XY and ZW species exhibit different sex-dependent demographics

Species with ZW/ZZ and XX/XY sex-determination systems exhibit significantly different ASRs (Fig 2A and 2B; Tables 1A and S1) consistently with our previous study that used 344 species of tetrapods [30]. ASR was male-skewed in ZW/ZZ systems and a female-skewed in XX/XY systems (S1 Table and Fig 2A and 2B). The type of GSD was also associated with sex differences in adult mortality, with more male-biased mortality (i.e., higher mortality in males than in females) in XY compared to ZW species, as expected (Tables 1A and S1 and Fig 2B). These effects of the sex chromosomes on ASR and adult mortality bias were medium-strong (Table 1A). However, XY and ZW species did not differ significantly in the other demographic traits, with only weak differences in birth sex ratio, juvenile mortality bias, and maturation bias (see effect sizes in Tables 1A and S1 and Fig 2A and 2B). Note that statistical power was low for the two variables with the lowest sample size (i.e., birth sex ratio, juvenile mortality bias) due to the small differences between XY and ZW species as well as large variances within each group (S2 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Phylogenetic generalized least squares analyses of the bivariate relationships between the type of genetic sex-determination system, population demographic traits, and adult sex ratio in tetrapods. (A) Difference between XY and ZW species in demographic traits. (B) Associations between demographic traits (predictors) and the adult sex ratio (ASR, response variable). Note that the continuous variables were standardized to 0 mean and 1 SD, so the model coefficients (b) are comparable across predictors. In each model, λ is the phylogenetic signal (Pagel’s lambda), R² is the proportion of variance explained by the model, n is the number of species (number of XY species and ZW species, respectively, presented in parentheses). Statistically significant (i.e., p < 0.05) relationships are typed in bold.

https://doi.org/10.1371/journal.pbio.3003156.t001

Association between demographic traits and ASR

Sex bias in both juvenile and adult mortality as well as maturation bias were associated with ASR (Tables 1B and S3 and Fig 2C), meaning that male-biased mortalities (where more males than females die) and male-biased maturation times (where males mature later than females) were associated with female-skewed ASRs. Birth sex ratio was not related to variation in ASR. The effect sizes were moderately strong for all three significant demographic predictors, whereas it was small for birth sex ratio (Table 1B). The power analyses suggested that the dataset provided statistical power to detect a moderate or stronger effect even in the case of birth sex ratio (S2 Table). None of these results changed qualitatively when outlier data points were excluded from the analyses (S4A Table).

To assess whether these relationships were consistent between XY and ZW species, we tested the two-way interactions of each demographic variable with GSD type (S5 Table and S1 Fig). The slope of the relationships between ASR and sex differences in both juvenile and adult mortality differed between XY and ZW species, as shown by the statistically significant interaction terms (GSD × juvenile mortality bias: F_1;102 = 4.3, p = 0.041; GSD × adult mortality bias: F_1;234 = 4.6, p = 0.034). Specifically, these effects were moderately strong in ZW species but weak in XY species (S5A Table and S1 Fig), although this difference was no longer statistically significant for adult mortality bias when outliers were excluded (S4B Table). The relationships of ASR with birth sex ratio and maturation bias did not differ between XY and ZW species (GSD × birth sex ratio: F_1;108 = 0.305, p = 0.582; GSD × maturation bias: F_1;353 = 2.056, p = 0.153; S5A Table and S1 Fig), and these results remained unchanged when outlier data points were excluded (S4B Table).

