Further Development of a Physical Habitat Index. for Clallam County Streams

Transcription

1 Further Development of a Physical Habitat Index for Clallam County Streams Mike McClean, Ed Chadd, and Steve Obrebski Clallam County Department of Community Development, July 2009 Introduction Modern concepts of stream ecology emphasize the importance of water quality, physical habitat, and land use patterns in affecting the biologic health of streams. Streamkeepers of Clallam County has acquired extensive data on each of these factors for numerous stream sites. Appropriate analysis and interpretation of these data should provide practical insights on the complex ecology affecting the health of County streams. The present report focuses on one aspect of the Streamkeeper dataset: measures of physical habitat and how they relate to stream biology as indicated by aquatic insect life. The work extends the previous analyses of Chadd et al. (2005) aimed at developing a physical habitat index (PHI). In this earlier study physical habitat measures were identified, relevant literature was reviewed, and computational procedures were applied to achieve approximate equivalence across measures. Quantitative studies of stream ecology often utilize global multimetric indices to summarize the relevant variation among large groups of variables. Streamkeepers has taken this approach in employing a water quality index (WQI) (Clallam County, 2004) and Benthic Index of Biological Integrity (B-IBI), a widely used measure of aquatic insect counts (Karr et al., 1986; 1

2 Karr & Chu, 1999). These types of indices are computed from a set of submetrics which reflect different dimensions of the general factor of interest (e.g., dissolved oxygen concentration and ph level are submetrics of a WQI). The two key decisions to be made when developing a multimetric index concern what submetrics to employ and how will measures be weighted and/or arranged within a computational framework to yield the final index. There is little uniformity in how different investigators have addressed these two questions with respect to physical habitat measures (e.g., McBride, 2001; Hall, et al., 1999; Wang et al., 1998). The present paper extends the work of Chadd et al. (2005) by utilizing multivariate statistical procedures to evaluate different approaches to computation of the PHI: one involving standardized multiple regression, another principal component analysis, and a third that combines the two techniques. A key element of the rationale for each method was that PHI computation should be optimized to provide the highest correlation possible with the biologic health of streams as indicated by B-IBI scores. This deviates somewhat from that of Chadd et al. (2005) who focused their selection and preliminary grouping of physical habitat variables on salmon health. Structure and Contents of the Statistical Spreadsheet Physical habitat measures were obtained by Streamkeeper volunteers on 45 monitoring sites between 1999 and The locations of these sites are shown in Figure 1. The field procedures used to acquire the raw data for each measure are described in the Streamkeeper Handbook. The parameters studied by Chadd et al. (2005) are large woody debris, pools, summer canopy cover, winter canopy cover, conifer stem density, percent fine sediment, degradation/aggradation, embeddedness, and bed substrate stability. They scaled various physical habitat measures in light of previous literature pertaining to threshold and saturation levels, and all measures were set to range from 0 to 1. These values are used in the analyses 2

3 described here. Bed substrate stability is not considered in the present analyses due to uncertainties about its computation using the present dataset. Also, preliminary indications are that other geologic measures were not well correlated with the B-IBI, suggesting that bed substrate stability would not be especially useful as a PHI submetric. Additional measures employed in the present analyses, which were not addressed by Chadd et al., are bank stability, aquatic noxious weeds, and riparian noxious weeds. These are all four-level equal-interval ordinal measures, and each was converted to a 0 to 1 scale. Measures of percent conifer density and percent deciduous density were available within the dataset, but the decision was made not to use these variables, because they were scaled at only three levels with unequal increments between levels. Correlation Analyses As mentioned above, it is assumed here that the most useful PHI will be well correlated with aquatic insect health. As an initial step in our analyses, simple correlations were performed relating B-IBI score to the 11 physical habitat measures. The resulting correlations are presented in Table 1. There it can be seen that the B-IBI was significantly correlated with large woody debris, conifer stem count, winter canopy, summer canopy, aquatic noxious weeds, and riparian noxious weeds. Notably, all these measures pertain to stream site vegetation, whereas the uncorrelated variables pertain to stream site physical geology. Only those variables showing significant correlations with the B-IBI were considered in the following analyses involved in computation of the PHI. Standardized Regression Analysis The goal of this analysis was to obtain a statistically valid model that assigned appropriate weighting coefficients to a subset of physical habitat measures. A best subsets multiple regression was first performed relating the B-IBI to those measures showing significant 3

