J Water Health 01 (2003) 209-223

Assessment of statistical methods used in library-based approaches to microbial source tracking

Kerry J. Ritter, Ethan Carruthers, C. Andrew Carson, R. D. Ellender, Valerie J. Harwood, Kyle Kingsley, Cindy Nakatsu, Michael Sadowsky, Brian Shear, Brian West, John E. Whitlock, Bruce A. Wiggins and Jayson D. Wilbur

Southern California Coastal Water Research Project, Westminster, CA 92863, USA Tel:+1/7143729210 Fax:+1/8949699 E-mail: kerryr@sccwrp.org

Department of Soil, Water & Climate, University of Minnesota, St Paul, MN 55108, USA

Department of Veterinary Pathobiology, University of Missouri, Columbia, MO 65211, USA

University of Southern Mississippi, Hattiesburg, MS 39406, USA

Department of Biology, University of South Florida, Tampa, FL 33620, USA

Applied-Maths, Inc., 512 East 11th Street, Suite 207, Austin, Texas 78701, USA

Department of Agronomy, Purdue University, West Lafayette, IN 47907 USA

Department of Soil, Water & Climate, University of Minnesota, St Paul, MN 55108, USA

Department of Veterinary Pathobiology, University of Missouri, Columbia, MO 65211, USA

Applied-Maths, Inc., 512 East 11th Street, Suite 207, Austin, Texas 78701, USA

Division of Math and Science, Hillsborough Community College, Tampa, FL 33614, USA

Department of Biology, James Madison University, Harrisonburg, VA 22807, USA

Department of Mathematical Sciences, Worcester Polytechnic Institute, Worcester, MA 01609, USA

Several commonly used statistical methods for fingerprint identification in microbial source tracking (MST) were examined to assess the effectiveness of pattern-matching algorithms to correctly identify sources. Although numerous statistical methods have been employed for source identification, no widespread consensus exists as to which is most appropriate. A large-scale comparison of several MST methods, using identical fecal sources, presented a unique opportunity to assess the utility of several popular statistical methods. These included discriminant analysis, nearest neighbour analysis, maximum similarity and average similarity, along with several measures of distance or similarity. Threshold criteria for excluding uncertain or poorly matched isolates from final analysis were also examined for their ability to reduce false positives and increase prediction success. Six independent libraries used in the study were constructed from indicator bacteria isolated from fecal materials of humans, seagulls, cows and dogs. Three of these libraries were constructed using the rep-PCR technique and three relied on antibiotic resistance analysis (ARA). Five of the libraries were constructed using Escherichia coli and one using Enterococcus spp. (ARA). Overall, the outcome of this study suggests a high degree of variability across statistical methods. Despite large differences in correct classification rates among the statistical methods, no single statistical approach emerged as superior. Thresholds failed to consistently increase rates of correct classification and improvement was often associated with substantial effective sample size reduction. Recommendations are provided to aid in selecting appropriate analyses for these types of data.

Full article (PDF Format)