USGS

Procedures for Adjusting Regional Regression Models of Urban-Runoff Quality Using Local Data

U.S. Geological Survey, Open-File Report 93-39

by A.B. Hoos and J.K. Sisolak

This report is available as a pdf below


Abstract

Statistical operations termed model-adjustment procedures (MAP’s) can be used to incorporate local data into existing regression models to improve the prediction of urban-runoff quality. Each MAP is a form of regression analysis in which the local data base is used as a calibration data set. Regression coefficients are determined from the local data base, and the resulting ‘adjusted’ regression models can then be used to predict storm-runoff quality at unmonitored sites. The response variable in the regression analyses is the observed load or mean concentration of a constituent in storm runoff for a single storm. The set of explanatory variables used in the regression analyses is different for each MAP, but always includes the predicted value of load or mean concentration from a regional regression model. The four MAP’s examined in this study were: single-factor regression against the regional model prediction, P, (termed MAP-lF-P), regression against P,, (termed MAP-R-P), regression against P, and additional local variables (termed MAP-R-P+nV), and a weighted combination of P, and a local-regression prediction (termed MAP-W).

The procedures were tested by means of split-sample analysis, using data from three cities included in the Nationwide Urban Runoff Program: Denver, Colorado; Bellevue, Washington; and Knoxville, Tennessee. The MAP that provided the greatest predictive accuracy for the verification data set differed among the three test data bases and among model types (MAP-W for Denver and Knoxville, MAP-lF-P and MAP-R-P for Bellevue load models, and MAP-R-P+nV for Bellevue concentration models) and, in many cases, was not clearly indicated by the values of standard error of estimate for the calibration data set. A scheme to guide MAP selection, based on exploratory data analysis of the calibration data set, is presented and tested.

The MAP’s were tested for sensitivity to the size of a calibration data set. As expected, predictive accuracy of all MAP’s for the verification data set decreased as the calibration data-set size decreased, but predictive accuracy was not as sensitive for the MAP’s as it was for the local regression models.

Table of Contents

PDF Files




FirstGov button  Take Pride in America button