4 correlations as indicated in Table 1. Given the size of the dataset, it was necessary to limit the number of predictor variables to four in selecting a final model. The four predictor variables in the best-fit equation that emerged from this analysis were large woody debris, conifer stem count, riparian noxious weed count, and aquatic noxious weed count. When these four variables are then employed as predictors in a multiple regression, the resulting equation is as follows: BIBI = LWD ConifSt AcqNoxW RipNoxW (1) This analysis showed an adjusted R-square of 44.4%, and the associated analysis of variance showed a p-level less than The variance inflation factor for each of the four variables was no greater than 1.1, indicating an absence of co linearity between variables (Allison, 1999). Because the four independent variables were on equivalent scales from 0 to 1, the resulting analysis produced standardized regression or Beta coefficients. Beta coefficients effectively quantify the relative contribution or weight of the different variables to prediction of the B-IBI. A PHI was effectively computed as the fitted or predicted values from this equation. In order to evaluate the general merit of this new variable, hereafter PHI-reg, further regression analysis was performed relating it to the B-IBI. The results of this analysis are shown in Figure 2. As indicated, the association between PHI-reg and the B-IBI was best fit with a quadratic equation resulting in an adjusted R-square of 52.5%. The quadratic equation improved the R- squared by four percent over a linear best fit. Principal Component Analysis A second approach taken to computation of the PHI involved the use of principal component analysis (PCA). PCA is a statistical technique for reducing the number of variables in a dataset to a smaller set of factors and for better understanding underlying relationships among variables. It is generally employed as an intermediate step in statistical analysis of multivariate datasets (Johnson & Wichern, 2002). Here we utilized PCA to clarify the 4

5 relationships among the six physical habitat variables showing positive correlations with the B- IBI. For our purposes, the essential results of the PCA are captured in the loading plot in Figure 3. This plot shows the association of the first two principal component values for the six physical habitat variables. The pattern of vectors of the six variables suggests two general dimensions or factors within the dataset. One of these is represented by the vectors involving aquatic noxious weeds and large woody debris. Their relatively large projections in opposite directions indicate that they are negatively correlated. The presence of a second dimension in the dataset is suggested by the vectors involving riparian noxious weeds and winter canopy cover which also appear to be negatively correlated. The orientations of the combined vectors of the two factors (LWD-AcqNoxW and WintCan-RipNoxW) are approximately at a right angle or orthogonal to one another, suggesting that they are independent. Working from this general interpretation of the loading plot in Figure 3, we inferred that a reasonable computing formula for a PHI would involve calculating the difference between the two variables in each factor and then summing the results as follows: PHI-pca = (LWD - AcqNoxW) + (WintCan - RipNoxW) To evaluate the merit of this approach, a regression analysis was performed relating PHIpca to the B-IBI, as was done for PHI-reg. The results of this analysis are shown in Figure 4. There it may be seen that a quadratic best fit produced an adjusted R-squared of 53.3 percent, essentially the same result as that obtained for PHI-reg. In this case, the R-squared associated with a quadratic best fit was eight percent higher than that obtained with a linear best fit. 5

