27 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist Vol. 12, Special Issue 4 Development and Validation of Habitat Models for the Threatened Blackside Dace, Chrosomus cumberlandensis, at Two Spatial Scales Tyler R. Black1,2,*, Brena K. Jones1,2, and Hayden T. Mattingly1 Abstract - Chrosomus cumberlandensis (Blackside Dace) is a small-bodied, freshwater fish endemic to the upper Cumberland River system in southeastern Kentucky and northeastern Tennessee. A detailed study of its habitat requirements using presence-absence data has not been published to date. Identification of important habitat variables at multiple spatial scales would facilitate proactive management and recovery of this federally listed species. Using logistic regression, we developed and validated habitat models of Blackside Dace presence as a function of environmental variables measured at two spatial scales (i.e., stream and reach). For model development, map-produced and fieldcollected variables were gathered for 91 waterways at the stream scale and 72 localities at the 200-m reach scale during summer. Our initial models predicted the likelihood of Blackside Dace presence to be optimized in streams with crude gradient between 1% and 6%, and in reaches with a turbidity ≤10 NTU, dissolved oxygen >8.5 mg/L, summer water temperatures between 14.6 oC and 18.5 oC, conductivity <240 μS, percent riffle between 35% and 50%, and link magnitude between 3 and 6, although the species was occasionally observed in locations with conditions outside of thi s predicted optimal range. We then validated the models by collecting additional data from 27 new streams and 47 new reaches. Model performance was assessed with Cohen’s kappa (κ). The strongest models included conductivity as a predictor variable, with the combination of conductivity and temperature producing the strongest performance (κ = 0.41). Models containing crude gradient, turbidity, dissolved oxygen, and percent riffle generally did not perform well upon validation. Our findings suggest that conductivity, water temperature, and link magnitude are three important reach-scale variables for resource managers to consider when conserving populations of Blackside Dace. Introduction Approximately 90% of temperate freshwater fish species found within the United States are considered nongame fishes (Warren and Burr 1994). These nongame species are popularly perceived as “minnows” whose existence is seldom recognized by the public and whose primary value to humans lies in their use as bait or food for game fishes (Sheldon 1988, Warren and Burr 1994). Among these “minnows” are two speciose families, Cyprinidae (minnows) and Percidae (darters, perches, sauger, and walleye), which comprise nearly half of the US fish fauna (Sheldon 1988, Warren and Burr 1994) and 61% of the Southeast’s fish fauna (Walsh et al. 1995). Within Cyprinidae, 49 species are considered 1Department of Biology, Box 5063, Tennessee Technological University, Cookeville, TN 38505. 2Current Address - North Carolina Wildlife Resources Commission, 1718 NC Highway 56 West, Creedmoor, NC 27522. *Corresponding author - tyler.black@ ncwildlife.org. Ecology and Conservation of the Threatened Blackside Dace, Chrosomus cumberlandensis 2013 Southeastern Naturalist 12(Special Issue 4):27–48 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist 28 Vol. 12, Special Issue 4 vulnerable, 20 are threatened, 47 are endangered, and 11 are extinct (Jelks et al. 2008). Many of these imperiled fishes are localized in the Tennessee-Cumberland and Mobile river basins; among these species, 70% are endemic or restricted to one of the two regions (Master et al. 1998). Chrosomus cumberlandensis (Starnes and Starnes) (Blackside Dace) is a federally protected (status = threatened) cyprinid endemic to the upper Cumberland River system within southeastern Kentucky and northeastern Tennessee (Black et al. 2013 [this issue], Eisenhour and Strange 1998; O’Bara 1988, 1990; Starnes and Etnier 1986; Starnes and Starnes 1978, 1981; USFWS 1987, 1988). Blackside Dace obtain an adult total length of 50–80 mm, have a moderately deep and slightly compressed body, and differ from congeners by exhibiting distinct coloration patterns (Etnier and Starnes 2001, Starnes and Starnes 1978, H.T. Mattingly, unpubl. data). Blackside Dace generally inhabit shallow pools in cool streams with ample refugia such as undercut banks, rootwads, sunken logs, and boulders. These small, 2- to 5-m-wide stream habitats often exhibit intact riparian zones and a percent riffle:pool ratio not exceeding 60:40 (O’Bara 1990; Starnes and Starnes 1981; USFWS 1987, 1988). Using presence-only data, Liang et al. (2012) recently modeled Blackside Dace habitat suitability in Kentucky at the stream-segment scale and found that stream gradient, mining density, and stream order most influenced their predictions of habitat suitability. Despite these generally recognized habitat affinities, a detailed study of Blackside Dace presence-absence and habitat use at multiple spatial scales has not been published to date. Development of coal and timber resources in the Cumberland River basin has led to degradation of streams and extirpation of numerous aquatic populations (Kentucky Division of Water 2000). Acid mine drainage has increased heavymetal concentrations and changed pH in streams; logging, road development, and cattle grazing have modified riparian zones, increased siltation, and altered temperature regimes; agriculture has contributed chemical runoff from crops; and rural homes have discharged untreated sewage and garbage into streams (O’Bara 1990; Starnes 1981; Starnes and Starnes 1978; USFWS 1987, 1988). These threats highlight the need for ongoing management actions to slow Blackside Dace habitat degradation. In particular, describing the relationship between the dace and its habitat is an important component of the species’ recovery plan (objective 1.3.1; USFWS 1988). Furthermore, knowledge of habitat ecology will play a vital role if new populations are to be discovered or reintroduced (objective 2; USFWS 1988). Porter et al. (2000) noted that conserving fish diversity requires reliable data on distribution and habitat, especially for species that are at risk because of low numbers, a fragmented or endemic distribution, or habitat perturbation issues. Therefore, predicting the distribution of rare species based on habitat parameters is often perceived to be a useful tool for conservation and ecological management (Fielding and Bell 1997). Models estimating species presence, population size, distribution, areas with high risks of extinction, and potential reintroduction locations are often constructed from survey data (i.e., biological, physical, and 29 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist Vol. 12, Special Issue 4 chemical components of ecosystems) using correlative univariate or multivariate methods such as logistic regression, discriminant analysis, and artificial neural networks (e.g., Fielding and Bell 1997, Joy and Death 2002, Kanno et al. 2012, Manel et al. 2001). Fielding and Bell (1997) and Manel et al. (2001) advocate that models should be tested with independent information to ensure their merit and accuracy. Comprehensive modeling investigations usually include a validation component, in which models are tested in new situations to assess their performance and applicability to management decisions (e.g., Mattingly and Galat 2004). Although model validation is often stressed as an important component of sound, predictive habitat models, it is frequently