Phylogenetic path analysis

The best-supported model included all identified bivariate relationships plus the direct link between GSD and ASR (Model 1.b) and yielded an acceptable fit with the data (Fisher’s C-statistic, p = 0.074; Fig 3 and Table 2A) according to the approach proposed by Santos [32]. Akaike Information Criterion corrected (AICc) values indicated substantially stronger support for this model than for any alternative (ΔAICc ≥ 3.8 in all cases, Table 2A). Note that birth sex ratio was excluded from all path models because it was consistently unrelated to both GSD and ASR in bivariate analyses (Table 1) and it was available only for a subset of species, which would restrict the analyses to a sample size too low for an informative path analysis (see Materials and methods for details). These results were consistent with the conclusions of pairwise comparisons between nested models using likelihood ratio tests (LRTs): Model 1.b outperformed Model 1.a (χ²= 3.9, df = 1, p = 0.049), justifying the inclusion of a direct link between GSD and ASR (Table 2 and S2 Fig). Furthermore, the more complex models that included additional effect(s) of GSD on demographic traits (Models 2-4.b) did not provide a better fit (χ²< 0.1, df = 1, p > 0.9 for all comparisons; ΔAICc > 8), supporting Model 1.b. According to this most supported model, sex bias in both adult and juvenile mortality as well as in maturation time all had negative effects of similar magnitude on ASR, while the type of GSD had weaker effects on the extent of male-biased adult mortality and ASR (S6 Table and Fig 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Comparison of phylogenetic path models produced (A) by the approach of Santos [32] using the implementation of ‘piecewiseSEM’, and (B) by the approach of von Hardenberg and Gonzalez-Voyer [33] using the implementation of ‘phylopath’. For each model, the table shows the number of independence claims (k), the number of parameters (q), Fisher’s C-statistic (C) for model fit, and its associated p-value. AICc and CICc are the Akaike and C-statistic information criterion, respectively, corrected for small sample sizes. ΔAICc (and ΔCICc) indicates the difference in AICc (or CICc) values between the most supported model (lowest AICc or CICc, model m1.b) and the focal models. ΔAICc (and ΔCICc) > 2 indicates substantially higher support for the best model than for the other models, whereas models with ΔAICc (and ΔCICc) < 2 have similar support (highlighted in bold type). The analyses include n = 95 tetrapod species with data available for all variables, excluding birth sex ratio. The model set and path coefficients of the most supported models are given in S2 Fig and S5 Table.

https://doi.org/10.1371/journal.pbio.3003156.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Visual interpretation of the most supported model in path analysis (Model 1b).

GSD, genetic sex determination; JMB and AMB, juvenile and adult mortality bias, respectively; MAT, maturation bias; ASR, adult sex ratio. Positive and negative relationships are shown as blue and red arrows, respectively, and numbers are the standardized path coefficients (solid lines: statistically significant relationships, dashed lines: statistically non-significant relationships). See results of phylogenetic path analysis in Tables 2 and S5, and the model set in S2 Fig. The data and phylogeny underlying the analyses and results displayed in this figure can be found in S1 Data and S1 File, 10.6084/m9.figshare.28562399.

https://doi.org/10.1371/journal.pbio.3003156.g003

These results were largely confirmed by a second set of path analyses using an alternative approach for model fitting developed by von Hardenberg and Gonzalez-Voyer [33]. The latter analyses confirmed that Model 1.b satisfactorily fits the data (Fisher’s C-statistic, p = 0.223), and model comparisons showed that this model had the lowest C-statistic Information Criterion corrected (CICc) value (Table 2B). Although two other models (Models 2.b and 3.b) had comparable support (ΔCICc < 2; Table 2B), Model 1.b was favored as the least complex one. Results for Model 1.b corroborated that the GSD had statistically significant effects on both adult mortality bias and ASR and that sex bias in both adult and juvenile mortality as well as in maturation time negatively influenced ASR (S2B Fig). The two other models with ΔCICc < 2 included an additional indirect link between GSD and ASR through either male-biased juvenile mortality or maturation time, respectively (S2A Fig). All three models showed moderately strong effects of GSD type on ASR and adult mortality bias, while the other effect sizes were smaller (S2B Fig).

Data collection

Sex-determination information was collected from the Tree of Sex database [44], with additional updates from recent primary publications (S1 Data). In this study, we only included species that have GSD systems: either XX/XY or ZW/ZZ sex-determination systems (with either homomorphic or heteromorphic sex chromosomes), or variants of these systems with demonstrated male or female heterogamety [44]. We did not include species with mixed-sex determination, i.e., where both temperature and sex-chromosome effects simultaneously operate during sex determination [45,46]. All birds were assigned to ZW/ZZ, and all mammals to XX/XY sex determination, respectively [47]. For amphibians, only species with exactly known GSD type were included, as there can be differences even between closely related species or between populations within a given species [48]. For reptiles, we included all species for which GSD was known at least either at the genus or family level and the type of GSD was not variable within the genus or family (see [30] for details). If a species (or its genus or family) for which we had demographic information did not occur in the Tree of Sex database, we searched for its GSD in primary literature. We did not identify reliable GSD data for seven amphibian species for which we had demographic information, so these species were excluded from the analyses of GSD effects on demographic variables, but were included in the analyses of demographic trait effects on ASR.

