How climatic variability is linked to the spatial distribution of range sizes: seasonality versus climate change velocity in sphingid moths

To map the spatial variation of range sizes within sphingid moths, and to test hypotheses on its environmental control. In particular, we investigate effects of climate change velocity since the Pleistocene and the mid‐Holocene, temperature and precipitation seasonality, topography, Pleistocene ice cover, and available land area.

Many hypotheses have been proposed to explain patterns of range sizes. Janzen (1967) suggested that species living in regions with high temperature stability throughout the year (i.e. the tropics) are tolerant to a narrower range of temperatures than species in highly seasonal regions. Stevens (1989) adopted this idea to explain a positive latitude-range size pattern (which he described as Rapoport's rule; Letcher & Harvey, 1994;McCain & Bracy Knight, 2013;Veter et al., 2013).
Besides such intra-annual variability, long-time climatic oscillations were also proposed to influence range sizes. Jansson (2000, Jansson &Dynesius, 2002) and Jansson (2003) connected several biological phenomena, including range size variation, with long-time climatic oscillations driven by changes in the Earth's orbit. These have stronger effects towards the poles and therefore cause larger temperature changes at higher latitudes. Dynesius and Jansson (2000) argued that areas of long-time climate stability allow for the persistence of small-ranged species, while only large-ranged species (which often have high climatic tolerances, are generalists and good dispersers) could survive in regions of low stability. Sandel et al. (2011) built on these ideas to connect the proportion of small-ranged species with the concept of climate change velocity (CVV).
Climate change velocity was developed as a measurement for long-time climate variability by Loarie et al. (2009) and adopted by Sandel et al. (2011). CCV expresses the speed at which species have to migrate to track a changing climate. CCV is influenced by a temporal gradient of change as well as the spatial change of climate in a region (which is high where there is high topographic variability).
Highest CCV occurs in flat landscapes with a high magnitude of climatic change, while it is lowest in mountainous regions with relatively stable climate (because there species do not need to travel far to reach a different climatic zone; Ackerly et al., 2010;Burrows et al., 2014;Loarie et al., 2009;Sandel et al., 2011). Focusing on temperature CCV since the Last Glacial Maximum (LGM), Loarie et al. (2009) andSandel et al. (2011) suggested a connection between small geographical range ('endemism') and low CCV, and discussed this in the light of species' vulnerability when exposed to high future CCVs.
Other potentially influencing factors include elevation range, available land area, and long-or short-term variation of climatic variables other than temperature. Elevational range affects habitat rarity (Hawkins & Diniz-Filho, 2006). Habitats found in highland regions typically have small area sizes. Hence, many species that occur there must be expected to have relatively small ranges. Similarly, available land area could be a relevant predictor in large-extent analyses (Ohlem€ uller et al., 2008). Terrestrial species can only have large ranges if there is sufficient land area available. Also, there is no ecological reason why seasonality (or CCV) effects should be related to temperature variation but not to precipitation (which is a relevant niche dimension for many species), or even more complex combinations of climatic variables.
Here we used the range size distributions of Old World sphingid moths, a family of herbivorous insects, to test the above hypotheses in a competitive manner for their explanatory power. We were especially interested in the recently published hypothesis of CCV effects (Sandel et al., 2011; that is, temperature change velocity) in comparison to the older hypothesis of temperature seasonality (Janzen, 1967). We also evaluated the evidence for different mechanisms acting simultaneously in shaping range sizes (i.e. their relative contribution in explaining patterns after accounting for the other hypothesized effects).
There is no a priori knowledge of what is an appropriate time window of CCV effects for a given taxon-we focus here on testing Sandel et al.'s (2011) specific hypothesis of post-Pleistocene effects (i.e. since LGM) as a general mechanism. However, we also use CCV calculations for a different time period (mid-Holocene to present), as mid-Holocene temperatures in many parts of the world were warmer than today. This will help to elucidate the suggested mechanism, which is not based on the direction but on the speed of climate change. Thus, we expect similar relationships of range size with CCV for both time periods. This also acknowledges that climate change since the LGM has not been linear (e.g. Thompson et al., 1998). Furthermore, we evaluated the placement of species' ranges within biomes (Olson et al., 2001). This will assist in assessing whether range size patterns are mainly due to large-scale habitat (i.e. vegetation) differences.