neglected in analyses (Manel et al. 2001). Olden et al. (2002) also report that researchers often perceive the modeling process as complete once construction and predictions are made, a problem reinforced by the reliance on statistical modeling software packages. Thus, the objectives of this study were to (1) develop logistic regression habitat models to predict Blackside Dace presence at stream and reach spatial scales, and (2) validate the effectiveness of habitat models to accurately predict presence at other sites within the upper Cumberland River drainage. Methods Model development Stream scale. A stream-scale analysis was conducted to distinguish streams that have a historic presence of Blackside Dace from those that do not. Ninety-one streams within the species’ historic range were selected for study, including 46 streams with known dace presence and 45 streams in which dace absence has been recorded (USFWS, Cookeville, TN, unpubl. data). The historic range of the Blackside Dace was divided into sub-basins based on 11-digit hydrological unit codes (HUC), yielding a total of 45 sub-basins. To choose streams representing dace absence, 45 streams were selected haphazardly from 40 of these 45 sub-basins. Variables measured for each stream were watershed (drainage) area, stream elevation at mouth and headwaters, stream length, stream gradient, stream order at mouth (Strahler 1957), and C-link (Fairchild et al. 1998). Data were collected from either topographic maps in MapTech’s® Terrain Navigator (MTN) or geographic information system (GIS) layers at a 1:24,000 scale. The drainage area for each stream was considered to be the area enclosed by the 14-digit HUC boundaries in ArcMap® GIS. C-link, an indicator of a stream’s position in the watershed, was calculated by counting the number of confluences downstream from the stream’s mouth until the Cumberland River mainstem was reached. Reach scale. Reach-scale models were constructed to distinguish contemporary differences between stream reaches that harbor Blackside Dace and those that do not. Reaches were defined as 200 m in length along the longitudinal stream axis. One to four such reaches were established in 28 streams with a known historic record of Blackside Dace presence, yielding a total of 72 study reaches surveyed by Black et al. (2013 [this issue]). Reaches were sampled via AC backpack electrofishing equipment in June through August 2003 and yielded T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist 30 Vol. 12, Special Issue 4 Blackside Dace presence at 25 of 28 streams and 52 of 72 reaches (Black et al. 2013 [this issue]). Twenty-two habitat variables were selected for reach-scale analysis, 19 of which were derived from field measurements, while the remaining 3 (stream order, link magnitude, and straight-line distance to coal exploration [km]) were assessed in ArcMap® GIS. Transects perpendicular to streamflow were established at each 200-m reach, spaced regularly at 20-m intervals. Not all variables were measured at each of the 72 reaches; however, a minimum sample size of at least 30 reaches for each variable was maintained. Water quality variables including turbidity (nephelometric turbidity units [NTU]), dissolved oxygen (mg/L), temperature (oC), conductivity (μS/cm), and pH were recorded once at the upstream and downstream limit of the reach. Turbidity was measured with an HF Scientific MicroTPI turbidimeter; dissolved oxygen, temperature and conductivity were recorded using a Yellow Springs Instrument (YSI)® Model 85 meter; and pH was determined using an Oakton Instruments pHTestr 3+ meter. Wetted-channel width, maximum and average water depth, and percent streambed gradient were measured at each transect, 11 times per reach. Percent gradient was measured using a handheld clino meter from one transect line upstream to the previous one. Values for each aforementioned variable were averaged to generate a single value for the entire reach. The coefficient of variation for maximum water depth and percent gradient was calculated once for each reach. Stream discharge (cm3/s), as outlined by Gallagher and Stevenson (1999), was assessed using a Marsh-McBirney Flo-Mate 2000 or 201D and a topset wading rod. Silt depth and percent of site composed of riffle habitat were recorded for each reach as a whole. Percent riffle was visually estimated for the entire reach cumulatively during electrofishing. Silt depth covering the rocky substrate was visually estimated at the upstream and downstream limits of each reach following a single calibration with a ruler at the beginning of the field season, and the highest value for the reach was used in analysis. The dominant substrate size along each transect line was visually estimated and classified as follows: 1 = silt, 2 = sand, 3 = gravel, 4 = cobble, 5 = rock, 6 = boulder; and 7 = bedrock (using a modified Wentworth scale from Bain 1999), and it was recorded 11 times per reach. The mode of the substrate values was calculated to generate a single substrate value for each reach. These categories were visually calibrated using a ruler at the beginning of the field season. Streambank vegetative and rock cover were evaluated 11 times per reach in the area where the transect line met the bank for a distance of 0.5 m both up- and downstream. Cover included any material observed between the water’s edge and the upper bankfull channel edge, and its extent was evaluated for both streambanks using the Daubenmire scale (1 = 0-5%, 2 = 5–25%, 3 = 25–50%, 4 = 50–75%, 5 = 75–95%, 6 = ≥ 95%; Daubenmire 1959) to estimate vegetative and rock cover separately. The 11 values for each bank were averaged, and then the two resulting values were averaged to create separate values for rock and vegetative cover for each reach. Woody debris 31 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist Vol. 12, Special Issue 4 was categorized by measuring the diameter (cm) of the largest piece in the stream channel that intersected the transect line: 1 = no debris present, 2 = <10.2 cm, 3 = 10.2–15.2 cm, 4 = 15.2–30.5 cm, 5 = >30.5 cm, and 6 = root mass or accumulated woody debris. A bank undercut was recorded as present (1) if the streambank extended over the edge of the wetted channel within ≤0.5 m from the water surface on either side of a transect. If neither side of the transect had an undercut bank, a value of 0 was recorded. The 11 observations per reach of woody debris and bank undercut were averaged to generate a single value for each vari able. Statistical analysis. For each environmental variable a frequency histogram was constructed comparing Blackside Dace presence and absence. Then, Spearman rank-order correlations were calculated in SAS® (SAS Institute, Inc.; www.sas.com) for all possible combinations of environmental variables, including Blackside Dace presence/absence. The sequential Bonferroni correction procedure was used to correct for experiment-wise error (Holm 1979) for habitat variable correlations; this procedure improves statistical power over the standard Bonferroni correction when testing multiple null hypotheses (Rice 1989). Correction reduces α values and thus reduces the probability of Type I error. However, for analysis of potential relationships between dace and habitat attributes, pre-correction values were used; because this species is threatened, taking an overly conservative statistical approach would be detrimental to recovery and management. Logistic regression models were built in SAS® with both single and multiple predictors relating the presence/absence of Blackside