Constructing and Testing Logistic Regression Models for Binary Data: Applicationsto the National Fire Danger Rating System

摘要

Logistic regression was used in examining the relationship between National FireDanger Rating System (NFDRS) indexes and historical fire occurrence data. Basic techniques of constructing and testing logistic regression models are presented at a modest mathematical level. The emphasis of the paper is on tests of fit for logistic regression models. Specific results of the study are reported elsewhere. The explanatory (independent) variable is an NFDRS index. The response (dependent) variable is fire-day, which has value 1 if one or more fires occur in an area of concern on a certain day and 0 otherwise. The application differs from classic dose/response studies in several ways: there is no control over the 'dose', the explanatory variable; indexes may range as high as 200 so it is necessary to group data into categories for tests of fit; and quick and easy methods for doing tests of fit are required for comparing 100 logistic regression models generated from the same set of weather data (20 fuel models x 5 indexes).