6 Standardized Regression Using PCA Results In a third approach to computation of the PHI, we utilized the above results of the PCA to select four predictor variables in a standardized regression analysis with the B-IBI as the dependent variable. The resulting equation is as follows: BIBI = LWD WintCan AcqNoxW 11.9 RipNoxW (2) The associated analysis of variance on this regression showed a p-level less than The variance inflation factor for each of the four variables was no greater than 1.4, indicating very limited co linearity between variables (Allison, 1999). A third form of PHI was taken as the predicted or fitted values from Equation 2. The merit of this new variable, PHI-pcar, was evaluated by regressing it with the B-IBI. The resulting best-fit was obtained with a quadratic equation showing an adjusted R-squared of 56.4%. This represents an approximate increase of 3% over the other two approaches. Further Analysis in Light of the B-IBI Distribution Examination of Figures 2 and 4 suggest that the nonlinear association of PHI-reg and PHI-pca were driven largely by B-IBI values less than 30. Examination of a frequency histogram of the B-IBI as shown in Figure 5 indicates a negatively skewed distribution with the tail being associated with B-IBI values less than 30. This result prompted an additional analysis to assess the association of physical habitat measures with B-IBI for those values above 30. For those rows in the spreadsheet where the B-IBI exceeded 30, a best-fit regression analysis was performed. The best four-variable regression model included fine sediment, large woody debris, conifer stem count, and riparian noxious weeds. This analysis showed an adjusted R-square of 50.7% and the associated analysis of variance showed a p-level less than Unlike the analyses involving the full dataset, the scatter plot relating the actual B-IBI scores to their predicted values was best-fit with a linear equation as shown in Figure 6. 6

7 Discussion The present analysis utilized standardized regression and principal component analyses for computation of three different forms of physical habitat index or PHI. Each approach was evaluated with respect to how well they predicted stream biologic health as quantified by the B- IBI. In each case analysis involved a subset of variables that were positively correlated with the B-IBI. These variables were related to stream site vegetation whereas geologic variables (e.g., embeddedness) were uncorrelated with the B-IBI. Roughly equivalent results were obtained with all three methods, each showing a quadratic association with the B-IBI that accounted for a little over 50% of the variance in the B-IBI. The negative skew seen in the distribution of B-IBI scores and the related nonlinearity in the regression results suggest a qualitative difference in the ecology of stream sites with B-IBI scores above and below 30. Regression analysis that included only values greater than 30 showed a similar set of predictors accounting for approximately 50% of the variance in the B- IBI, although fine sediment became one of the four best predictors. In thinking about possible modifications and future use of the physical habitat suite in Streamkeeper monitoring procedures, the two key issues concern which submetrics to employ and how those submetrics be weighted and arranged for computation of the PHI. The present results suggest that measures might effectively be limited to large woody debris, aquatic noxious weeds, riparian noxious weeds, and either winter canopy cover or conifer stem count. As to the computational approach, several considerations suggest that PHI-pca would most suitable. The selection of variables for PHI-pca was based on an analysis that revealed two distinct and intuitively plausible dimensions within the set of habitat variables. The negative associations of the two poles of these dimensions generally agree with our intuitions about associated stream ecology. That is, one would expect increased levels of aquatic noxious weeds in areas with 7

8 reduced large woody debris, and riparian noxious weeds would be more evident in areas with reduced winter canopy cover. The use of PHI-pca is also appealing because of the simplicity of its computation. Unlike the other two measures, it does not depend on regression coefficients and the error inherent in their modeling equations. An important limitation of the present results concerns the focus on aquatic insect health as the criteria for development and evaluation of the PHI. Chadd et al. (2005) were thinking chiefly of salmon health in selecting the variable set for development of a PHI for Clallam County streams. While there is reason to expect that the B-IBI is associated with salmon health, geologic measures have strong external validity for salmon health, traditionally they have been a major component of the PHI, and there is direct evidence for their association with salmon-based biological indices (e.g., Wang et al., 1998). One possible strategy for future assessment of physical habitat in Clallam County streams would be to utilize two physical habitat indices, one that focused on vegetation and a second on geology. 8