Dace to environmental variables. Models were in the following format (Hosmer and Leme show 2000): π (BSD) = e f / (1 + e f ), where f = β0 + β1x1 + β2x2 + … + βixi , π (BSD) equals the probability of Blackside Dace presence, β0 equals the model intercept (constant), βi equals the parameter estimate, and xi is the environmental variable, where i ≥ 1. Model performance was evaluated internally using Cohen’s kappa (κ): κ = ([a + d] – [([a + c][a + b] + [b + d][c + d]) / n]) , (n – [([a + c][a + b] + [b + d][c + d]) / n]) where a is the number of true positives, b is the number of false positives, c is the number of false negatives, d is the number of true negatives, and n is the total number of all cases (Fielding and Bell 1997, Manel et al. 2001, Mattingly and Galat 2004). Cohen’s kappa is a statistic that calculates the proportion of all presence/ absence cases that are correctly predicted by a model after taking random chance into consideration (Manel et al. 2001). A probability decision threshold was chosen individually for each model, based on the distribution of output values, in order to maximize κ. Any output value above this threshold was considered to be a prediction of dace presence, while values below were classified as a prediction of absence. For example, at a probability decision threshold of 0.70, a model output of 0.43 for a stream reach would be classified as indicating dace T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist 32 Vol. 12, Special Issue 4 absence in that particular reach. SAS-generated concordance values were also recorded; a higher concordance value indicates that a model is better able to correctly predict the outcome. Odds ratios were generated for variables exhibiting significant correlations with dace presence. All models were ranked according to their Akaike information criteria (AIC); a low AIC indicates a more valuable model (Burnham and Anderson 2002). All R2 values reported are ratios of loglikelihood values for each null and fitted model (Nagelkerke 1991). R2 values for logistic regression are often lower than those reported for linear regression models due to the binary nature of the response variable, and they can be difficult to interpret for a single equation (Hosmer and Lemeshow 2000). However, these values can be useful when comparing the applicability of competing logistic regression models (Hosmer and Lemeshow 2000). For all statistical tests, α = 0.05. Certain continuous habitat variables were transformed to categorical variables to provide stronger models and to highlight intervals of disproportionate dace occupancy. Discrete intervals were identified using each variable’s frequency histogram; specifically, if several adjoining categories had “present” frequency bars consistently higher than “absent” bars, this interval was assigned a value of 1. For example, streams with a gradient between one and six were assigned a value of 1 and streams outside this range were assigned a value of 0 to create the categorized crude gradient variable. Similarly, for reaches, categorized turbidity is a dichotomous variable whose value was 1 if turbidity was ≤10 NTU and 0 if it was >10 NTU. Categorized dissolved oxygen was equal to 1 (>8.5 mg/L) or 0 (≤8.5 mg/L). Categorized percent riffle was equal to 1 (between 35 and 50%) or 0 (outside of aforementioned range). Categorized link magnitude was equal to 1 if the link magnitude of the reach was between 3 and 6, and 0 if it had any value outside of this range. The only exception to this categorization procedure occurred with conductivity: first, the selected interval did not occur exactly at the transition point between higher present or absent bars (shifted from 200 μS to 240 μS); second, conductivities above 240 μS were assigned a value of 1 and conductivities below 240 μS were assigned a value of 0 to generate a negative parameter coefficient during the modeling process. Model validation Significant models generated during the model-development phase were subsequently validated with additional, independent data collected during June through August, 2005 and 2006 (Black et al. 2013 [this issue]). Specifically, 47 new 200-m reaches within 27 new streams were sampled via pulsed-DC backpack electrofishing equipment (Smith-Root® Model LR-24; Vancouver, WA) to assess presence of Blackside Dace (Fig. 1; Black et al. 2013 [this issue]). One difference between model development and validation was that development data were from streams historically inhabited by Blackside Dace, while validation data were from watersheds known to harbor dace, but not restricted to inhabited streams. Also, more streams in Tennessee were included in the validation phase than in the development phase (Fig. 1). All validation variables were assessed as 33 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist Vol. 12, Special Issue 4 Figure 1. Upper Cumberland River system in Kentucky and Tennessee depicting 200-m study reaches sampled during model development in summer 2003 (triangles) and model validation in summers 2005 and 2006 (circles). T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist 34 Vol. 12, Special Issue 4 outlined previously; however, models that included dissolved oxygen were omitted from analysis due to erratic oxygen-meter readings in 2005. Deviations from the aforementioned variable assessments include omitting 2 streams from the stream-scale analysis and reach-scale link magnitude because they were not present on topographic maps. In addition, percent riffle was obtained by two methods (visual estimation and physical measurement) to thoroughly evaluate models that included percent riffle as one of the parameters. The operator of the backpack electrofishing unit visually estimated percent riffle during single-pass electrofishing efforts, and physical measurements were taken after the electrofishing pass by recording the total length of riffle within a reach to determine actual percent riffle. Values from the habitat variables assessed at the stream and reach scale were inserted into models to generate the predicted probability of Blackside Dace presence at new localities. Predicted probability of dace presence values were arranged in order of increasing magnitude for each model and graphed to facilitate visual assessment of model performance. Performance was quantitatively assessed using the Cohen’s kappa statistic (κ). A confusion matrix (Fielding and Bell 1997, Manel et al. 2001, Mattingly and Galat 2004) was constructed to calculate κ, and probability decision thresholds were assessed to include a minimum of 20% of the total number of streams or reaches. Probability decision thresholds were set where sensitivity (the percentage of true positives correctly predicted) was maximized. If visual assessment of a model graph indicated that a probability decision threshold should be adjusted, the threshold was then reset where misclassification rate (the proportion of false positives and the false negatives within the data set) was minimized. Sensitivity and misclassification rate were calculated using the following equations: Sensitivity = a / (a + c), Misclassification rate = (b + c) / n, where a, b, c, and n are the same variables as described for Cohen’s kappa (Fielding and Bell 1997). Kappa values range from -1 to 1, with higher values representing stronger model performance and values below zero indicating poor performance. According to Manel et al. (2001), kappa values indicate model performance as follows: 0.0–0.4 = slight to fair, 0.4–0.6 = moderate, 0.6–0.8 = substantial, and 0.8–1.0 = almost perfect (with 1 meaning all sites were correctly classified). Results Model development Stream scale. The number of significant correlations between Blackside Dace and habitat attributes was reduced from 17 to 10 after using the sequential Bonferroni procedure. Before Bonferroni correction, dace presence was positively correlated with one habitat variable at the stream scale, categorized crude gradient (|rs| = 0.32, P = 0.002); dace were 3.9 times more likely to be found in 35 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist Vol. 12, Special Issue 4 streams with a crude gradient between one and six, yielding a significant logistic regression model (G = 9.57, P = 0.002; Table 1). Six other variables (watershed area, stream length, stream order, stream elevation at mouth or headwaters, and C-link) were not significantly correlated with Blackside Dace presence and did not result in any significant models (Jones 2005). Reach scale. The number of significant correlations between Blackside Dace and habitat attributes was reduced from 84 to 24 after using the sequential Bonferroni procedure. Before Bonferroni correction, dace presence was correlated with five habitat variables. There was a negative correlation with water temperature (|rs| = 0.29, P = 0.01), turbidity (|rs| = 0.28, P = 0.02), link magnitude (|rs| = 0.26, P = 0.03), and conductivity (|rs| = 0.24, P = 0.04). Variation in average maximum water depth was positively correlated with dace presence (|rs| = 0.28, P = 0.04), but no significant regression model could be constructed for this variable. Seventeen other habitat variables were not correlated with Blackside Dace presence (Jones 2005). Fourteen significant logistic regression models were created using turbidity, dissolved oxygen, water temperature, conductivity, percent riffle, link magnitude, and selected combinations of these 6 variables (Table 1, Fig. 2). Blackside Dace were especially likely to be found in reaches with a turbidity ≤10 NTU (odds ratio = 3.44), dissolved oxygen >8.5 mg/L (odds ratio = 4.16), summer temperatures between 14.6 oC and 18.5 oC (odds ratio = 0.63), conductivity <240 μS (odds ratio = 0.11), percent riffle between 35 and 50 (odds ratio = 3.25), and a link magnitude between 3 and 6 (odds ratio = 9.43). The strongest model (AIC = 71.05; Table 2) incorporated categorized versions of both conductivity and link magnitude. Of the significant variables, the strongest single predictor of Blackside Dace presence was categorized percent riffle (AIC = 75.29); categorized turbidity was the weakest (AIC = 83.91). The model including both categorized turbidity and categorized dissolved oxygen had the highest Kappa value (κ = 0.44), indicating a higher proportion of agreement between model predictions and actual field data used to create the model. The two-variable models had kappa values ranging from 0.30 to 0.44, while the single variable models had a wider range, from 0.19 to 0.40. Of the single-variable models, link magnitude produced the highest Kappa (κ = 0.40), and categorized dissolved oxygen had the lowest value (κ = 0.19). Model validation Most (58%) of the Blackside Dace logistic regression habitat models had slight performance (i.e., kappa of 0.0–0.4; Table 2). One stream-scale model (categorized gradient; kappa = -0.18), and three reach-scale models that included turbidity and link magnitude individually or jointly had poor performance. Models with slight performance had kappa values that ranged from 0.05 to 0.39. The weakest model with slight performance was turbidity and temperature with a kappa of 0.05, followed by turbidity and categorized percent riffle (κ = 0.09), categorized percent riffle (κ = 0.09), temperature (κ = 0.19), temperature and T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist 36 Vol. 12, Special Issue 4 Table 1. Logistic regression models predicting Blackside Dace presence at two spatial scales. The Wald χ2 statistic tests the significance of each coefficient. The odds ratio indicates by how much the probability of dace pr esence will increase given a corresponding increase of one in t hat term. Model/Term Coefficient SE Wald χ2 P > χ2 Odds Ratio AIC G P R2 Stream-scale model A: Categorized crude gradient 120.57 9.57 0.002 0.13 Constant -0.7730 0.3490 4.9059 0.0268 Categorized crude gradient 1.3548 0.4515 9.0057 0.0027 3.8760 Reach-scale model A: Categorized conductivity and categorized link magnitude 71.51 19.58 <0.0001 0.34 Constant 3.4275 1.0653 10.3525 0.0013 Categorized conductivity -2.7007 0.9384 8.2832 0.004 0.0670 Categorized link magnitude 2.2436 0.8347 7.2241 0.0072 9.4270 Reach-scale model B: Categorized turbidity and categorized perc ent riffle 71.68 9.70 0.0078 0.20 Constant -0.2559 0.4706 0.2958 0.5865 Categorized turbidity 1.4405 0.6407 5.0542 0.0246 4.2230 Categorized percent riffle 1.3605 0.6428 4.4805 0.0343 3.8980 Reach-scale model C: Categorized percent riffle 75.29 4.10 0.0430 0.09 Constant 0.4308 0.3563 1.4621 0.2266 Categorized percent riffle 1.1787 0.6057 3.7861 0.0517 3.2500 Reach-scale model D: Temperature and categorized conductivity 77.43 13.65 0.0011 0.25 Constant 11.1949 4.0611 7.5988 0.0058 Temperature -0.4175 0.2073 4.0558 0.0440 0.6590 Categorized conductivity -1.9965 0.7774 6.5956 0.0102 0.1360 Reach-scale model E: Categorized turbidity and link magnitude 77.82 12.61 0.0018 0.23 Constant 1.2419 0.4954 6.2850 0.0122 Categorized turbidity 1.5569 0.6372 5.9700 0.0146 4.7440 Link magnitude -0.1718 0.0714 5.7819 0.0162 0.8420 Reach-scale model F: Categorized turbidity and temperature 77.95 13.13 0.0014 0.24 Constant 10.7487 4.0671 6.9845 0.0082 Categorized turbidity 1.5175 0.6190 6.0106 0.0142 4.5610 Temperature -0.5531 0.2154 6.5931 0.0102 0.5750 37 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist Vol. 12, Special Issue 4 Table 1, continued. Model/Term Coefficient SE Wald χ2 P > χ2 Odds Ratio AIC G P R2 Reach-scale model G: Temperature and link magnitude 79.56 10.87 0.0044 0.20 Constant 10.1498 4.0630 6.2405 0.0125 Temperature -0.4471 0.2095 4.5549 0.0328 0.6390 Link magnitude -0.1357 0.0704 3.7214 0.0537 0.8730 Reach-scale model H: Categorized conductivity 79.90 9.19 0.0024 0.17 Constant 3.5009 0.9303 14.1613 0.0002 Categorized conductivity -2.1741 0.7573 8.2416 0.0041 0.1140 Reach-scale model I: Categorized dissolved oxygen and link magn itude 80.32 10.11 0.0064 0.19 Constant 1.3501 0.4935 7.4838 0.0062 Categorized dissolved oxygen 1.3481 0.7023 3.6839 0.0549 3.8500 Link magnitude -0.1414 0.0712 3.9433 0.0471 0.8680 Reach-scale model J: Categorized turbidity and categorized diss olved oxygen 80.41 10.68 0.0048 0.20 Constant -0.0442 0.4015 0.0121 0.9123 Categorized turbidity 1.3183 0.5868 5.0469 0.0247 3.7370 Categorized dissolved oxygen 1.5130 0.7087 4.5576 0.0328 4.5400 Reach-scale model K: Link magnitude 82.68 5.75 0.0165 0.11 Constant 1.7199 0.4615 13.8904 0.0002 Link magnitude -0.1404 0.0662 4.4982 0.0339 0.8690 Reach-scale model L: Temperature 82.69 6.40 0.0114 0.12 Constant 9.7152 3.7528 6.7020 0.0096 Temperature -0.4631 0.1958 5.5941 0.0180 0.6290 Reach-scale model M: Categorized dissolved oxygen 83.86 5.22 0.0223 0.10 Constant 0.5680 0.3036 3.5006 0.0613 Categorized dissolved oxygen 1.4243 0.6862 4.3079 0.0379 4.1550 Reach-scale model N: Categorized turbidity 83.91 5.17 0.0230 0.10 Constant 0.4055 0.3450 1.3810 0.2399 Categorized turbidity 1.2368 0.5639 4.8104 0.0283 3.4400 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist 38 Vol. 12, Special Issue 4 Figure 2. Relative frequency histograms depicting turbidities, dissolved oxygen levels, temperatures, conductivities, percent riffles, and link magnitudes associated with Blackside Dace presence or absence during model development. Solid bars represent dace presence and open bars represent dace absence. 