A Surfeit of Studies: What Have We Learned from All the Box Turtle ( Terrapene carolina and T. ornata ) Home Range Studies?

: Home range (HR) studies are a particularly common approach to investigations of animal habitat use, resource availability, and response to management manipulation such as relocations. Terrapene carolina (Eastern box turtle) and its sister taxon T. ornata (Ornate box turtle) are especially popular subjects of HR studies because they are relatively easily tracked. Terrapene HR studies have revealed a wide variation in HR sizes within and between populations, due to factors such as di ﬀ erences in ecoregion and analytical approach (e.g., minimum convex polygons, kernel analysis, bivariate normal, multivariate Ornstein–Uhlenbeck stochastic process, harmonic means). We performed a meta-analysis of the available literature, including unpublished work to avoid bias due to under-publication, to explore the causes for variation in HR size. We found 19 studies reporting T. carolina HR sizes and seven studies reporting T. ornata HR sizes; the resulting meta-analysis revealed patterns that are not visible in the individual studies. We found important di ﬀ erences between the species: female T. ornata had smaller HRs than males, whereas the opposite is true for T. carolina , and T. ornata HRs were inﬂuenced by ecoregion, while T. carolina HRs were not similarly inﬂuenced. Not surprisingly, we found that choice of analysis technique a ﬀ ected HR estimate; analyses using ellipses resulted in larger HR estimates than all the other techniques, while kernels were smaller than minimum convex polygons. Although not indicated by individual studies, our meta-analysis showed that the HRs of relocated T. carolina females were signiﬁcantly larger than those of non-relocated females. Although the number of individual turtles in studies varied from three to 25, the sample size did not signiﬁcantly a ﬀ ect HR size.


Introduction
Home range (HR) studies have a long, complex, and contentious history, in part because the concept of home range is attractive yet ambiguous. One of the earliest HR studies defined HR as "the area an individual uses to carry out its life functions" (e.g., feeding, mating, overwintering, etc.) [1]. This broad definition, along with the lack of a generally accepted subsequent narrowing of this definition, has led to a wide range of uses for the term HR, which is reflected in the varied methods that are used for calculating home ranges. For example, Kernohan et al. [2] described HR as an area where an individual is most likely to be present during a given period of time, which is a probability density also referred to as the utilization distribution (UD) [3]. Some studies attempt to include all the especially sensitive to data with low sample size and autocorrelation, often inflating the HR estimate as compared to minimum convex polygon techniques [16,[18][19][20]. However, MCP estimates lack information on centers of activity that other estimators provide, including kernel density, harmonic mean, and 95% ellipses. Dodd [7] listed 10 T. carolina HR studies and nine T. ornata HR studies, but there has never been a systematic review, including all the studies with HR data, to quantitatively compare what is known about HR size for these well-studied species. We conducted a meta-analysis of data (Table S1) from these papers to quantify the effect of sex, latitude, sample size, tracking method, analysis method, and habitat on T. carolina and T. ornata HR size. We also considered the effects caused by relocation, which is a conservation technique that is often used with Terrapene.

Method-Meta-Analysis Outline
In 2014, we used a variety of approaches to locate studies that reported HR for either T. carolina or T. ornata. We used the terms and various combinations of the terms "terrapene", "carolina", "ornata", "home range", and "homerange" using Google Scholar. We obtained all papers cited as containing home range data from Dodd [7]. We also searched for PhD dissertations and master's theses in the ProQuest Dissertations and Theses database, and searched for published papers that resulted from theses and dissertations. Where theses and dissertations were not published, we used the original studies. We examined the literature cited sections of all the studies that we found for additional studies.
We selected all the studies that reported HR for either adult T. carolina or adult T. ornata and provided a mean, variance, and sample size, or the data for which these could be calculated. For two studies [21,22], we emailed authors to clarify and supplement data. We used mean estimates of home range data as reported, or we calculated these from individual data as reported. Where other forms of variance were reported (e.g., standard error), we converted them to standard deviation for analysis. Study sites were found on Google Maps, and the GPS coordinates for each site were plotted in ArcMap. We used the Level II Ecoregion shapefiles (EPA.gov) [23] to identify the ecoregion for each study. Sex was used as reported: "male", "female", or "mixed" where sex specific data was not specified. We attempted to collect the appropriate data to test for the effect of length of study and number of locations/individual.
We also noted whether turtle locations were determined through radio telemetry, visual search, or string trailers. However, all the studies with appropriate data used radio telemetry, so this factor was left out of the analysis. We placed all the HR studies into one of four categories based on analytical technique: kernel (95% kernel), minimum convex polygon (minimum convex polygon, 95% minimum convex polygon, minimum area, minimum polygon, or convex polygon), ellipse (95% bivariate normal or 95% ellipse), and harmonic mean (harmonic mean or 95% harmonic mean). If a paper reported results for more than one gender, location, or analytical technique, we included all the HR estimates and controlled for non-independence by including a random term (Author/paper ID) in the model.

