual Hermit Thrush repertoires with respect to frequency characteristics. For example, in the Northern subspecies C. g. faxoni, the introductory note frequencies of each individual's song types typically span 3,000 Hz or more (Roach et al. 2012). To capture such variation and avoid losing information, we calculated not just the mean but also the maximum and minimum frequency for the song types in each recording. With mean values for all 8 acoustic variables and maximum and minimum values for the 6 frequency-related variables, there were a total of 20 acoustic variables used in the analyses.

Statistical Analyses

Because of the uncertainty regarding Hermit Thrush subspecies designations based on genetic information, statistical analyses were conducted mainly on the level of the 3 major geographic groups (Northern, Western Mountain, and Western Lowland). Analyses were completed in IBM SPSS Statistics 22 with an α value of 0.05, unless otherwise noted.

To reduce the number of variables and examine how they contributed to song structure variation, we conducted a principal component analysis (PCA) based on the correlation matrix of the 20 variables. Principal components with eigenvalues greater than 1.00 were used in subsequent analyses. One of the resulting principal components (PCs) lacked within-class normality, as assessed by examining the data's residuals, and could not be made normal by transformation (with log, ln, or square root). Thus, the principal component data did not the meet the multivariate normality assumption of the planned analyses (MANCOVA and discriminant function analysis). One multivariate outlier was identified based on examination of both the raw data and of plots showing residuals of regression for the principal components. When the principal component analysis was rerun with the remaining 100 individuals, the non-normal distribution was still present.

Because the PC data did not meet the requirements for MANCOVA and discriminant function analysis, data were first examined using a classification method that does not require a normal distribution, k-nearest neighbor analysis, which classifies cases into groups based on the group membership of the most similar other cases. It correctly predicted the group membership of 89% (89/100) of individuals. Among the 11 misclassified individuals were 5 of the 6 individuals belonging to C. g. slevini, a Western Lowland subspecies. Similarly, the algorithm misclassified 5 of the 6 members of the Northern subspecies C. g. euborius that were recorded at high altitudes (>1,900 m) in the Canadian Rockies (7 other C. g. euborius were recorded at lower altitudes (<1,500 m) in the surrounding areas). On the basis of these consistent misclassifications, the 100 birds were reclassified into 5 groups: Northern (n = 39 individuals), Western Lowland (n = 17), Western Mountain (n = 32), high-altitude C. g. euborius (consisting of the individuals recorded >1,900 m; n = 6), and C. g. slevini (consisting of the members of this subspecies; n = 6). This new classification exhibited within-class normality and was therefore used for the subsequent analyses.

Differences in song structure across these 5 groups were assessed via MANCOVA using the principal components that resulted from the PCA. Because the sampling period of the recordings was so large (1951–2015), year of recording was treated as a covariate. Between-group differences were examined using Bonferroni-corrected post-hoc comparisons. In order to test whether individuals could be accurately classified on the basis of song structure, we then conducted a discriminant function analysis (DFA) using the same principal components. Geographic group was used as the grouping variable, with the five-group classification described above.

To assess how ecological factors may have influenced song variation, we used Mantel and partial Mantel tests to measure correlations among matrices representing geographic distance (the distance between recording sites), ecological distance (how different recording sites are with respect to climatic variables), and song distances (how different recordings are from each other with respect to song structure) between recordings. For ecological distance, we used climate data related to various aspects of rainfall and temperature; such data have been used by other studies of geographic song variation (e.g., Ruegg et al. 2006b, Wei et al. 2015) in order to provide a broad picture of environmental effects on habitat. Climatic data representing the period 1950–2000, which corresponds to this study's sampling period, were downloaded from  http://worldclim.org, and 19 climatic variables representing temperature and precipitation data were extracted using ArcGIS (Esri, Redlands, California, USA) for each recording site at a resolution of ∼1 km². A principal component analysis based on the correlation matrix was conducted in order to reduce the number of climatic variables. A matrix of the Euclidean ecological distances between each pair of recordings was calculated in SPSS using the principal components with eigenvalues over 1.00; likewise, a matrix of Euclidean song distances was calculated from the first 2 discriminant functions resulting from the discriminant function analysis. Using the longitude and latitude of each recording site, the package fields in the statistical analysis program R was used to calculate a matrix containing the great circle distance (the shortest route over the earth's surface between any 2 points) between each pair of recording locations.

The analysis program PASSaGE 2 ( www.passagesoftware.net/) was used to conduct Mantel tests testing for correlations between song distance and each of geographic and ecological distance, thus measuring the degree to which song differences varied in concert with geographic location and ecological factors. We also conducted partial Mantel tests, wherein a correlation between 2 matrices is calculated while the influence of a third is controlled.