39 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist Vol. 12, Special Issue 4 Table 2. Performance statistics for Blackside Dace logistic regression habitat models during development and validation. C = concordance, S = sensitivity, and M = misclassification rate. Model development Model validation Decision Decision ID Models AIC ΔAIC threshold κ C threshold S M κ Stream-scale: A Categorized crude gradient 120.57 n/a 0.50 0.31 42.7 0.32 0.71 0.48 -0.18 Reach-scale: A Categorized conductivity and categorized link magnitude 71.51 0.00 0.70 0.31 62.8 0.67 0.85 0.29 0.39 B Categorized turbidity and categorized percent riffle 71.68 -0.17 0.50 0.40 62.2 0.77 0.19 0.49 0.09 C Categorized percent riffle 75.29 -3.78 0.70 0.23 40.1 0.61 0.19 0.49 0.09 D Temperature and categorized conductivity 77.43 -5.93 0.70 0.35 77.1 0.82 0.65 0.30 0.41 E Categorized turbidity and link magnitude 77.82 -6.31 0.60 0.41 71.7 0.89 0.62 0.49 -0.02 F Categorized turbidity and temperature 77.95 -6.44 0.60 0.33 77.8 0.44 0.58 0.47 0.05 G Temperature and link magnitude 79.56 -8.05 0.70 0.30 74.3 0.33 0.88 0.36 0.22 H Categorized conductivity 79.90 -8.39 0.50 0.35 33.0 0.79 0.73 0.34 0.30 I Categorized dissolved oxygen and link magnitude 80.32 -8.81 0.60 0.34 69.5 n/a n/a n/a n/a J Categorized turbidity and categorized dissolved oxygen 80.41 -8.90 0.60 0.44 62.0 n/a n/a n/a n/a K Link magnitude 82.68 -11.17 0.70 0.40 62.4 0.81 0.39 0.53 -0.03 L Temperature 82.69 -11.18 0.70 0.26 71.1 0.27 0.85 0.38 0.19 M Categorized dissolved oxygen 83.86 -12.35 0.70 0.19 36.0 n/a n/a n/a n/a N Categorized turbidity 83.91 -12.40 0.70 0.24 41.7 0.60 0.65 0.55 -0.16 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist 40 Vol. 12, Special Issue 4 link magnitude (κ = 0.22), categorized conductivity (κ = 0.30), and categorized conductivity and categorized link magnitude (κ = 0.39). The strongest model (temperature and categorized conductivity) had a kappa of 0.41, which represents moderate model performance. The three best-performing models included water conductivity as a predictor variable. Five models and one habitat variable were adjusted for final assessment of model performance. The following models were assessed at the decision threshold where misclassification rate was minimized: categorized percent riffle, temperature and categorized conductivity, turbidity and link magnitude, turbidity and temperature, and link magnitude. Also, models using categorized percent riffle were initially assessed with two separate data sets (visual estimation and physical measurements), but performance for both data sets was in the lower range of the slight performance category. Therefore, only visually estimated data were retained for final analysis, because original model construction was based on visual estimation of percent riffle within a reach. The rank order of regression model performance changed considerably after model validation. In general, models that included temperature or categorized conductivity as variables increased in placement, while models that incorporated turbidity or categorized percent riffle decreased in ranking order (Fig. 3). Turbidity as a single variable produced the weakest model during development and again during validation. Figure 3. Adjusted placement of reach-scale models after validation. Solid bars represent the 5 strongest performing models, while open bars represent weakest performing models. Models on Y-axis are listed from strongest (A) to weakest (N) based on Akaike information criteria generated during development. For example, model D was the fourth strongest model at development, but upon validation it was adjusted up four places to become the strongest model. 41 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist Vol. 12, Special Issue 4 Discussion Numerous studies have shown that there is value in examining the interactions between fish species and biotic and abiotic variables on a variety of spatial scales. Interest in these differing spatial scales arose with the recognition that events occurring at small scales within a watershed (i.e., in a single reach or pool) usually do not affect the dynamics of the entire system; however, largescale effects (i.e., on the entire watershed) could cascade down to the smaller systems nested within (Frissell et al. 1986). Porter et al. (2000) suggested that little was gained in terms of prediction from field-based data on smaller (reach, run-riffle-pool complexes) scales. This, however, was not the case in our study. Highly fragmented Blackside Dace populations live in an area of heavily and patchily disturbed habitat. If they are extirpated from a stream or reach, the nearest population available for recolonization may be prohibitively far away, even given this species’ ability to move relatively long distances (>1 km) in a year’s time (Detar and Mattingly 2013 [this issue]). The reason for extirpation is often unclear, particularly if there are no data associated with the time of population disappearance to suggest a cause. Furthermore, these small, shallow streams have far more ephemeral habitat than higher-order streams and rivers; they are more prone to drought, temperature, and oxygen stress. All of these factors lead to a habitat that is temporally and spatially heterogeneous. Despite this heterogeneity, we found Blackside Dace presence at the 200-m reach scale to be fairly predictable based on as few as three environmental variables. Our validation of Blackside Dace habitat models provides a good example of why validation of habitat models for rare species is extremely important for conservation efforts. The single stream-scale model, crude gradient, had poor performance after testing with independent data. Although our model construction phase (O’Bara 1988) provided support that Blackside Dace inhabit streams with a gradient somewhere between 1% and 6%, this pattern was not especially apparent during validation. Liang et al. (2012) found that low-gradient segments of second- and third-order streams were particularly suitable for Blackside Dace presence in Kentucky. The segment spatial scale is situated between the stream and reach scales (Frissell et al. 1986), and could represent a better spatial resolution for modeling the influence of gradient on dace presence-abs ence. We also observed poor performance for the 3 reach-scale models that included turbidity and link magnitude individually or jointly. Caution is recommended when assessing models that include turbidity because dace in reaches with low abundance could be difficult to detect in turbid waters. After field-testing, it was found that the model including turbidity could only predict occurrence at a rate that was equal to chance. Additionally, rain events prior to sampling would undoubtedly inflate turbidity values above normal circumstances. Therefore, future efforts to evaluate turbidity should follow strict guidelines for timing of measurements (e.g., a set number of days after a rain event or after a rain event of a certain magnitude). Link magnitude as a single variable may have performed poorly because of variation between model-construction and validation reaches, T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist 42 Vol. 12, Special Issue 4 or because it is a single factor in a complex ecosystem. Poor performance of these four models should not imply that the variables