| Sphingid moths and range size data
Sphingid moths are a family of large, mobile and, in some cases, extremely dispersive Lepidoptera (Kitching & Cadiou, 2000). Caterpillars are folivorous with a moderate degree of host plant specialization (typically to plant family or order). Due to their popularity with amateur collectors, more is known about their taxonomy, distribution and life history than for many other taxa, making them a suitable model group for global-scale biogeographical studies on insects (Ballesteros-Mejia, Kitching, Jetz, & Beck, 2017).
Distribution maps for all 972 sphingid species found in the research region were available at 5 km grain size. These data were based on a carefully processed multi-source compilation of specimen records, combined with species distribution modelling techniques (SDM; based on climatic variables and vegetation cover) and then expert-edited to account for dispersal limitation (for details and validation see Ballesteros-Mejia et al., 2017; maps can be browsed in Map of Life, www.mol.org). For each species, range size was calculated (Appendix S1). We then used up-scaled distribution maps to a 100 km grain to calculate the median range size of the species cooccurring in each pixel (i.e. a 2-D version of 'Steven's method'; Gaston, Blackburn, & Spicer, 1998;Sizling, Stroch, & Keil, 2009). To provide comparable data with published studies (Hawkins & Diniz-Filho, 2006;Morueta-Holme et al., 2013) we also calculated average range sizes after log-transforming the data.

| Environmental predictors
The calculation of the climate change velocity (CCV) followed Sandel et al. (2011) and Loarie et al. (2009), who described CCV as the temporal gradient of temperature change divided by its spatial gradient, resulting in a unit of distance per time. We used mean annual temperatures for current conditions (averages 1950-2000), mid-Holocene (6,000 years before present) and Last Glacial Maximum (LGM; 21,000 years before present). LGM data were derived from two cou- The PMIP2 data has a lower resolution (2.5°), which misses more localized climatic effects caused, for example, by topography. To obtain data at the same resolution as for WorldClim data (2.5 arcmin), we interpolated the raster. The same interpolation was applied to the PMIP2 current temperature data. We then calculated the difference in current temperature between WorldClim and PMIP2 data and added it to the mid-Holocene data from PMIP2 database to account for elevation effects (and other time-stable anomalies). This method (A. Wilson, pers. comm.) follows the assumptions that (1) local adiabatic effects on temperatures have not changed much since the mid-Holocene, and (2) that topographies have remained stable.
To verify this procedure we repeated all steps to calculate a map for the LGM with data from PMIP2 database and obtained a high correlation with WorldClim data (Pearson's r = .987; equivalent Mid-Holocene data are now also available at www.worldclim.org).
We calculated the temporal gradient as the difference between present and past temperatures in each cell after converting all temperatures to Kelvin (K). We converted differences to absolute values to retrieve positive velocities independent of their sign. We calculated the spatial gradient as the slope of the temperature over distance, that is, rate of change for each cell from a 3 9 3 cell neighborhood. Spatial change rates are mainly driven by elevation difference due to the adiabatic relationship of temperature with air pressure. We used the slope of current temperatures because the temperature slopes of past climates correlate very well with these (e.g. current versus LGM rate, Pearson's r = .997). Values <0.01 K/ km were changed to 0.01 K/km to avoid dividing by zero (or nearzero). The temporal change rates (K/year) were then divided by the spatial change rates (K/km) and then multiplied with 1,000 to yield data in units of m/year. We denote the velocity from LGM to present as CCV21, the velocity from the mid-Holocene to the present as CCV6.
To allow for a multidimensional interpretation of climate and CCV (i.e. extending the original hypothesis of temperature change effects towards precipitation changes and other climatic variation), we applied methodology developed by Hamann, Roberts, Barber, Carrol, and Nielsen (2015). This method for predicting CCV effects involves an ordination of climatic data, so it does not allow a direct test of the 'temperature CCV'-hypothesis of Sandel et al. (2011).
However, it assures that other, more complex and multivariate CCV effects within the given time window are not overlooked. As these analyses did not alter our conclusion, we present them in the Appendix (S6).
Temperature seasonality (T seas ) is measured as the standard deviation of monthly mean temperatures throughout the year, and precipitation seasonality (Precip seas ) as its coefficient of variation (data from www.worldclim.org; Hijmans et al., 2005).
We calculated available land area as the area of land cells in a given radius around a cell (Morueta-Holme et al., 2013). As it is somewhat arbitrary what radius is to be used for this calculation, we tested (in preliminary trials) different radii that lead to circles with maximum areas of the lower quartile, median and upper quartile of range sizes. We found that the radius related to the upper quartile of ranges sizes (i.e. 955 km) led to the best model fits and we used the resulting area calculations for further analyses.
All further GIS manipulations and analyses were carried out in Mollweide World equal area projection at 100 km grain size. Climate data, originally processed at 2.5 arcmin, were aggregated and projected to this grid. Pleistocene ice extent (Ehlers et al., 2011) was coded as one (ice) and zero (no ice). Elevation range was calculated from a digital elevation model (Stein et al., 2015). Furthermore, we used a broad classification of zoogeographical realms from Holt et al. (2013).
We restricted the study region in various ways to reduce unwanted variability and bias. First, all smaller islands were excluded to avoid effects of dispersal limitation of island endemics on range data (as these will not contribute to our understanding of the general drivers of range size). Exceptions were made for the British Isles, Sumatra, Borneo, Madagascar and New Guinea, because they are either large enough to develop range size variability within the island, or were connected to continental regions in the relevant past (i.e. LGM). Since the connectivity argument cannot be made for Madagascar (but for all others of the large islands), we also re-run analyses without Madagascar (which did not change conclusions; data not shown).
Second, we excluded cells with a species richness <5, because random effects in the data have great potential to introduce noise into ecological patterns. This restriction affected mostly desert regions in North Africa and Western Australia, as well as much of north-eastern Siberia. Third, we excluded coastal cells to avoid unwanted effects due to area size variation of cells. After applying these restrictions, 762 species continued to contribute to range size data in 7,108 pixels.