9 References Allison, P.D. (1999) Multiple regression: a primer. Pine Forge Press, Thousand Oaks, CA. Chadd, E.A., Clancy, E.M., & Mowe, J.H. (2005). A physical habitat index focusing on salmonids in the Pacific Northwest, Streamkeepers of Clallam County, unpublished manuscript. Hall, L.W., Morgan, R.P., Perry, E.S., & Waltz A. (1999). Development of a provisional physical habitat index for Maryland freshwater streams. Maryland Department of Natural Resources. Johnson, R.A. & Wichern, D.W. (2002) Applied multivariate statistical analysis. Prentice Hall, Upper Saddle River, NJ. Karr, J.R., K.D. Fausch, P.L. Angermeier, P.R. Yant, and I.J. Schlosser (1986). Assessing biological integrity in running waters, a method and its rational. Illinois Natural History Survey, Special Publication 5. Karr, J.R. & Chu, E.W. (1999). Restoring Life in Running Waters: Better Biological Monitoring. Island Press, Washington, D.C. McBride, M. (2001). Spatial effects of urbanization on physical conditions in Puget Sound Lowland streams. Unpublished Master s thesis, University of Washington. Wang, L., Lyons, J., Kanehl, P., & Gatti, R. (1997). Influences of watershed land use on habitat quality and biotic integrity in Wisconsin streams. Fisheries, 22, Wang, L., Lyons, J., & Kanehl, P. (1998) Development and evaluation of a habitat rating system for low-gradient Wisconsin streams. North American Journal of Fisheries Management, 18,

10 Table 1. Correlations relating B-IBI to physical habitat variables. For most variables, including the B-IBI, the sample size was 45; otherwise the sample size is indicated in parentheses. Statistically significant correlations are indicated with bold type. Physical Variable Correlation Coefficient (r) Probability Pools Fine Sediment Large Woody Debris Conifer Stem Count Winter Canopy Summer Canopy Aggradation/Degradation (38) Bank Stability (37) Embeddedness (42) Aquatic Noxious Weeds Riparian Noxious Weeds

11 Figure 1. Forty five streams sites for which physical habitat and B-IBI data were obtained. 11

12 Figure 2. Line plot relating B-IBI the PHI-reg. The scatter plot was best fit with a quadratic equation Fitted Line Plot BIBI = PHI-reg PHI-reg**2 S R-Sq 54.7% R-Sq(adj) 52.5% BIBI PHI-reg

13 Figure 3. Loading plot or Gabriel s biplot of the first and second principal component of the physical variable data. The lines or vectors represent projections of the original variables on a plane defined by the first two components, and their lengths are correlations between the original variables and the components represented in the plot. The longer the vector length, the closer is its representation of it in the plane of the two components. Vectors that are close together (separated by small angles) represent variables that are highly correlated. Vectors that are at right angles to each other represent variables that are uncorrelated. Vectors with angles greater than 90 degrees represent variables that are negatively correlated with maximum negative correlation occurring at 180 degrees. Vectors that are parallel to the component axes are highly correlated with those axes AcqNoxW WintCan 0.25 Second Component ConifSt SumCan RipNoxW LWD First Component

14 Figure 4. Scatter plot relating B-IBI to PHI-pca. The scatter plot was best fit with a quadratic equation. 50 Fitted Line Plot BIBI = (LWD-AcqNoxW)+(WintCan-RipNoxW) (LWD-AcqNoxW)+(WintCan-RipNoxW)**2 S R-Sq 55.5% R-Sq(adj) 53.3% 40 BIBI PHI-pca

15 Figure 5. Frequency histogram of B-IBI Frequency BIBI

16 Figure 6. B-IBI versus predicted value for cases greater than 30 (n=36) 50 S R-Sq 52.1% R-Sq(adj) 50.7% 45 BIBI Predicted (BIBI = FineSed LWD ConifSt RipNoxW) 16