We suggest that partitioning data in a biologically meaningful way (as opposed to geographically) can help to overcome high omission rates in distribution models of widespread species [17, 18, 28]. Despite having a relatively high AUC value (0.899), the model built upon the whole species' distribution failed to predict known localities of P. polionotus, which can be observed by comparing Figure 1 to Figure 2a. Spatially partitioning the data into quadrants produced a much better distribution model after combining four regional models with AUC values ranging from 0.844 to 0.993, which can be seen by comparing Figure 1 to Figure 2b. The 14 AUC scores for the SDMs based on subspecies partitioning ranged from 0.851 to 1.0, which is not substantially different from AUC values obtained for the quadrant models. However, the accuracy and increased resolution of the composite of the subspecies models can be seen in comparing the three panels of Figure 2 and by comparing Figure 1 to Figure 2c.
Limitation of AUC values for assessing predictive performance
AUC values are commonly used as indicators of model fit [9, 17, 18, 27], and high values for all three methods in our study would suggest that each method produced highly accurate models. Furthermore, the modest increase in AUC values with ever greater data partitioning would suggest that each successive partitioning scheme produced, at best, only slightly better fitting models. However, this finding is misleading when one compares the predicted distributions to the known distribution for the species. The trend of increasing AUC scores may indicate the direction of change in accuracy, but it fails to capture the magnitude of improvement in the predicted distributions of the quadrant composite method and the subspecies composite method. This failure is, in part, due to the fact that the AUC scores for the geographic quadrant method and subspecies method are composites of 4 and 14 combined models (respectively), and the accuracy (or inaccuracy) of individual models is compounded when they are combined. This fact alone, however, cannot fully explain the observed discrepancy between the vastly improved model prediction and the modestly better AUC scores.
AUC values can be misleading when assessing a model's predictive ability for several reasons. The AUC measures discrimination and not accuracy per se, thus ignoring the goodness of fit of a model . The AUC value also takes into account the performance of the model at the extreme left (as well as the right) of the ROC curve (see  for a details), a region that is not operationally meaningful in our case. We are only interested in thresholds of predicted probability of occurrence greater than 0.50 because that would equal the probability of occurrence of a null model. This inclusion of the area under the extreme left of the ROC curve can inflate AUC values, which can be further inflated when the total geographic extent of the model is considered. If the ratio between areas of presence and the total extent is high, true positives are more likely to occur by chance alone . Because this ratio changes with each of the individual models built on different data partitions, AUC values may not be useful in accurately comparing relative model performance between or among our subspecies and regional models.
It is also possible that the inflation of AUC values observed in the models we present results from the interaction between geographic and environmental space. Because of the narrow geographic space at which most subspecies of P. polionotus occur, locality information is geographically clustrered; therefore, the environmental space sampled by the models may show spatial autocorrelation in some of the environmental variables used. In SDM, spatial autocorrelation occurs when the values of the variables sampled at nearby locations are not independent from each other  and as a result, measures of accuracy (e.g. AUC) can be inflated [31, 32]. In our case, the geographic clustering of narrowly distributed subspecies of P. polionotus may cause spatial autocorrelation and thus inflate AUC values (see Figure 4 and 5). Nevertheless, within these narrow extents, we included samples spanning the entire geographic space representing a significant portion of the environmental space occupied by each subspecies. The resulting models are accurate to the true distribution of the subspecies and are able to detect even subtle local environmental conditions likely affecting each subspecies differently, despite the seemingly geographic clustering. This further emphasizes the point that AUC values provide an unreliable way to accurately compare relative model performance. The only exceptions to the issue of inflated AUC values in our dataset are P. p. colemani (AUC = 0.917) and P. p. polionotus (AUC = 0.851), which are the only two widely distributed subspecies spanning a more heterogeneous environmental space where locality information for the subspecies is not geographically clustered (Figure 2). Because of the larger geographic space occupied by these subspecies, the resulting models from P. p. colemani and P. p. polionotus are unlikely affected by spatial autocorrelation and therefore do not show inflated AUC values.
Finally, the AUC does not provide information as to the spatial distribution of errors. It also weighs omission and commission errors equally, both of which vary in interpretive meaning and importance with the intended use of the model . Because we do not have true absence data, we cannot quantify our commission error rate. However, the omission curve shows how well the model performs at different thresholds (i.e. the distribution of omission errors). Therefore, the omission curve can be as important as the AUC value in terms of assessing model performance, if not more-so. A model with relatively lower omission errors at higher predicted probabilities of occurrence is preferred.