Principal Component Analysis

Analyses included recordings of 100 individuals, containing 873 song types (Figure 2). Examples of songs from each subspecies included are shown in Figure 3. A principal component analysis conducted to reduce the initial 20 acoustic variables generated 5 principal components with eigenvalues greater than 1.00 (see Table 1 for factor loadings). Together, these principal components explained 79.57% of the variation observed in the original 20 variables. The raw data for the 20 acoustic variables, classified by geographic group, are summarized in Appendix Table 3.

FIGURE 3.

Spectrograms showing example song types from subspecies within the Northern (N), Western Lowland (WL), and Western Mountain (WM) groups.

TABLE 1.

Factor loadings of acoustic variables for principal components with eigenvalues >1. High factor loadings are in bold.

Comparison across Geographic Groups

A MANCOVA using the above-noted principal components revealed an overall difference in song structure across the 5 groups (Pillai's Trace = 2.026, F = 19.087, p < 0.001) while controlling for year of recording. There were significant differences by group with respect to PC1 (F = 342.563, df = 4 and 100, p < 0.001), PC2 (F = 23.568, df = 4 and 100, p < 0.001), PC3 (F = 8.311, df = 4 and 100, p < 0.001), and PC4 (F = 7.732, df = 4 and 100, p < 0.001), but not for PC5 (F = 2.058, df = 4 and 100, p = 0.07).

Bonferroni-corrected post-hoc comparisons between groups revealed that PC1, which had high loadings for the mean and minimum values of introductory note frequency (FI1 and FI2) and absolute frequency change within the introductory note (FIΔ) (see Table 1), was higher in the Western Lowland group than the Western Mountain group (p < 0.001), and higher in the Western Mountain group than the Northern group (p < 0.001) (Figure 4A). The introductory note frequencies of these groups occupied dramatically different ranges of values, as can be seen by plotting the introductory note frequencies of all song types measured (Figure 5). Specifically, the introductory note frequencies of Northern birds spanned a large range (approximately 1,700–5,000 Hz in terms of FI1), while those of Western Lowland (approximately 4,000–5,500 Hz), and Western Mountain (approximately 3,000–4,000 Hz) birds occupied frequency ranges that were smaller, but distinct from each other (Figure 4A). Figure 5 also demonstrates the differences in frequency change within the introductory notes (FIΔ), the magnitude of which is reflected by the distance from the line that represents no frequency change (slope = 1): FIΔ was largest in Western Lowland birds, smaller in Western Mountain birds, and very small in the Northern group. With respect to PC1, high-altitude C. g. euborius birds were intermediate between Western Lowland and Western Mountain individuals, while C. g. slevini birds were most similar to the Western Mountain group.

FIGURE 4.

Boxplots showing group data for PCs 1–4 (A–D). Significant differences between groups (Bonferroni-corrected post-hoc comparisons; α = 0.05) are indicated by asterisks (*); for PC1, all between-group differences were significant except for one (marked by “N.S.”).

FIGURE 5.

Scatterplots showing FI1 and FI2 for all song types of individuals from the Northern (n = 369 song types), Western Lowland (n = 129), and Western Mountain (n = 270) groups. Song types from high-altitude (>1,900 m) C. g. euborius individuals from the Northern group (n = 45) and birds from C. g. slevini (n = 60) are shown separately. Diagonal lines indicate where FI1 and FI2 would be equal, representing no change in frequency within the introductory note (FIΔ = 0); thus, distance from the lines reflect the magnitude of FIΔ.

For both PC2 and PC3 (Figures 4B and 4C), most strongly associated with post-introductory frequency measures (the mean and maximum values for FPMin and FPPeak, as well as all FPBand values; Table 1), the Northern group was higher than the Western Mountain group (PC2: p < 0.001; PC3: p < 0.001), as was the Western Lowland group (PC2: p < 0.001; PC3: p = 0.003). The Northern and Western Lowland groups did not differ from each other. High-altitude members of C. g. euborius had lower PC2 values (FPMin and FPeak) than Northern (p = 0.002) and Western Lowland birds (p = 0.003), and did not differ from any groups with respect to PC3 (FBand). C. g. slevini birds were lower than Northern (p = 0.003) and Western Lowland birds (p = 0.005) with respect to PC3, but similar to those groups with respect to PC2 (Figures 4B and 4C). Values for PC4 (Figure 4D), associated with minimum post-introductory peak frequency (FPPeak _Min) and introductory note duration (TI), were significantly lower in C. g. slevini compared to the other groups (p < 0.002 for each); the other groups did not differ from each other on PC4.