To illustrate evolutionary changes in the sex chromosome systems in the studied clades, we mapped the transitions between sex-determination systems using the ‘make.simmap’ function in the ‘phytools’ package [49] in R [50] with 100 simulations. We used two different models: the ‘equal rates’ (ER) model, where only equal transition rates are allowed between the character states; and the ‘all rates different’ (ARD) model, where different transition rates are allowed between the states, and compared the two models according to their Akaike Information Criterion (AIC) value. Transition reconstruction resulted on an average of 16.8 transitions for ARD and 18.17 for ER model between XY and ZW systems in our dataset, with an average of 1.04 (ARD) or 7.01 (ER) transitions from XY to ZW systems, and an average of 15.76 (ARD) or 11.09 (ER) transitions from ZW to XY systems. ARD had a lower AIC value than ER (∆AIC = 3.308), hence we present the results of the ARD analysis (S3 Fig). We successfully utilized this variation in sex chromosome systems among tetrapods in our earlier studies as well. For example, Pipoly and colleagues [30] showed that the sex chromosome system is associated with ASR both in tetrapods as a whole and its taxonomic subsets (amphibians and reptiles). By using a subset of 81 reptiles, Pipoly and colleagues [13] demonstrated the association between sex chromosome systems and the occurrence of multiple paternity. Finally, Bókony and colleagues [35] found differences between genetic versus environmental sex-determination systems in various demographic traits from a dataset including information on 181 reptiles. These results show that tetrapods are a suitable model system to study associations between the type of sex chromosome systems and various demographic and reproductive traits.

ASR estimates (expressed as the proportion of males in the adult population [51,52]) were taken from Pipoly and colleagues [30] for 344 tetrapod species living in the wild. This initial dataset was augmented with ASR data for additional amphibian, reptile, and mammal species, taken from existing comparative datasets (reptiles [35], mammals [17,53]), or from primary literature searched in Web of Science or in Google Scholar (updated until June 2020; S1 Data). We assessed the quality of the available data records and used only those ASR estimates that were based (1) on an accurate sexing method (i.e., molecular sexing or sexual organ identification) and (2) on a sampling method for male and female numbers that was presumably not sex-biased (see further details on data quality and repeatability of ASR estimates in refs [14,17,30,35]). The final dataset includes 453 species with ASR estimates (49 amphibians, 93 reptiles, 187 birds, and 124 mammals).

We then collected information on further demographic traits from populations in the wild. First, we collected birth sex ratio (BSR) data (the proportion of males at birth or hatching among the neonates in the population). We focused on the species listed in our ASR dataset and searched in Web of Science and Google Scholar using the search terms ‘(birth OR secondary) sex ratio*’ and the taxonomic name of the species. We evaluated these studies by similar criteria as the ASR studies (see above), and we only kept reliable estimates. We found reliable BSR data for 2 amphibian, 28 reptile, 55 bird, and 27 mammal species (for which we also had ASR information). We included BSR estimates where the whole clutches/litters were sampled soon after hatching/birth, but sexing techniques could be variable, e.g., molecular or based on sexual organs. Wilson and Hardy [51] showed that analyzing sex ratios as a proportion is reliable when sex ratios are estimated from samples of ≥10 individuals and the dataset has ≥50 sex ratio estimates, which met the requirements in our analyses for both ASR and BSR.

Second, we collected data on annual juvenile mortality rates and annual adult mortality rates for males and females separately. We used data either from existing comparative datasets (birds [2]; mammals [20]), or from primary publications (S1 Data). In the latter case, we searched the literature through the search engines Web of Science and Google Scholar, using the search terms ‘male AND female AND (mortality OR survival)’ together with the taxonomic names of the species for which ASR estimates were compiled. We categorized mortality estimates as either juvenile or adult mortality in the same way as these were provided in the original sources, i.e., as juvenile mortality when the source stated it was estimated for juveniles, immatures, or subadults, and as adult mortality when the source stated that it was calculated for adults or mature individuals. Overall, juvenile mortality corresponded to the mean annual mortality prior to the age at first reproduction, whereas adult mortality corresponded to annual mortality from the age at first reproduction onward. Mean-weighted age-specific mortalities were used when fully age-dependent estimates of adult survival were available. We calculated the sex difference in annual juvenile and adult mortality rates (juvenile mortality bias and adult mortality bias, respectively) from sex-specific estimates as log₁₀(male annual mortality rate/female annual mortality rate) [17,54,55], leading to measure the magnitude of the sex-biased mortality. Thus, positive values of mortality bias mean higher male than female mortality, whereas negative values mean higher female than male mortality. We were able to estimate juvenile mortality bias for 108 tetrapod species (3 amphibians, 17 reptiles, 52 birds, and 36 mammals), and adult mortality bias for 244 species (23 amphibians, 41 reptiles, 128 birds, and 52 mammals).

