5163.0: Wednesday, November 15, 2000 - Board 6

Abstract #13541

Evaluating non-linearities in the exposure-response relationship using nonparametric smoothing and conditional logistic regression

Patricia A. Sullivan, MS1, Ellen E. Eisen, ScD2, David Kriebel, ScD2, Susan R. Woskie, ScD2, and David H. Wegman, MD, MPH2. (1) Division of Respiratory Disease Studies, National Institute for Occupational Safety and Health, 1095 Willowdale Road, Morgantown, WV 26505-2888, 304 285-5813, pcs5@cdc.gov, (2) Department of Work Environment, University of Massachusetts Lowell, One University Avenue, Lowell, MA 01854

This paper applies nonparametric smoothing techniques in exploratory epidemiologic analysis to help describe exposure-response relationships. Typically, dose-response models assume that the relation between exposure and response is linear on some scale. Many disease mechanisms, however, such as sensitization or carcinogenesis, may produce non-linearities in the dose-response curve. Moreover, linear models may be inappropriate in occupational epidemiology studies where the healthy worker effect can lead to an apparent plateau or even down-turn in risk among the more highly exposed. Occupational epidemiologists typically resort to categorical exposure variables to avoid linearity assumptions, but results are not robust to changes in cut-points. Nonparametric graphing methods make no a priori assumption about the shape of the exposure-response curve and so can identify empirical cut-points between homogeneous exposure categories. As illustrated using data from a study of stomach cancer risk among auto workers, exposure categories based on empirically identified cut-points were evaluated in conditional logistic regression models that controlled for confounding. Model fit was better and the risk estimates higher than in models based on traditional cut-points (selected a priori). For example, initial categorical analysis based on quartiles of the exposure distribution found an odds ratio of 1.4 (95% CI 0.8-2.5) in the highest category of exposure (>1.9 mg/m3). Empirical cut-points identified after smoothing resulted in a model with better fit, a higher cut-off for the highest exposure category, and an odds ratio of 1.9 (95% CI 1.0-3.6) among those exposed to at least 4 mg/m3. These methods have potential widespread application in epidemiologic analysis.

Learning Objectives:

  • Develop awareness of how choice of categorical cut-points can affect results of epidemiologic exposure-response analysis.
  • Assess the usefulness of nonparametric smoothing techniques in exploratory epidemiologic analysis.

    • Keywords: Epidemiology, Methodology

    Presenting author's disclosure statement:
    Organization/institution whose products or services will be discussed: None
    I do not have any significant financial interest/arrangement or affiliation with any organization/institution whose products or services are being discussed in this session.

    The 128th Annual Meeting of APHA