Because neither of the temporal characteristics loaded highly on PCs that explained substantial variation in the original variables, each characteristic was assessed separately via ANCOVA, with year of recording treated as a covariate. Duration of both the introductory note (TI) and post-introductory song portion (TP) differed significantly across the groups after controlling for year of recording (TI: F = 12.433, df =4 and 95, p < 0.001; TP: F = 9.406, df = 4 and 95, p < 0.001) (Figure 6). Bonferroni-corrected post-hoc comparisons revealed that introductory note duration was longer in Western Mountain birds compared to Northern birds; in contrast, post-introductory note song duration was greater in Northern birds than in Western Lowland and Western Mountain birds.

FIGURE 6.

Boxplots showing group data for duration of the introductory note (A) and the post-introductory song portion (B). Significant differences between groups (Bonferroni-corrected post-hoc comparisons; α = 0.05) are indicated by asterisks (*).

Discriminant Function Analysis

A discriminant function analysis (DFA) was conducted using the 5 principal components and with geographic group (Northern, Western Lowland, and Western Mountain, plus C. g. slevini and high-altitude C. g. euborius in their own groups) as the grouping variable further demonstrated that song structure was distinctly different across the groups: it correctly classified 97.0% (97/100) of the individuals (Figure 7). The 3 incorrectly classified individuals included a C. g. slevini bird misclassified as Western Mountain, a Western Mountain bird misclassified as C. g. slevini, and a Western Lowland bird misclassified as high-altitude C. g. euborius (see filled-in symbols in Figure 7).

FIGURE 7.

Plot of discriminant functions 1 and 2 representing results of discriminant function analysis. Symbols represent group affiliation; empty symbols represent correctly classified birds, black-colored symbols represent misclassified birds, and green-colored symbols represent group centroids.

The position of group centroids and the distribution of individuals indicated that song structures across the 3 major geographic groups (Northern, Western Lowland, and Western Mountain) were highly distinct from each other, with Northern and Western Lowland song structure more similar to each other than either was to the Western Mountain group (Figure 7). C. g. slevini song structure was similar to that of the Western Mountain group (this similarity accounted for 2 of the 3 misclassifications), while song structure in high-altitude C. g. euborius was intermediate between the Western Lowland and Western Mountain groups.

Relationships between Song Structure, Geographic Distance, and Ecological Factors

To examine possible connections between ecology and song structure, we conducted Mantel tests designed to measure the correlations between differences in song structure, geographic distance, and ecology (assessed using 19 climatic variables; Appendix Table 4) across individuals and their recording locations. A principal component analysis using the climatic variables generated 4 PCs with eigenvalues above 1.00, which combined to explain 91.20% of the variation across recording sites (Appendix Table 4).

The results of the Mantel tests are summarized in Table 2. Across all recordings, song distance was strongly correlated with both geographic distance and ecological distance, even when partial Mantel tests were used to compare song distance to each of those variables while excluding the influence of the other. In the Northern group, there was a significant correlation between song distance and geographic distance, which remained after controlling for ecological distance; there was also a significant correlation between song distance and ecological distance, although it did not remain after controlling for geographic distance. When high-altitude C. g. euborius birds were excluded from the Northern group, Mantel tests revealed no significant associations. In the Western Lowland group, song distance was correlated with both geographic distance and ecological distance; as with the Northern group, only the association between song distance and geographic distance remained significant following partial Mantel tests. Exclusion of C. g. slevini reduced the strengths of the associations observed in the Western Lowland group. Song distance was correlated with neither geographic nor ecological distance within the Western Mountain group.

TABLE 2.

Summary of the associations between song distance and each of geographic distance and ecological distance, as measured by Mantel tests. Bold formatting indicates that the association remained significant after excluding the influence of the third factor (e.g., controlling for geographic distance in the relationship between song distance and ecological distance) via Partial Mantel test. “N.S.” indicates that the Mantel test did not indicate a significant association.

Evolutionary Divergence of Major Geographic Groups

Some of the observed variation in Hermit Thrush song structure likely dates back to separation of Hermit Thrush into 2 lineages, as instigated by the emergence of ice sheets during the Pleistocene era. Genetically, Hermit Thrush can be divided into 2 distinct groups representing eastern (stretching west to eastern British Columbia, roughly corresponding to the Northern group) and western (corresponding to the Western Lowland/Western Mountain groups) populations (Kerr et al. 2007, Topp et al. 2013, Alvarado et al. 2014) that also differ from each other morphologically (mainly with respect to body size; Alvarado et al. 2014). Overall differences in song structure align with this eastern–western division, as there were distinct differences between the Northern and Western Lowland/Western Mountain groups for both frequency (e.g., introductory note frequency) and temporal characteristics (i.e. duration of the introductory and post-introductory parts of the song).