Third, we collected data on the sex-specific age of sexual maturity (in days). We used either existing datasets (for anurans [56], for reptiles [35], for birds [15], and for reptiles and mammals [57]), or searched the literature using the search terms ‘male AND female AND age AND (matur* OR first reproduction)’ together with taxonomic names. We quantified the sex differences in maturation time (referred to in this paper as maturation bias) as log₁₀(male maturation age/female maturation age). Positive values mean that males mature later than females, whereas negative values mean that females mature later than males. We were able to estimate maturation bias for 366 species (37 amphibians, 70 reptiles, 175 birds, and 84 mammals).

Most data on sex ratios, mortality, and maturation were collected blindly with respect to the hypotheses we test in the current analyses because these data were originally collected for the purposes of other comparative studies. Data on different demographic traits for the same species were often available only from different populations or studies. For all demographic traits, we used the mean value for each species whenever reliable data were available from more than one population or study.

Statistical analyses

All analyses were run in R (version 4.0.0; R Core Team 2020]. We standardized all numeric variables by centring to the mean and dividing by standard deviation, and we used these values for all analyses in the main text. We built phylogenetic generalized least squares (PGLS) models to conduct bivariate analyses using the ‘pgls’ function of the ‘caper’ [58] R package. To control for the phylogenetic relatedness among taxa, we used the composite phylogeny applied in [30] with the addition of new species according to a family-level reptile phylogeny [48] and other recent phylogenies (Squamata [59,60], Testudines [61], birds [62], mammals [63,64]). Since composite phylogenies do not have true branch lengths, we used Nee’s method to generate branch lengths using Mesquite software [65]. The S1 File contains the phylogenetic tree we used in the analyses. In each model, we used the phylogenetic signal (i.e. Pagel’s λ [66]) as estimated by the maximum-likelihood method for that model.

Using all species with available data, we first applied two sets of bivariate PGLS models. In the first set of analyses, we tested whether each of the examined demographic traits differed between the two types of GSD (i.e., XY versus ZW species). In these models, the type of GSD was the predictor variable, and one of the four demographic trait (birth sex ratio, juvenile mortality bias, adult mortality bias and maturation bias) was the dependent variable. In the second set of analyses, we assessed whether demographic traits predict ASR. In these models, ASR was the response variable and one of the four demographic traits was the predictor variable. To check whether a non-linear relationship would fit the data better than a linear model, we re-ran this second set of models with adding the quadratic term of the predictor. However, these latter models never provided a better fit than the ones without the quadratic term (S7 Table). We conducted statistical power analyses for these models using simulation. We calculated power to detect small, medium, and large effects corresponding to Cohen’s d effect sizes of 0.3, 0.5., and 0.8 in the first set of models, and correlation effect sizes of 0.1, 0.3, and 0.5 in the second set of models (see also Table 1). We used an implementation of simulation power analysis for linear models (R function est_lm_power_slope, published at https://public.wsu.edu/~jesse.brunner/classes/bio572/Lab5_SimulationAndPower.html), which we applied to phylogenetically transformed data (using the same transformation as in the phylogenetic path analyses, see below).

We also tested whether the relationships of ASR with the demographic predictors differed between GSD types. For this analysis, we built PGLS models using the ‘gls’ function of the ‘ape’ [67] R package, with the corPagel correlation structure accounting for phylogenetic relationships among species. In these models, ASR was the response variable, and one of the demographic traits, GSD type, and demographic trait × GSD type interaction were included as predictors. We used type-3 ANOVA tables (R function ‘anova.gls’ with the argument type = ‘marginal’) to test whether the interaction was statistically significant in each of the above models. Then, for each interaction model, we tested whether the slope of the relationship between ASR and the demographic variable differed from zero in either of the two GSD types (XY or ZW species) using the ‘emtrends’ function of package ‘emmeans’ [68].

