Comparison of Logistic Regression and Discriminant Analyses for Stock Identification of Anadromous Fish, with Application to Striped Bass (Morone saxatilis) and American Shad (Alosa sapidissima)
We examined the applicability of logistic regression to stock identification studies and compared its performance on two data sets to that of linear and quadratic discriminant functions. Logistic regression can be used to model a categorical dependent variable associated with continuous or discrete independent variables, and is preferred to discriminant analyses when the explanatory variables are not multivariate normal. Our examples were American shad (Alosa sapidissima) from the Connecticut River and Hudson River estuaries, and striped bass (Morone Saxatilis) from the Hudson River, Chesapeake Bay, and Roanoke River estuaries. In the examples we used a resampling method to assess classification and allocation errors of the two methods on new data. For the shad data, the logistic model classified significantly more fish correctly, and provided a significantly better estimate of stock composition. For the striped bass data, the two methods classified about the same proportion of fish correctly, but the logistic model gave a significantly better estimate of stock composition.