| Statistical analyses
All data (predictors and response) were standardized to a mean of zero and a standard deviation of one, which makes model coefficients directly comparable. All variables except land area were log 10transformed prior to standardization to reach normality. We tested predictor data for collinearity, finding that all correlations between variable pairs were weak (r 2 <.26). In a preliminary analysis step, we used model selection (Burnham & Anderson, 2002) to compare a full model with all hypothesized effects (no interactions) to simplified models of various subsets of these predictors. We present both the Bayesian information criterion (BIC) to account for very large sample sizes, as well as Akaike's information criterion (AIC) for comparison.
We based the main analyses on univariate regression of the above predictors, as well as on their combination in a multivariate linear model. Because our dataset contained spatially structured data, ordinary least square (OLS) models are most likely biased in significance assessments and possibly also in coefficient estimates (Bini et al., 2009). Therefore, we also present results from a spatial simultaneous autoregressive error (SAR) model (function errorsarlm in R-package 'spdep', Bivand & Piras, 2015; neighbourhood distance of 5,000 km, based on preliminary trials; residual autocorrelation remained high only over very short distances of <200 km (concluded from correlograms of residuals), which we considered acceptable at our data resolution and extent). By comparing results for OLS and SAR the reader can appreciate the potential effects of spatial structure in our data.
Zoogeographical regions differ in their evolutionary history, but it is unclear to what extent this may affect range sizes (e.g. whether or not range size data carry phylogenetic signal; Jablonski, 2008;Cardillo, 2015). Furthermore, available zonations are based on vertebrate taxa, which may be different to the appropriate (yet unknown) zonation for sphingid moths. For these reasons, we calculated all analyses with and without additional consideration of zoogeographical regions (as binary dummy predictors). Because they led to the same conclusions, we present only models without zoogeographical regions in the main text (see Appendix S4 for inclusion).
Analyses were carried out in R 3.3 (R Core Team, 2016). We present pseudo-R 2 values of the (likelihood-fitted) models, calculated from a linear regression of model prediction versus observed data.

| RESULTS
Estimated range sizes varied over seven orders of magnitude, from 75 to 53.4 9 10 6 km 2 (for raw data see Appendix S1). Range size data resembled a left-skewed log-normal distribution ( Figure S2.1 in Appendix S2), as commonly observed in such data (Gaston, 2003); there are many more small-ranged than large-ranged species. the research region. There is a clear north-south decline in range sizes (Pearson's r = .656), while a correlation of range size with absolute latitude, as expected by Rapoport's effect (Stevens, 1989), is weak (Pearson's r < .225).