The validity of the models was checked by inspection of residual diagnostic plots to ensure that all statistical assumptions were met. We also checked the sensitivity of the results to the influence of outlier data points, which were identified from the standardized predictor variables as being more than three standard deviations distance from the mean (i.e., have high leverage). Outliers were identified separately for each group (XY or ZW) within each model. We repeated all PGLS analyses with demographic predictors after excluding these potentially highly influential data points (see S3 Table for number and identity of species excluded). Throughout the paper, we present figures using the untransformed data points (i.e., non-standardized values) to facilitate interpretation. We repeated the bivariate analyses using non-standardized variables (S3 Table) to allow for the interpretation of the effect sizes on the original data scales. For illustration purposes, we added regression lines to the figures calculated from PGLS models using non-standardized data, and plotted the uncertainty of the slopes using the ‘evomap’ [69] package in R. In tables of the main text, we present the results with the transformed (i.e., standardized) variables.

We used phylogenetically controlled confirmatory path analyses to investigate the fit of models representing different sets of pathways linking GSD and ASR (S2A Fig). First, we constructed a path model that included links (1) between GSD and adult mortality bias, and (2) between ASR and three demographic traits: juvenile and adult mortality bias, and maturation bias (Model 1.a). This model aimed to simultaneously estimate the relationships that were detected in the bivariate PGLS analyses (see Results and Table 1), and also corresponded to the existing knowledge on the effect of GSD on sex-specific life span [18,19] and the association of ASR with sex-biased mortality and maturation [15,25,29]. We did not include the path leading from BSR to ASR because it was consistently not supported both by our current results (see Table 1) and earlier analyses [2]. Furthermore, BSR was available only for a relatively small subset of species, thus by excluding this variable we were able to double the sample size of the path models (n = 43 and 95 species, with BSR included or excluded, respectively), resulting in a more suitable sample size for path analyses [70]. Then, we created an extended model by adding the direct link between GSD and ASR (Model 1.b). Finally, we created three further expanded versions of both Models 1.a and 1.b, each of which contained additional paths between GSD and (1) juvenile mortality bias, (2) maturation bias or (3) both juvenile mortality and maturation bias (Models 2-4.a and 2-4.b). By these latter models, we tested whether the fit of path models can be improved by including relationships that were not supported in the bivariate PGLS analyses. Note that Model 1.a is nested in Model 1.b, and both of these models are nested in Models 2-4a and 2-4b, respectively, facilitating model comparison (see below). We used those species for path analyses for which we had data for all included variables (i.e., ASR, GSD, juvenile and adult mortality bias, and maturation bias; n = 95 species).

We used two methods to fit the path models to the data. Firstly, we followed the approach proposed by Santos [32] that applies phylogenetic transformation on the data before the path analysis to control for the effects of phylogenetic relatedness among species. For this purpose, Pagel’s λ was estimated separately for each variable by maximum likelihood, and this variable-specific λ value was used to re-scale the phylogenetic tree. Then, we used this transformed tree to calculate phylogenetically independent contrasts for the variable using the ‘pic’ function of the R package ‘ape’ [67]. We repeated this process for each variable, and the resulting phylogenetically transformed variables were used for fitting the path models by the R package ‘piecewiseSEM’ [71], which implements the d-separation method [72]. This approach uses Fisher’s C-statistic to test the goodness of fit: the model is rejected (i.e., it does not provide a satisfactory fit to the data) if the C-statistic is statistically significant, and conversely, a high p-value means satisfactory fit [71]. Model support was compared between different path models by their AICc (Akaike Information Criterion corrected for low sample sizes) values. Since information-theoretic model comparisons (e.g., those based on AICc) can be unreliable for path models [70,73] (see also Appendix 3 in ref [17] for an example in the context of comparative analyses), we also conducted pairwise model comparisons using LRTs. LRT is straightforward to conduct for pairs of nested models, where a statistically significant difference supports the model with higher likelihood (better fit) [74]. For LRTs, we conducted the model fitting using the ‘lavaan’ R package [75] and used the ‘anova’ function to compare pairs of nested models.

Secondly, to test the robustness of the results, we repeated the path analysis using the method developed by von Hardenberg and Gonzalez-Voyer [33]. Unlike Santos’ method [32], in this latter approach, a single value of Pagel’s λ is estimated for each pair of traits using PGLS models, rather than a value of λ for each variable. Model fitting and model comparisons were conducted using the ‘phylopath’ package [76]. Model fit was evaluated using the d-separation method [72] and compared between path models by their CICc for small sample sizes [33]. CIC is a modified version of AIC developed for d-separation path analysis framework [77] and is used in the same way as AIC to compare model fit. Confidence intervals for the standardized path coefficients were estimated by bootstrapping, using the ‘boot’ function with 500 iterations.