Some of the song structure differences within the western lineage (i.e. between the Western Lowland and Western Mountain groups) may be related in part to further separation events. In examining the evolutionary histories of boreal superspecies, Weir and Schluter (2004) pointed not just to the above-noted separation event one million years ago but also to a second one, which involved more southern glacial advances and took place around 700,000 years ago. For such superspecies complexes, this second period of separation led to division of the western groups into so-called Pacific Coast and Rocky Mountain fragments, which geographically align with this study's Western Lowland and Western Mountain groups, respectively. While there is no evidence of further genetic divergence within the western lineage of Hermit Thrush, it is possible that this second separation event contributed to differences in song structure between the 2 western groups. Any song divergence that accumulated (e.g., via cultural drift) may have been maintained in part by the group's different migration patterns: Western Lowland birds travel along the Pacific coast, while Western Mountain birds are altitudinal migrants that winter in valleys near their Rocky Mountain breeding grounds (Dellinger et al. 2012).

Frequency differences between the 2 western groups may relate in part to body size, which has been negatively correlated with song frequency in a number of songbird species (Ryan and Brenowitz 1985). The frequency of both introductory and post-introductory song components was higher in Western Lowland birds, a group reported by Aldrich (1968) to contain the smallest of all Hermit Thrush subspecies, than in the Western Mountain group, which includes the largest subspecies. This variation in body size may also contribute to the comparatively larger bandwidth of the post-introductory portion of song in Western Lowland birds, as bandwidth has also been negatively correlated with body size in some species (e.g., Mason and Burns 2015).

Role of Altitude and Habitat

Overall, song structure was relatively consistent within each of the 3 major geographic groups. One interesting exception involved high-altitude members of the Northern subspecies C. g. euborius located in the Canadian Rockies. While a portion of C. g. euborius birds sang songs that were similar to the rest of the Northern group, song structure in birds at high elevations (1,910–2,275 m) was unlike that of other Northern birds and intermediate between Western Lowland and Western Mountain song structure, as indicated by their DFA misclassification (Figure 7). Compared to the Northern group, these birds had songs with higher introductory note frequencies and lower post-introductory frequencies. Such elevation-related differences in song structure have been documented in some other songbirds, such as the Gray-breasted Wood-Wren (Henicorhina leucophrys) (Caro et al. 2013).

These differences may relate in part to habitat structure at higher elevations. Compared to the high tree density at lower elevations in the Rockies, the habitat around 2,000 m is more open, with trees giving way to shrubs like mountain heathers as the subalpine forest transitions toward the alpine tundra (Luckman 1990). Particularly for the introductory note, which may function in long-range communication and therefore be more subject to the habitat's effects on sound transmission, this open habitat may have led to comparatively higher frequencies, as predicted by the acoustic adaptation hypothesis (Wiley and Richards 1978). Due to the limited number and coverage of the recording sites examined here, it remains unclear whether song structure changes gradually with increasing elevation, as in other species (e.g., Caro et al. 2013), or more abruptly. Recording individuals at various altitudes along this elevation gradient, possibly in conjunction with playback experiments to examine responsiveness to different song forms, will clarify the nature and significance of song structure variation in this part of the species range.

Birds from C. g. slevini, a Western Lowland subspecies, were also frequently misclassified by the DFA, indicating that their song structure differs from their parent geographic group. Both the introductory and post-introductory portions of the song were lower in frequency among C. g. slevini individuals compared to other Western Lowland birds. These differences may also relate to habitat. While most Western Lowland birds occupy an ecoregion (Marine West Coast Forests) characterized by very high levels of precipitation, C. g. slevini occupy a comparatively dry ecoregion (Mediterranean California) along the Pacific coast, where precipitation is seasonal and overall lower than in other Western Lowland habitats. It may be that the higher song frequencies among most Western Lowland birds reflects the influence of sound interference, in the form of constant rainfall (see Slabbekoorn 2004), a selection pressure that would be less prevalent for C. g. slevini. It is also possible that the more open habitats within the Mediterranean California ecoregion contribute to the lower song frequencies observed in C. g. slevini.