are unimportant in this species’ survival. Stream gradient has been associated with keeping substrate free of sediment (Waters 1995, Wood and Armitage 1997), turbidity is linked to land-use patterns and increased sedimentation (Waters 1995), and link magnitude may be related to drought situations when headwater streams lose conne ctivity. Although Starnes and Starnes (1981) and our model development analysis suggested that dace occurred more often in streams with a percent riffle less than 60%, validation of models using percent riffle resulted in low kappa values. Weak performance of models including percent riffle may have resulted from inaccurate visual estimates of percent riffle during model development or differences among survey crews. Therefore, future efforts to model percent riffle should focus on physical measurements of riffles after electrofishing has been completed. Models including temperature tended to have midrange kappa values within the slight performance category, with the exception of the model containing temperature and conductivity (see below). Thermal preferences of freshwater fish are well documented because changes in temperature often limit the distribution, abundance, and behavior of fishes (Huff et al. 2005, Richardson et al. 1994). Elevated stream temperatures are often associated with land-use changes within the watershed, such as riparian-zone vegetation removal due to logging, mining, or agricultural practices, which increase the amount of solar radiation entering a waterway. Increased water temperature can cause decreased dissolved oxygen, which can result in elevated stress levels and lower competitive ability in fish (Diana 2004, Helfman et al. 1997). Helfman et al. (1997) suggested that environmental changes cause production of heat-shock proteins (hsps), which are thought to help fish adapt to stressful situations. Heat-shock proteins stabilize biochemical functions by reconfiguring other proteins that have been denatured by high temperatures. Therefore, production of hsps may explain the continued existence of Blackside Dace in streams that exceed normal thermal boundaries. Invasion of Blackside Dace watersheds by Adelges tsugae Annand (Hemlock Woolly Adelgid) has been discussed by Blomquist et al. (2010) as a threat to Blackside Dace populations. Tsuga canadensis (L.) Carriere (Eastern Hemlock) currently serves a vital role by buffering water temperature changes, reducing fluctuations in stream flow, and decreasing sediment input from surrounding landuse disturbances, simply due to its abundance and canopy contribution in riparian areas of many streams occupied by Blackside Dace (Blomquist et al. 2010). Substantial mortality of Eastern Hemlock due to Hemlock Woolly Adelgid infestations could further complicate recovery efforts, given the likelihood that summer water temperatures would be elevated under more open canopies due to solar heating. A temperature increase of even a few degrees could be detrimental to Blackside Dace persistence (Fig. 2: lower left panel), especially if other environmental parameters were also altered in a less-than-optimal direction for this species. 43 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist Vol. 12, Special Issue 4 The strongest performing models incorporated categorized conductivity as an indicator variable for Blackside Dace presence/absence. After validation with independent data, it was apparent that Blackside Dace have an affinity for waters with conductivities less than 240 μS (Fig. 2; upper right panel). A number of different land-use activities can result in elevated water conductivity, but mining is believed to be the primary activity in Blackside Dace streams that has produced such changes in past and present times (McAbee et al. 2013 [this issue]). Even after mining activities have ceased, elevated conductivity in streams can persist for long periods (reviewed by Bernhardt and Palmer 2011). In a West Virginia study, for example, Bernhardt et al. (2012) reported mean conductivities of 64 μS in reference streams, 118 μS in unmined streams, and 626 μS in mined streams. These authors found biological impairment was likely to occur when surface coal mines exceeded 5.4% of their contributing watershed area, or when conductivity exceeded 308 μS, a value very close to the 300-μS benchmark value recently devel oped for Central Appalachian ecoregion streams by USEPA (2011). Physicochemical properties may vary with location in a watershed, which may also determine the distributional patterns of fishes (Koel and Peterka 2003, Polivka 1999). Polivka (1999) reported that proportionally more individuals of Notropis girardi Hubbs and Ortenburger (Arkansas River Shiner) were observed in a narrow range of conductivities, although seasonal variations were observed. In Spain, conductivity was the most significant variable that affected the condition (i.e, weight-length relationship) of Barbus sclateri Günther (Sclater’s Barbel) (Oliva-Paterna et al. 2003). The abundance of Blackside Dace has also been known to decline through time as conductivity increased in a stream affected by mining disturbance (M. Floyd, USFWS, Frankfort, KY, pers. comm.). Mortality and declines in abundance attributed to high water conductivity may be due to increased energy usage associated with osmotic ionic regulation and poor reproductive success (Koel and Peterka 2003, Oliva-Paterna et al. 2003). However, conductivity may only be a longer-term indicator of population declines, and the true problem may relate to shorter-term episodic events, such as the release of metal ions due to decreased pH during a rain event. Finally, temperature and link magnitude played supportive roles to conductivity in the performance of models A and D, and the importance of their contribution should not be overlooked because these variables further refine the predictive ability of t he models. Overall, most (58%) Blackside Dace habitat models had a kappa value between 0.0–0.4 (slight performance). Manel et al. (2001) recorded high rates of slight to poor performance for Himalayan river birds, with 76% of models indicating poor performance (κ < 0.4) based on criteria proposed by Fielding and Bell (1997). Additionally, Manel et al. (2001) reported that few models were classified as good (21%) and even fewer cases were identified as having excellent performance (3%), so it is not surprising that excellent performance was not obtained during validation with independent data in this study. Nevertheless, our overall predictive success (i.e., percentage of correctly identified cases) of the top five performing models exceeded 61% and the overall mean for the top three models T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist 44 Vol. 12, Special Issue 4 was 69.1% (± 2.8% SD). Given the numerous impacts that could affect Blackside Dace distribution, being able to correctly predict presence or absence 70% (i.e., 33 of 47 sites) of the time with only 2 environmental variables is a noteworthy consideration that merits the attention of resource managers. Holding the decision threshold constant for model D (categorized conductivity and temperature), 5 more correct predictions (i.e., 38 of 47 sites) would have been needed to elevate model performance to the substantial category (κ > 0.6) and 10 more correct predictions (i.e., 43 of 47 sites) would have been required to elevate kappa to almost perfect (κ > 0.8) model performance. Although reach-scale models were developed from streams that historically harbored Blackside Dace, validation was completed using streams