| Environmental correlates of range size variation
Patterns of all environmental predictors are mapped in Figure 2. Climate change velocities derived for the period from LGM to the present (CCV21) exhibit a very different pattern to that from the mid-Holocene to the present (CCV6). Likewise, temperature seasonality is distributed differently than precipitation seasonality. Univariate models with median range sizes ( Figure 3, Table 1) indicate strong positive effects of temperature seasonality and land area, slightly weaker, negative correlations with elevation range, and almost no effect of precipitation seasonality. CCV measures are relatively weak and inconsistent in strength (positive for CCV21, negative for CCV6; note that this is not due to opposite temperature gradients, as velocities are based on absolute change).
Model selection based on AIC as well as BIC (Appendix S3) showed that models containing all (or nearly all) predictors were most useful for predicting range size variability. In particular, the full model (7 predictors) was considered best if biogeographical regions were not included (DAIC = 16, DBIC >8 to second-best model; Table S3.1). For models including biogeographical regions as predictors (Table S3.2), all top-models had 5-7 predictors (additional to biogeographical regions), whereas the full model was ranked second (DAIC = 1.7) or third (DBIC = 8.6), depending on the information criterion. For consistency among the following analyses, we therefore chose to always use the full models for in-depth assessments of predictor effect. Table 2 shows results for multivariate models containing all predictor variables. Both modelling approaches (spatial and non-spatial) confirmed strong positive effects of temperature seasonality and land area, and negative effects of elevation range. Positive effects of LGM ice extent were weaker but consistent, whereas we could neither find unequivocal support for partial effects of precipitation seasonality, nor for both CCV measures. Relatively weak effects of CCV21, in particular, changed sign depending on whether OLS or  All modelling was carried out on standardized data (except "Ice", a binary variable); sample size N = 7,108 pixels, grain size = 100 km. All ordinary least square (OLS) regressions were highly significant (not shown). Spatial autoregressive models (SAR) were used to control for autocorrelation. Pseudo-R 2 values for SAR were based on a regression of modelled versus observed data.
assessments. Using a multivariate metric of CCV (Hamann et al., 2015), based on six climatic variables, did not alter our conclusion of weak CCV effects compared to those of temperature seasonality alone (Appendix S6). Range size patterns were not related to biome area sizes ( Figure S7.2 in Appendix S7), hence biome size does not provide an alternative, arguably more parsimonious explanation of range size patterns.

| DISCUSSION
Our data indicated that, for sphingid moths within the geographical restrictions of our analyses (i.e. continental Old World and Australia), current intra-annual temperature variability explains statistically the spatial variation of geographical range sizes much better than longerterm variation as captured by CCV since the LGM or mid-Holocene.
Available land area and elevation range proved important covariates in the system, whereas Pleistocene ice extent had relatively low impact across the research region and precipitation seasonality apparently played no role.