Meta-Analysis
All the calculations and statistical analyses were executed using functions included in the metafor package (version 2.0-0; [24]) of the R open-source software environment (version 3.4.1; [25]) within RStudio (version 1.0.153; [26]). We provide all the data, codes, and output used in this experiment as digital appendices [see Supporting Information]. We used the raw means, standard deviations, and sample sizes of Terrapene home range estimates that were extracted from the literature to calculate the sampling variance needed for weighting the outcomes in the meta-analysis. We used the measure = "MN" argument within the escalc() function for this purpose. Unlike many traditional meta-analytical effect sizes such as Hedge's d or log odds, the raw mean home range (ha) is retained as the effect size response rather than a zero-centered and/or transformed response.
Using restricted maximum likelihood (REML) estimation in the function rma.mv(), we ran multi-level random-effects (MLRE) models to understand relationships between Terrapene HR sizes and intrinsic or extrinsic factors. Many factors (e.g., population genetics, researcher error, and climate) are more likely to be similar within than across experiments and locations. MLRE-type meta-analytical models are specifically designed to account for non-independence among effect sizes by allowing for the addition of a random term in the model [24,27]. We included "Author" as a random term in each model to account for the potential lack of independence among effect sizes derived from the same article and/or the same experimental infrastructure. Where appropriate, we used the glht() function in the multicomp package (version 1.4-8; [28]) to perform post hoc multiple comparisons using the Holm-Bonferroni method. Heterogeneity among effect sizes within datasets was assessed using the Q-statistic. Large Q values suggest that differences between effect sizes within a dataset do not estimate a common population mean, and thus could vary for reasons other than sampling error (e.g., due to some environmental factor; [29]). We tested for the significance of Q using a χ 2 distribution.
We tested for publication bias using Egger's regression test [30,31] by modifying the MLRE models to include the standard error of the effect sizes as a moderator. When the intercept of this regression test significantly deviates from zero, the overall relationship between the precision and size of studies included in the dataset is considered asymmetrical, and therefore biased [31]. We considered analyses to be biased if the intercept differed from zero at p = 0.10 (as in [30]).
Parameter estimates in meta-analyses, similar to other regression-based tests, are sensitive to outliers and influential data points. However, diagnostic tests for identifying-and rules for excluding-these types of cases are still evolving, particularly for multivariate/multi-level meta-analytical models [32]. We evaluated the sensitivity of our analyses as in Habeck and Schultz (2015) [33] by comparing fitted models with and without effect sizes that we defined as influential outliers. We considered influential outliers to be effect sizes with hat values (i.e., diagonal elements of the hat matrix) greater than twice the average hat value (i.e., influential) and standardized residual values exceeding 3.0 (i.e., outliers; [32,34,35]). Finally, we used the function lme in the package nlme (version 3.1-131.1; [36]) to test whether the sample size of experiments influenced Terrapene HR sizes.

Data Set
We found 23 publications that reported T. carolina HR estimates, and eight publications that reported T. ornata HR estimates. Five of these did not report sufficiently detailed HR data. We extracted 72 Terrapene HR estimates (T. carolina = 50, T. ornata = 22) from the remaining 26 publications ( [21,22,). The original experiments were conducted within 10 Level II ecoregions as defined by Omernik [61] and the U.S. Environmental Protection Agency. Only 42 (58.3%) of the HR estimates were from published studies, 22 (30.6%) were from unpublished undergraduate and master's theses or PhD dissertations, and eight (11.1%) were from unpublished government agency reports. All the studies that met our criteria for HR mean, variance, and sample size, or the data for which these could be calculated, used radio telemetry as their tracking technique. We were unable to collect meaningful data to test for the effect of length of study or number of locations/individual because very few publications reported these data in sufficient detail.