with and without a historical occurrence of dace; however, all surveyed streams were within watershed systems known to harbor dace. Therefore, we reevaluated all models using only historical streams to ensure that original performance assessments were accurate. All models remained in their original classification category except models B (turbidity and categorized percent riffle) and C (categorized percent riffle), which decreased from slight to poor performance, and model E (turbidity and link magnitude), which increased from poor to slight performance. However, the primary focus of validation was centered on the one stream-scale model (categorized crude gradient) and the 3 strongest performing models (D, A, H). The 4 primary focus models remained in the same performance classification category and the majority (>60%) of the eliminated reaches for the 3 reach-scale models were correctly classified (i.e., as true positives and true negatives). The stream-scale model, crude gradient, had a reduced kappa value (-0.18 to -0.19), and only 40% of the eliminated streams were correctly classified. Model A (categorized conductivity and categorized link magnitude) also had a reduced kappa value, which decreased from 0.39 to 0.35. Reach-scale models H and D both experienced increased kappa values, with model H (categorized conductivity) increasing from 0.30 to 0.33, and model D (categorized conductivity and temperature) increasing from 0.41 to 0.49. Overall, only slight variation was observed when model performance was assessed with the two data sets; therefore, all sites were retained during validation as reported in the results. In addition, when models are evaluated, researchers often overlook issues associated with species prevalence (Manel et al. 2001). Therefore, the use of evaluation statistics like Cohen’s kappa and normalized mutual information statistic are better suited for situations where prevalence could influence results. Furthermore, it is also important to ensure that the performance of all models exceeds the expectations based on chance (Olden et al. 2002). Performance of all models in this study exceeded the expectations based on chance, except for model N, which used turbidity as a single variable. Management implications Habitat model validation is an essential process to ensure that models are reliable and that the appropriate model is selected for management policies. 45 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist Vol. 12, Special Issue 4 Untested or poorly performing models could lead to wasted resources and efforts focused on conservation goals that would inevitably fail. After field-testing habitat models, it is evident that Blackside Dace habitat management could focus on protecting or restoring stream reaches that meet at least 2 of 3 optimal conditions (summer conductivity less than 240 μS, temperature less than 18.5 oC, and link magnitude between 3 and 6). Of course, many other habitat conditions are necessary for Blackside Dace populations to thrive, and our results merely highlight 3 variables as the most important habitat features elucidated at a 200-m spatial scale. Identifying important habitat variables at larger and smaller spatial scales (e.g., landscape, pool-riffle, microhabitat) also would be a valuable part of future efforts to conserve Blackside Dace populations. Abnormally high conductivity within the upper Cumberland River watershed is most likely linked to numerous anthropogenic perturbations. This region is located amidst vast coal, natural gas, and timber reserves where land-use disturbances are widespread and commonplace. Therefore, continued monitoring of conductivity should be undertaken to better understand its relationship to land use and Blackside Dace persistence. Our results clearly showed conductivity to be a predictable indicator of Blackside Dace presence and persistence in stream reaches. However, the mechanistic or causal link between conductivity and the likelihood of Blackside Dace persistence has yet to be determined. Laboratory experiments on the influence of conductivity on various life stages of dace could identify tolerance thresholds for each life stage and provide a better understanding of dace vulnerability to elevated conductivity and its associated ionic components. The supportive roles played by temperature and link magnitude also warrant monitoring and use of best management practices (BMPs) within watersheds inhabited by Blackside Dace. Maintenance of intact riparian corridors would help ensure that summer temperatures remain within the optimal range (14.6–18.5 oC), and that flow regimes associated with a link magnitude of 3–6 remain conducive to dace occupancy. Furthermore, it is widely known that sufficiently robust riparian buffers provide the added benefit of reducing input of sediment and other pollutants into a waterway. Acknowledgments This research project was supported by the US Fish and Wildlife Service, US Geological Survey, and the Center for the Management, Utilization, and Protection of Water Resources and the Department of Biology at Tennessee Technological University (TTU). Completion of the manuscript was facilitated by a TTU Faculty Non-Instructional Assignment during 2011–2012. We especially thank Jason Detar, Jason Hunt, and Anthony Smith for their extensive field assistance, and the numerous private landowners, Kentucky State Nature Preserves Commission, and the US Department of Agriculture Forest Service for allowing us to access their properties to conduct surveys. The manuscript was improved by comments from C.A. Brown, J.B. Layzer, D.D. Smith, and three anonymous reviewers. T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist 46 Vol. 12, Special Issue 4 Literature Cited Bain, M.B. 1999. Substrate. Pp. 95–104, In M.B. Bain and N.J. Stevenson (Eds.). Aquatic Habitat Assessment: Common Methods. American Fisheries Society, Bethesda, MD. 216 pp. Black, T.B., J.E. Detar, and H.T. Mattingly. 2013. Population densities of the threatened Blackside Dace, Chrosomus cumberlandensis, in Kentucky and Tennessee. Southeastern Naturalist 12(Special Issue 4):6-26. Bernhardt, E.S., B.D. Lutz, R.S. King, J.P. Fay, C.E. Carter, A.M. Helton, D. Campagna, and J. Amos. 2012. How many mountains can we mine? Assessing the regional degradation of Central Appalachian rivers by surface coal mining. Environmental Science and Technology 46:8115–8122. Bernhardt, E.S., and M.A. Palmer. 2011. The environmental costs of mountaintop mining valley-fill operations for aquatic ecosystems of the Central Appalachians. Annals of the New York Academy of Sciences 1223:39–57. Blomquist, S.M., T.D. Johnson, D.R. Smith, G.P. Call, B.N. Miller, W.M. Thurman, J.E. McFadden, M.J. Parkin, and G.S. Boomer. 2010. Structured decision-making and rapid prototyping to plan a management response to an invasive species. Journal of Fish and Wildlife Management 1(1):19–32. Burnham, K.P., and D.R. Anderson. 2002. Model Selection and Inference: A practical Information- Theoretic Approach, 2nd Edition. Springer-Verlag, New York, NY. 488 pp. Daubenmire, R. 1959. A canopy-coverage method of vegetational analysis. Northwestern Science 33:43–64. Detar, J.E., and H.T. Mattingly. 2013. Movement patterns of the threatened Blackside Dace, Chrosomus cumberlandensis, in two southeastern Kentucky watersheds. Southeastern Naturalist 12(Special Issue 4):64-81. Diana, J.S. 2004. Biology and Ecology of Fishes, 2nd Edition. Cooper, Traverse City, MI. 498 pp. Eisenhour, D.J., and R.M. Strange. 