| Temperature seasonality fits better than CCV
Unlike earlier studies (e.g. Sandel et al., 2011), we did not investigate the CCV-range size link in isolation. If we had, we would have concluded a moderately positive effect (Table 1). By comparing CCV against effects of other hypothetical drivers of range size variation, in univariate and multivariate models, we can evaluate more fully the empirical support for CCV as a mechanism shaping range size distributions. Even though broad spatial patterns of temperature seasonality and CCV resemble each other (Figure 2), collinearity should not have seriously biased analyses (e.g. T seas versus CCV21, Pearson's r = .382). Our data suggest that hypothetical mechanisms acting through adaptations to current climates (i.e. seasonality, Janzen, 1967) are better-supported explanations of range size patterns than those that invoke climatic dynamics of the past (i.e. CVV). This view is also suggested by the inconsistent direction of effects of LGMand mid-Holocene CCV in our models (but see discussion below).
However, our analyses carry the caveat that current temperature seasonality may be correlated to climate variation (hence, CVV) at an unspecified point in the past. Thus, statistical support for seasonality does not rule out more complex causal pathways-it only rejects CCV effects as tested (i.e. temperature during the two tested time periods).
Temperature variation between the LGM and the present is one of the strongest climatic changes of the Quaternary (Ruddiman, 2001). However, temperatures did not change linearly (as implied by CCV calculations), but included many smaller shifts and oscillations, as evident from locally studied stable isotopes from ice cores, or from pollen records (Thompson et al., 1998;Claussen et al., 1999;Davis, Brewer, Stevenson, & Guiot, 2003). During the mid-Holocene, temperatures in some areas (e.g. northern Europe) were higher than today (Davis et al., 2003). We would have expected that CCV calculations of both time periods should have similar effects on range sizes if the velocity, not the direction, of climate change mattered.
Negative links indicate that high velocity regions are associated with small range size. In the absence of reasonable ecological interpretation, this is possibly a spurious finding. Model misspecification is always a possibility with messy ecological data (in this study and others). Furthermore, climatic variation since the Holocene was of smaller magnitude than that since the Pleistocene, so CCV21 effects may have overridden CCV6 effects in some parts of the world, leading to unclear patterns. Univariate models (Table 1) showed that CCV since the mid-Holocene had only a very low explanatory power as a single variable. This supports the assessment that temperature change velocity since the mid-Holocene did not influence species range sizes. We had also considered CCV effects from the LGM to the mid-Holocene (not shown), which did not lead to further insights. In conclusion, finding consistent effects of CCV21 and CCV6 would have strengthened the case for the proposed mechanism of CCV acting through selection of species' migration speed and mobility, or their niche breadth. Not finding them in our analyses, however, may be due to a range of methodological issues that do not allow clear inference.
The correlation with current seasonality, however, does not rule out the possibility that seasonality patterns of past times shaped range sizes (as the seasonal pattern did not change much through time; for example, WorldClim LGM seasonality versus current seasonality, Pearson's r = .999). However, we find it intriguing that range size effects of long-term climatic variability can be theoretically explained in an elegant manner as the outcome of selection for mobile, wide-niched taxa (Dynesius & Jansson, 2000;Sandel et al., 2011), while the exact mechanism behind a seasonality effect, which we support here empirically, is somewhat unclear (Janzen, 1967;Stevens, 1989Stevens, , 1992McCain & Bracy Knight, 2013). A combination of physiological niche measures and spatially explicit evolutionary modelling may be useful to disentangle the various pathways of how adaptation to high local, or range-wide, climate variability may lead to wider niches and larger geographical ranges (Gaston, 2003). All modelling was carried out on standardized data (sample size N = 7,108 pixels, grain size 100 km). Ordinary least square (OLS) model fit was R 2 adj = .571, SAR had a pseudo-R 2 = .691.
While the data in this study reject the CCV hypothesis in the tested timeframes, it may be argued that these were not appropriate to the evolutionary history, migration ability, generation length or other biological traits of the studied taxon. For example, high mobility in sphingids may have led to new equilibria much faster after climatic disturbance than, for example, in poorly dispersing amphibians (Sandel et al., 2011). Thus, our results cannot reject the general idea that CCV at any, unspecified time window had effects on today's range size distribution. However, without an a priori hypothesis on a specific, appropriate time window, rigorous scientific testing is impossible (we are not aware of any specific CCV hypothesis for alternative timeframes, for sphingids or any other taxon). Data-mining for links between any CCV and range size data for a given taxon may give interesting exploratory clues to relevant drivers, but this cannot be viewed as hypothesis testing (see Forstmeier, Wagenmakers, & Parker, 2016, for a general critique of post hoc 'testing' in biological science).

| Available habitat area matters
Habitat area, as pointed out by Morueta-Holme et al. (2013), is an important pre-condition for the development of species range sizes.
Without land, there is no potential for expansion in terrestrial species. This effect is strong and obvious on small, isolated islands, where many endemics are typically found. However, after excluding these from our analyses we still recovered relatively strong land area effects on median range sizes (Tables 1 & 2). Many small-ranged species in Madagascar, New Guinea and Australia's east, in particular, are associated with small areas of available land in the vicinity.
Land area, however, is only a crude proxy for suitable habitat. We can expect that the availability of homogeneous, suitable habitat Habitat rarity is also one (of several) potential explanations for the effects of elevation range. Highlands have smaller areas than lowlands and, as there tends to be taxonomic turnover from lowland to highland regions (for sphingids: Beck, Holloway, Chey, & Kitching, 2012), highlands will contain species adapted to those rare habitats. Additionally or alternatively, elevation gradients may act as dispersal barriers or ecotones that facilitate speciation (Doebeli & Dieckmann, 2003). Highlands may therefore be associated with the presence of young, yet small-ranged taxa. Furthermore, elevation gradients act as buffers to climatic change (Hawkins & Diniz-Filho, 2006). The latter effect is essentially the suggested mechanism of CCV, as climatic stability (due to easy migration up and down a mountain) would lower extinction rates and gives small-ranged species a higher chance to survive (Burgess et al., 2007). In line with this, in other taxa phylogenies (Smith, de Oca, Reeder, & Wiens, 2007) and richness patterns (Colwell, Brehm, Cardelus, Gilman, & Longino, 2008) on mountains seem to support the idea of highest survivability at mid-elevations on mountains. Elevation range of grid cells was a very weak univariate predictor of range sizes, but had a strong effect in multivariate models. Thus, while an additional effect of mountains on range size is evident (irrespectively of the mechanism) it is not a factor that can serve as a main determinant of the global-scale pattern (given that much variability occurs also across lowland regions, Figure 1).