Analysis
Previously reported T. carolina HR sizes ranged from 0.205 ha (males only, [54]) to 28.370 ha (mixed sexes, [50]), which is a factor difference of 138. The overall mean of the 50 studies, including those in which turtle sex was identified and those in which sex was not identified, was 7.54 ha (95% CI = 5.86 to 9.21 ha). Previously reported T. ornata HR sizes ranged from 0.825 ha (females only, [49]) to 19.097 ha (females only, [40]), which is a factor difference of 23. The overall mean of the 22 reports included was 4.94 ha (95% CI = 2.69 to 7.19 ha).
Home range size did not differ between T. carolina and T. ornata (Table 1). However, within species, home range size differed between males and females (Table 1). For T. carolina, the female home range was 27% larger than that of the males. For T. ornata, the female home range was 6% smaller than that of the males. Table 1. Raw mean home range (HR) estimates with standard error (SE) and 95% confidence intervals (CI.L (=lower limit of confidence interval) and CI.U (=upper limit of confidence interval)) for T. carolina and T. ornata by species and sex, using only studies where turtle sex was identified. Superscript letters show within-species differences between females and males, based on Holm-Bonferroni pairwise post hoc comparisons. For T. carolina, HR size did not differ among ecoregions ( Table 2). The HR size of T. ornata within the west-central semi-arid prairies and western Sierra Madre piedmont ecoregions differed from all other ecoregions (Table 3). HR size differed among home range estimator techniques. HR estimates calculated using the ellipse technique were larger than those calculated using all other techniques, and MCP estimates differed from kernel estimates (Table 4). The number of turtles in studies varied from three to 25 individuals (mean = 9.85, s.d. = 7.54). The sample size in experiments did not influence either Terrapene HR size (t = 0.54, df = 45, p = 0.59, Figure 1) or the s.d. of HR (β = −0.061, t = −0.316, df = 45, p = 0.754, Figure 2). The number of turtles in studies varied from three to 25 individuals (mean = 9.85, s.d. = 7.54). The sample size in experiments did not influence either Terrapene HR size (t = 0.54, df = 45, p = 0.59, Figure 1A) or the s.d. of HR (β = −0.061, t = −0.316, df = 45, p = 0.754, Figure 1B). Relocation status influenced HR size in T. carolina, and there were no appropriate data for T. ornata. Individuals that were relocated had a HR that was on average 50% larger than individuals that were not relocated (Table 5). Relocation status influenced HR size in T. carolina, and there were no appropriate data for T. ornata. Individuals that were relocated had a HR that was on average 50% larger than individuals that were not relocated (Table 5).

HR of T. Carolina and T. Ornata
In agreement with our expectations, HR did not differ between T. carolina and T. ornata, which was due both to their similar mean HR values and to a wide HR variation within both species ( Table 2). The sister taxa are ecologically similar, although adult T. carolina are typically slightly larger than T. ornata [12]. The species have broad differences in typical habitat: T. carolina primarily inhabits woodlands, and T. ornata primarily inhabits prairies. However, both are habitat generalists and broadly omnivorous. They are sympatric over eastern Texas and Oklahoma and parts of Missouri and Illinois, where they can be found in the same microhabitat and sometimes hybridize [62][63][64].

Do the HRs of Male and Female Terrapene Differ in size?
Males have larger HRs than females in many species of terrestrial turtles ([13], Lawson 2006); in some species, the female HR is larger (e.g., [15]), and in some species, there is no difference [65]. In contrast, some T. carolina HR studies have found that females have larger HRs than males (e.g., [54]), while others have found that males have larger HRs than females (e.g., [49]). Dodd [7] reviewed Terrapene HR studies and concluded that male and female Terrapene sp. HRs are "similar in size". However, we found that within species, the HR size differed significantly between males and females ( Table 2). In T. carolina, the average female HR was 27% larger than that of males, while for T. ornata, the average female home range was 6% smaller than that of males. This surprising difference between these closely related species has not been previously detected, and should be investigated where they are in sympatry and allopatry.

Is HR Influenced by Ecoregion?
Numerous researchers have found that the HRs of some species are influenced by resource availability (e.g., olive baboon (Papio anubis): Barton et al. [66]; grizzly bears (Ursus arctos): Edwards et al. 2013 [67]; African striped mice (Rhabdomys pumilio): Schoepf et al. 2015 [68]). Dodd [7] suggested that local resource levels and habitat structure might influence Terrapene HR. These data were not available to us, so we used ecoregions as crude proxies. While we found no significant difference in T. carolina HR among the five ecoregions they inhabit (Table 3), we note that the average T. carolina HR in the central USA plains ecoregion is 3.6 times larger than that of the ecoregion with the next highest average T. carolina HR. However, this unusually high HR is because of a single T. carolina HR study in the central USA plains ecoregion [58]. In contrast, the average T. carolina HR in the other four ecoregions was very similar.
In contrast, the average T. ornata HR size in the west-central semi-arid prairies and Western Sierra Madre piedmont ecoregions was significantly larger than that in the other four ecoregions for which we had data ( Table 4). The mean HR for the west-central semi-arid prairies ecoregion was exceptionally large, 2.7 times larger than that of the western Sierra Madre piedmont ecoregion. This again was the result of a single HR study from that ecoregion [56].

