Comparison of three statistical procedures for classifying the presence-absence of an aquatic macrophyte from microhaibitat observations

Abstract

Although a variety of statistical techniques exist to evaluate the presence-absence of taxa relative to habitat parameters, few studies have compared the performance of multiple techniques. We used multiple logistic regression (MLR; full and reduced models), linear discriminant function analysis (LDFA), and nearest neighbor discriminant analysis (NNDA) to assess the presence-absence relationship between eelgrass (Vallisneria americana) and microhabitat predictors in the upper Chipola River, Florida. In addition, we performed variable selection procedures to assess if MLR and LDFA identify the same variables as most discriminating. Data were collected and analyzed separately from four sampling events (summer and fall, 1999 and 2000), and their performance was judged using Cohen's kappa. Linear discriminant function analysis was consistently the poorest classifier, whereas MLR (full and reduced models), and NNDA always performed well and comparably. However, with the exception of the poor performance of LDFA, no technique performed with relative consistency. The full MLR model classified observations slightly better than the reduced model MLR. Variable selection procedures of MLR and LDFA selected different sets of discriminating variables for three out of four sampling events. We have shown that statistical classification techniques may perform relatively inconsistently with highly similar samples. Future presence-absence studies should examine multiple models and compare them using Cohen's kappa to determine which discriminating model may be best for their application. © 2002, Copyright Taylor & Francis Group, LLC.