| Ice cover and precipitation
The extent of the glaciation is an effect of Pleistocene history that goes beyond temperature effects, as it determines the available land area for all taxa than depend on plant growth. Glaciation history undoubtedly affects species richness and composition in Europe and in particular in North America, where glaciation was more extensive (Morueta-Holme et al., 2013). By adding Pleistocene ice extent as a separate predictor to our analysis, we recovered consistent, although not particularly strong effects in the multivariate model ( Table 2).
Given that sphingid moths are generally very mobile, that extensive glaciation was restricted to northern Europe, and that southern European species are also relatively wide-ranging (Figure 1), it is perhaps not surprising that the ice effect was not overly strong. However, glaciation history, in combination with high CCV, may be a reason for different latitudinal range size clines in Western Europe compared to East Asia (Figure 1), a pattern also evident in data from Sandel et al. (2011). Pleistocene refuge areas, such as Iberia, Italy and the Balkans (Hewitt, 1999;Sommer & Nadachowski, 2006), had clearly lower CCV (Figure 2).
A surprising result was the apparent irrelevance of precipitation seasonality in explaining range size variation. The mechanisms suggested for effects of temperature seasonality should also be relevant for precipitation, and the map of range size variation suggests higher values in low-precipitation regions at least in the subtropics and tropics (e.g. fringes of Sahara, Namib, Australian deserts). Although there are options for artefacts-for example, niche modelling may have an inherent tendency to overestimate the range filling (or occupancy) of desert species (who may be dependent on water sources other than precipitation, unknown to the niche models), and Worldclim precipitation data may lack precision in tropical regions-we find it surprising that this absence of a precipitation effect has so far not been a topic of the scientific discourse.
Our multivariate model explained a substantial part of the nearglobal range size variation studied here (OLS: 57%, SAR: 69%; Table 2) from only a few environmental correlates, and results clearly supported some variables while deeming others irrelevant. Nevertheless, statistical as well as principal issues remain to be solved before we can optimistically claim to understand how climate and other factors shape range sizes and endemism. For example, large-ranged species generally contribute overly to pixel-based analyses (a phenomenon of pseudo-replication; Jetz & Rahbek, 2002), but it is far from trivial to overcome this effect. Sizling et al. (2009) pointed out how geometric effects alone can lead to (in parts) counter-intuitive patterns of range size and species richness. Furthermore, phylogeny may link species' occurrences (i.e. closely related taxa tend to occur in nearby regions) with their range sizes (Beck, Kitching, & Linsenmair, 2006;Cardillo, 2015;Jablonski, 2008). It is not straightforward to control analyses simultaneously for spatial and phylogenetic effects of non-independent data. Last, and most important in our view, unclear ideas on mechanisms lead to vague hypothesis predictions, which reduces the inference value of tests. This highlights the need to investigate more thoroughly how seasonality affects niche evolution, and what testable predictions can be derived from that.

| CONCLUSIONS
We found a distinct spatial pattern of range size variation that does not conform to Rapoport's effect, but showed an across-tropics North-South pattern (cf. Di Marco & Santini, 2015). This fits with the long-standing observation that northern hemisphere studies tend to find support for a Rapoport pattern while southern hemisphere studies do not . Our findings confirmed that regions directly or indirectly associated with high climatic instability selected for species with large range sizes. However, among variables of climatic instability, temperature seasonality was the strongest empirical predictor of the range size distribution, while measures of CCV received much weaker support. This illustrates the inference value of testing competing hypotheses in comparison to each other (McGill, 2003).
Although our models explained a substantial proportion of the measured range variability across a near-global study extent, we see need for caution. Without deeper insights (e.g. from physiology and evolutionary modelling) into evolutionary mechanisms of how niche evolution responds to climatic variability (e.g. seasonality), it is difficult to move from statistical pattern search towards true testing of mechanistic hypotheses.