1998. Threatened fishes of the world: Phoxinus cumberlandensis Starnes & Starnes, 1978 (Cyprinidae). Environmental Biology of Fishes 51:140. Etnier, D.A., and W.C. Starnes. 2001. The Fishes of Tennessee, 2nd Printing. The University of Tennessee Press, Knoxville, TN. 689 pp. Fairchild, G.W., R.J. Horwitz, D.A. Nieman, M.R. Boyer, and D.F. Knorr. 1998. Spatial variation and historical change in fish communities of the Schuylkill River drainage, southeast Pennsylvania. American Midland Naturalist 139:282–295. Fielding, A.H., and J.F. Bell. 1997. A review of methods for the assessment of prediction errors in conservation presence/absence models. Environmental Conservation 24:38–49. Frissel, C.A., W.J. Liss, C.E. Warren, and M.D. Hurley. 1986. A hierarchical framework for stream habitat classification: Viewing streams in a watershed context. Environmental Management 10:199–214. Gallagher, A.S., and N.J. Stevenson. 1999. Substrate. Pp. 149–157, In M.B. Bain and N.J. Stevenson (Eds.). Aquatic Habitat Assessment: Common Methods. American Fisheries Society, Bethesda, MD. 216 pp. Helfman, G.S., B.B. Collette, and D.E. Facey. 1997. The Diversity of Fishes. Blackwell Science, Malden, MA. 528 pp. Holm, S. 1979. A simple, sequentially rejective multiple test procedure. Scandinavian Journal of Statistics 6:65–70. 47 T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist Vol. 12, Special Issue 4 Hosmer, D.W., and S. Lemeshow. 2000. Applied Logistic Regression, 2nd Edition. Wiley- Interscience, New York, NY. 392 pp. Huff, D.D., S.L. Hubler, and A.N. Borisenko. 2005. Using field data to estimate the realized thermal niche of aquatic vertebrates. North American Journal of Fisheries Management 25:346–360. Jelks, H.L., S.J. Walsh, N.M. Burkhead, and 13 co-authors. 2008. Conservation status of imperiled North American freshwater and diadromous fishes. Fisheries 33(8):372–407. Jones, B.K. 2005. Predictive habitat models for conservation of the threatened Blackside Dace (Phoxinus cumberlandensis). M.Sc. Thesis. Tennessee Technological University, Cookeville, TN. 75 pp. Joy, M.K., and R.G. Death. 2002. Predictive modeling of freshwater fish as a biomonitoring tool in New Zealand. Freshwater Biology 47:2261–2275. Kanno, Y., C.U. Schmidt, S.B. Cook, and H.T. Mattingly. 2012. Variation in microhabitat use of the threatened Spotfin Chub (Erimonax monachus) among stream sites and seasons. Ecology of Freshwater Fish 21:363–374. Kentucky Division of Water. 2000. Cumberland River basin and Four Rivers region status report. Kentucky Division of Water, Frankfort, KY. Koel, T.M., and J.J. Peterka. 2003. Stream fish communities and environmental correlates in the Red River of the North, Minnesota and North Dakota. Environmental Biology of Fishes 67:137–155. Liang, L., S. Fei, J.B. Ripy, B.L. Blandford, and T. Grossardt. 2012. Stream habitat modelling for conserving a threatened headwater fish in the upper Cumberland River, Kentucky. River Research and Applications. Available online at doi: 10.1002/ rra.2605. Manel, S., H.C. Williams, and S.J. Ormerod. 2001. Evaluating presence-absence models in ecology: The need to account for prevalence. Journal of Applied Ecology 38:921–931. Master, L.L., S.R. Flack, and B.A. Stein (Eds.). 1998. Rivers of life: Critical watersheds for protecting freshwater biodiversity. The Nature Conservancy, Arlington, VA. 71 pp. Mattingly, H.T., and D.L. Galat. 2004. Predictive performance of a summer microhabitat model for the threatened Niangua Darter, Etheostoma nianguae. Journal of Freshwater Ecology 19:109–114. McAbee, K.T., N.P. Nibbelink, T.D. Johnson, and H.T. Mattingly. 2013. Informing recovery management of the threatened Blackside Dace, Chrosomus cumberlandensis, using a Bayesian-belief network model. Southeastern Naturalist 12(Special Issue 4):143–161. Nagelkerke, N.J.D. 1991. A note on a general definition of the coefficient of determination. Biometrika 78:691–692. O’Bara, C.J. 1988. Ecological and behavioral characteristics of the Blackside Dace Phoxinus cumberlandensis. US National Forest Service Report, Winchester, KY. 36 pp. O’Bara, C.J. 1990. Distribution and ecology of the Blackside Dace, Phoxinus cumberlandensis (Osteichthyes: Cyprinidae). Brimleyana 16:9–15. Olden, J.D., D.A. Jackson, and P.R. Peres-Neto. 2002. Predictive models of fish species distributions: A note on proper validation and chance predictions. Transactions of the American Fisheries Society 131:329–336. Oliva-Paterna, F.J., A. Andreu, and M. Torralva. 2003. Water quality affects the condition of Barbus sclateri Günther, 1868 (Pisces, Cyprinidae) in semi-arid reservoirs from the Iberian Peninsula. Anales de Biologia 25:3–11. T.R. Black, B.K. Jones, and H.T. Mattingly 2013 Southeastern Naturalist 48 Vol. 12, Special Issue 4 Polivka, K.M. 1999. The microhabitat distribution of the Arkansas River Shiner, Notropis girardi: A habitat-mosaic approach. Environmental Biology of Fishes 55:26 5–278. Porter, M.S., J. Rosenfeld, and E.A. Parkinson. 2000. Predictive models of fish species distribution in the Blackwater Drainage, British Columbia. North American Journal of Fisheries Management 20:349–359. Rice, W.R. 1989. Analyzing tables of statistical tests. Evolution 43:223–225. Richardson, J., J.A.T. Boubee, and D.W. West. 1994. Thermal tolerance and preference of some native New Zealand freshwater fish. New Zealand Journal of Marine and Freshwater Research 28:399–407. Sheldon, A.L. 1988. Conservation of stream fishes: Patterns of diversity, rarity, and risk. Conservation Biology 2:149–156. Starnes, L.B., and W.C. Starnes. 1981. Biology of the Blackside Dace, Phoxinus cumberlandensis. The American Midland Naturalist 106:360–372. Starnes, W.C. 1981. Listing package for the Blackside Dace, Phoxinus cumberlandensis. US Fish and Wildlife Service, Asheville, NC. 77 pp. Starnes, W.C., and D.A. Etnier. 1986. Drainage evolution and fish biogeography of the Tennessee and Cumberland rivers drainage realm. Pp. 325–361, In C.H. Hocutt and E.O. Wiley (Eds.). The Zoogeography of North American Freshwater Fishes. John Wiley, New York, NY. 866 pp. Starnes, W.C., and L.B. Starnes. 1978. A new cyprinid of the genus Phoxinus endemic to the upper Cumberland River drainage. Copeia 1978:508–516. Strahler, A.N. 1957. Quantitative analysis of watershed geomorphology. Transactions of the American Geophysical Union 8(6):913–920. US Environmental Protection Agency. 2011. A field-based aquatic-life benchmark for conductivity in Central Appalachian streams. Office of Research and Development, National Center for Environmental Assessment, Washington, DC. EPA/600/R- 10/023F. US Fish and Wildlife Service (USFWS). 1987. Endangered and threatened wildlife and plants; determination of threatened species status for the Blackside Dace. Federal Register 52(113):22,580–22,585. USFWS. 1988. Recovery plan for Blackside Dace (Phoxinus cumberlandensis). Atlanta, GA. 23 pp. Walsh, M.J., N.M. Burkhead, and J.D. Williams. 1995. Southeastern freshwater fishes. Pp. 144–147, In E.T. Laroe, G.S. Farris, C.E. Puckett, P.D. Doran, and M.J. Mac (Eds.). Our Living Resources: A Report to the Nation on the Distribution, Abundance, and Health of US Plants, Animals, and Ecosystems. US Department of the Interior, Washington, DC. 530 pp. Warren, M.L., and B.M. Burr. 1994. Status of freshwater fishes of the United States: Overview of an imperiled fauna. Fisheries 19(1):6–8. Waters, T.F. 1995. Sediment in Streams: Sources, Biological Effects And Control. American Fisheries Society, Bethesda, MD. 251 pp. Wood, P.J., and P.D. Armitage. 1997. Biological effects of fine sediment in the lotic environment. Environmental Management 21:203–217.