5141.0: Wednesday, October 24, 2001 - 12:30 PM

Abstract #20792

Analyzing landmine incidents via zero-inflated Poisson models

Lawrence H. Moulton1, Aldo A. Benini2, Charles E. Conley2, and Shawn Messick2. (1) Departments of International Health and Biostatistics, The Johns Hopkins University School of Public Health, 615 N. Wolfe St., Baltimore, MD 21205, 410-955-6370, lmoulton@jhsph.edu, (2) Global Landmine Survey, Survey Action Center, 2001 S St. NW, Suite 310, Washington, DC 20009

Zero-inflated Poisson (ZIP; also called “zero-altered” or “hurdle”) models are becoming popular in the health sciences for modeling counts of events. These models are formed by a mixture of two distributions: a point mass distribution at zero, and a Poisson (or overdispersed Poisson) distribution. ZIP models are particularly appropriate for modeling the occurrence of incidents caused by landmines. Some geographically-defined observational units do not have any exposure (no landmines) and others do; covariates of interest may be related to existence of landmines, or to distribution of the landmines, geography, population, or to all of these factors. Separate covariate vectors permit identification of variables related to landmine existence (or existence of “safety” factors) and to heightened risk among those areas with landmines. We describe analyses of two sets of mine incident data: from 327 districts of Kosovo, and from 592 communities in Yemen. Extensive GIS work provided some of the critical variables, and responses to standardized survey instruments provided others. We demonstrate the utility of fitting ZIP models in these settings, and of incorporating population measures directly in the estimated part of the linear predictor, as contrasted with using them as offset terms. Relationships to logistic regression results are discussed. Perhaps because of the relatively small numbers of mine-related events occurring per geographic unit, we did not observe extra-Poisson variability in the Poisson distribution components. The following website describes the Global Landmine Survey-- See www.vvaf.org/gls/index.shtml

Learning Objectives: At the conclusion of this presentation, listeners will be able to:

  1. 1. Describe the rationale for and components of a zero-inflated Poisson model.
  2. 2. Identify data-analytic features that call for the use of a ZIP model.
  3. 3. Recognize situations that require explicit estimation of the exposure-incidence relationship.

Keywords: Landmines, Statistics

Presenting author's disclosure statement:
Organization/institution whose products or services will be discussed: None
I have a significant financial interest/arrangement or affiliation with any organization/institution whose products or services are being discussed in this session.
Relationship: I am a paid consultant to the Global Landmine Survey.

Handout (.ppt format, 337.5 kb)

The 129th Annual Meeting of APHA