Quadrant versus subspecies partitioning
Because Maxent draws pseudo-absence data at random to calculate AUC scores, it is possible that in our quadrant analysis it drew false pseudo-absences from areas outside the quadrant being tested, especially in the case of the northwest quadrant. Whether this occurred, and if it did, whether it contributed to the observed underprediction is debatable. First, there was far less underprediction in the other three quadrants. For example, the model for the southwest quadrant, which includes peninsular Florida predicted occurrences in regions where the species does not occur. Second, the two northern quadrants cover roughly the same geographic area as two of the subspecies (P. p. colemani and P. p. polionotus), yet the two northern quadrants failed to predict areas of known occurrence that the two subspecies models predicted accurately. It is possible that poor quadrant-based models resulted because pseudo-absences were being generated by Maxent in areas with true presences for either of the two subspecies just outside of the quadrant being modeled. That is, the arbitrary boundary between the quadrants obfuscates biologically meaningful boundaries between populations or subspecies. Thus, in using quadrant-based partitioning, niche information was lost for the two subspecies, which emphasizes our point that biologically relevant data partitioning informs species distribution models.
Molecular data and population distribution models
It is easy to see how molecular data, such as DNA sequences, can be used to delineate biologically meaningful groups (i.e. clades) within a species, and that those clades might be partitioned separately for species distribution modeling, especially if they are geographically discrete. But, just as molecular data can improve methods of generating SDMs, the findings associated with SDMs can also inform the work done by molecular biologists studying population genetics, phylogenetics, or phylogeography. When SDMs are nonoverlapping for populations within a species, they may be revealing cryptic patterns of divergence that would be interesting to study with molecular data. Conversely, when molecular data uncover population structure or limits to gene flow, SDMs can be used to test hypothesized mechanisms of divergence such as niche differentiation. Examining both molecular data and SDMs together has been explored only recently [6, 9].
Building SDMs for P. polionotus by partitioning data into subspecies and building a composite distribution model mitigated the problem of high omission rates that usually occurs when modeling the distributions of widely distributed species. This suggests that the SDM based on biologically relevant partitions (subspecies in our case) could accommodate variability in the niches of subspecies, whereas modeling the whole species distribution together could not. This is supported by the fact that spatial partitioning of data into quadrants produced models that had regions of both under- and over-prediction (Figure 3), whereas the models based on partitioning by subspecies showed no signs of underprediction and only modest overlap in the distributions of adjacent individual subspecies caused by overprediction (Figure 5). The evidence of high levels of population structure between locally adapted populations [19, 21–23, 25, 26] might be driving the improvement we see in the composite model based on subspecies distributions. In our case, we show that molecular data at the population level improved model accuracy. Furthermore, in the absence of detailed molecular information on the populations studied, researchers could generate relevant data partitions using alternative data sources such as subspecies delimitations, morophological differences or other phenotypic traits.
Phylogeographic implications for P. polionotus
Guisan & Zimmermann  encourage collaboration with evolutionary biologists and population geneticists in cases where widespread species are being modeled. More recently, Rödder et al.  discussed how a variety of techniques including molecular ecology and environmental niche modeling can be complimentary in answering phylogeographic questions. The case is such here, where our method of partitioning data was based largely on the literature, which includes population genetic studies that have been conducted on P. polionotus. Conversely, as molecular work helped to inform our models, our models also shed light on and confirm results of studies on the species' genetic structure, and possibly evolutionary trajectory. For example, the SDMs for P. p. rhoadsi and P. p. nivieventris do not overlap and are geographically discrete (Figure 4 and 2c), which is consistent with the genetic results of Degner et al.  and current taxonomy (Figure 1). The southeastern quadrant model, however, has very poor resolution of this finescale distinction (Figure 3c). Similarly, the models for P. p. polionotus and P. p. colemani capture the known extent of their respective ranges (Figure 5), while the two northern quadrant models do not (Figure 3, a and 3b), implying that these two subspecies occupy different climatic niches.
Climate appears to play an important role in defining inland and inland vs. coastal subspecies (e.g. Figure 4A-D). However, despite there being genetic differences in the beach mouse subspecies located in the Florida panhandle (Figure 5A-F), there was considerable overlap in their predicted distributions, suggesting that climate may not be the primary factor defining the range of these subspecies. Predicted niche overlap usually occurred between adjacent coastal subspecies (Figure 5A-F) and only once between coastal and inland subspecies (see P. p. albifrons and P. p. sumneri, Figure 5F-G). Studies have shown that the coastal beach mouse populations reflect patterns of local adaptation and strong selection favoring cryptic coloration [22–25]. Therefore, in cases where climatic habitat on adjacent coastal beaches might be similar, vicariance (i.e. coastal inlets) and strong selection for coat coloration are more likely than climate to maintain the distinctiveness of coastal beach mouse subspecies.