Comparison of song structure and ecological characteristics using Mantel tests revealed a positive correlation between differences in song structure and differences in habitat. This association suggests that environmental factors drive song variation to some degree. However, although ecological differences likely contribute in some cases (as in the cases of C. g. euborius and C. g. slevini described above), the overall differences among the main geographic groups argue against a dominant role for habitat. For example, birds in the Western Mountain group, which occupy open, high-altitude habitats, sang at the lowest frequencies—opposite of what would be predicted by the acoustic adaptation hypothesis (Wiley and Richards 1978). It may be that the association between environment and song structure is somewhat spurious: other factors such as cultural drift or morphological differences may covary with ecological differences and play a larger role in song divergence. Future work that examines morphology, genetics, and song structure in concert may help to resolve this uncertainty.

Independent Divergence of Song Components

Our analysis suggests that each of the 2 components of Hermit Thrush song, the introductory note and the post-introductory portion, have been shaped somewhat independently and subjected to a different mix of influences. If the 2 components were shaped by the same factors and in the same way, we would expect the differences among groups to be similar for both parts of the song. Instead, the patterns of geographic divergence are quite different for each of the components. For example, introductory note frequencies were highest in Western Lowland birds and lowest in Northern birds, and occupied dramatically different ranges of the spectrum; in contrast, post-introductory frequencies of Western Lowland and Northern birds were similar to each other and higher than those of Western Mountain birds. With respect to temporal characteristics, Western Lowland and Western Mountain birds had the longest introductory notes, whereas Northern birds possessed the longest post-introductory song durations.

This disconnect may be partly a result of the structurally simple introductory note and more complex post-introductory portion serving different functions. Research in other species has shown that whistles of consistent frequency, such as the introductory note, are more resistant to degradation than other types of song elements (Brown and Handford 2000) and their transmission may even be aided by forest reverberations (Slabbekoorn et al. 2002). In Blackbirds (Turdus merula) (Dabelsteen and Pederson 1993, Dabelsteen et al. 1993) and Eastern Towhees (Pipilo erythrophthalmus) (Richards 1981), the introductory song portion, which in each case is structurally simple and therefore relatively resistant to degradation by the environment, is important for both long-range communication and conspecific recognition. In contrast to those introductory portions, more complex parts of Blackbird and Eastern Towhee songs, which are more vulnerable to degradation, are thought to serve close-range functions related to individual recognition. Since Hermit Thrush have a comparable song structure to the above species, with a whistled introductory note and complex post-introductory portion, it may be that a similar functional division exists, with the introductory note serving in long-range species recognition (possibly directed mainly at females) and the rest of the song more associated with short-range individual recognition by neighboring males.

In partial support of this idea, research in another thrush species, the Veery (Catharus fuscescens), suggests such a role for its introductory note. Playback experiments with Veeries indicate that the introductory note is not essential for eliciting territorial responses (MacNally et al. 1986), and male Veeries often use songs lacking introductory notes in close-range territorial interactions with other males (Belinsky et al. 2015). Like Veeries, Hermit Thrush males often sing songs lacking an introductory note in response to territorial intrusions (Roach 2016). If the introductory portion functions primarily for long-range signaling in Hermit Thrush, it may be subject to unique influences, such as sexual selection. Further playback studies, in combination with transmission experiments, will provide further information on the respective roles of the introductory and post-introductory portions of Hermit Thrush song. This, in turn, will help in understanding the forces that have shaped each part of the signal.

Conclusion

In summary, this study has documented reliable differences in song structure across the breeding range of Hermit Thrush. Overall, the observed patterns of geographic song variation in Hermit Thrush are likely the result of a complex mix of factors and selection pressures that have influenced song structure at various points during the species' evolutionary history. Broadly, the most distinctive differences between the Northern and the Western Lowland/Western Mountain groups (e.g., introductory note frequency) likely date back to the separation of Hermit Thrush into eastern and western lineages, and arise via genetic differences and cultural drift. Song divergence within the western lineage may also have resulted from cultural drift, maintained by differences in migration and overwintering between the Western Lowland and Western Mountain groups; it is also possible that secondary consequences of natural selection's effects upon body size have contributed to spectral differences between these groups. Selection for optimal sound transmission in different habitats has also likely contributed to song variation, especially in the case of within-group variation (i.e. C. g. slevini and high-altitude C. g. euborius).

The large-scale song differences between eastern and western lineages parallel established genetic and morphological differences in some respects, although variation within and between the geographic groups highlight more extensive influences on song structure. Future work that examines song structure, morphology, and genetics in concert—especially in areas of interest such as the hybrid zone identified by Alvarado et al. (2014) and the Canadian Rockies—will help to clarify the evolutionary history of Hermit Thrush as well as the forces that have shaped its song over that span. Such information has the potential to add to our overall understanding of how song structure variation contributes to speciation and associated processes in songbirds.