How Does Choice of Analysis Technique Affect HR Estimate?
The analytical approaches that are used for HR size estimation have gone through numerous changes since Burt [1]. There has been vigorous discussion over the optimal approach to the analysis of individual location points (e.g., [20,[69][70][71]). Since Terrapene have been the subjects of numerous HR studies since the 1920s, Terrapene HRs have been estimated using a wide variety of techniques. We combined Terrapene HR studies into four major categories (kernel, minimum convex polygon, ellipse, and harmonic mean), and found that HR studies that used ellipse techniques had significantly larger HRs than those that used any other techniques, and further, that MCP estimates indicated, on average, significantly larger HRs than kernel estimates (Table 5). These results show that studies using different HR analysis techniques cannot be combined in a single analysis without adjusting for these differences.

Does Sample Size Affect HR Size Estimation?
We hypothesized that sample size (number of turtles) would affect the HR size and the standard deviation of HR size, because small samples sizes can increase the relative effect of individuals with unusual HR. We found that differences in sample sizes among experiments did not influence mean HR size (Figure 1) or the standard deviation of HR sizes (Figure 2). This suggests that typical sample sizes (mean = 7.85 individuals) for Terrapene HR studies are adequate. We were unable to test for the effect of length of study and number of locations/individual because these data were rarely reported. This is unfortunate because, despite the relative ease of Terrapene HR studies, resource use could be optimized if these aspects of study design were better understood. The analytical approaches that are used for HR size estimation have gone through numerous changes since Burt [1]. There has been vigorous discussion over the optimal approach to the analysis of individual location points (e.g., [20,[69][70][71]). Since Terrapene have been the subjects of numerous HR studies since the 1920s, Terrapene HRs have been estimated using a wide variety of techniques. We combined Terrapene HR studies into four major categories (kernel, minimum convex polygon, ellipse, and harmonic mean), and found that HR studies that used ellipse techniques had significantly larger HRs than those that used any other techniques, and further, that MCP estimates indicated, on average, significantly larger HRs than kernel estimates (Table 5). These results show that studies using different HR analysis techniques cannot be combined in a single analysis without adjusting for these differences.

Does Sample Size Affect HR Size Estimation?
We hypothesized that sample size (number of turtles) would affect the HR size and the standard deviation of HR size, because small samples sizes can increase the relative effect of individuals with unusual HR. We found that differences in sample sizes among experiments did not influence mean HR size (Figure 1) or the standard deviation of HR sizes (Figure 2). This suggests that typical sample sizes (mean = 7.85 individuals) for Terrapene HR studies are adequate. We were unable to test for the effect of length of study and number of locations/individual because these data were rarely reported. This is unfortunate because, despite the relative ease of Terrapene HR studies, resource use could be optimized if these aspects of study design were better understood.

Do Relocated Terrapene Have Larger HR?
Studies of relocated turtles often attempt to test whether they have a HR size that is different from non-relocated turtles (e.g., [51,[72][73][74]). Testing for this relationship is often difficult, because

Do Relocated Terrapene Have Larger HR?
Studies of relocated turtles often attempt to test whether they have a HR size that is different from non-relocated turtles (e.g., [51,[72][73][74]). Testing for this relationship is often difficult, because usually only small numbers of relocated individuals and resident individuals can be tracked in any single study. Typically, the HRs of both relocated and non-relocated turtles are highly variable, raising the probability of Type 2 errors. Meta-analysis can address both of these problems. We only found appropriate data for relocated T. carolina, which had considerably higher HR than those that had not Diversity 2019, 11, 68 9 of 12 been relocated. We recognize that this is a fairly crude analysis, because a better analysis would be to compare the HRs of relocated turtles to the HRs of local conspecifics that were not relocated. However, this test clearly shows a dramatic difference due to relocation, which is an effect that had not been shown consistently in the individual studies.

Conclusions
Individual HR studies of T. carolina and T. ornata contributed a more thorough understanding of the habitat use and behaviors of individuals within local populations. Combining these studies within a meta-analysis allowed us to reveal larger patterns in HR across T. carolina and T. ornata populations, and between these two species, patterns that were often not detected in individual studies. For example, meta-analysis revealed an increase in HR in relocated T. carolina compared to those not relocated. This analysis also revealed species-level differences regarding the influence of sex and ecoregion on Terrapene HR, with sex affecting the HR in both species but in opposite directions, and ecoregion playing a larger role in T. ornata HR than T. carolina HR. Furthermore, our meta-analysis revealed that while tracking technique (number of locations) did not affect HR, the HR estimators (MCP, ellipse, etc.) did affect HR. We note that we were unable to test whether the study length or number of locations/individual affected HR size; however, we suggest that the raw data from an unusually long-term study with very large sample sizes (e.g., [21,45]) or novel studies using current GPS-tracking technology create samples that vary frequently and with high accuracy (e.g., [75,76]) could be sub-sampled to address these questions, as well as to address important issues of spatial autocorrelation. As shown in this study, meta-analysis is a powerful tool that can detect large-scale patterns across studies and provide evidence for or against conclusions made about species within individual studies.

Conflicts of Interest:
The authors declare no conflict of interest.