Estimating adjusted NNT measures in logistic regression analysis

Authors

  • Ralf Bender,

    Corresponding author
    1. Department of Medical Biometry, Institute for Quality and Efficiency in Health Care (IQWiG), Cologne, Germany
    2. Faculty of Medicine, University of Cologne, Germany
    • Department of Medical Biometry, Institute for Quality and Efficiency in Health Care (IQWiG), D-51105 Cologne, Germany
    Search for more papers by this author
    • Statistician.

  • Oliver Kuss,

    1. Institute for Medical Epidemiology, Biostatistics, and Informatics (IMEBI), Martin-Luther-University Halle-Wittenberg, Halle (Saale), Germany
    Search for more papers by this author
  • Mandy Hildebrandt,

    1. Department of Medical Biometry, Institute for Quality and Efficiency in Health Care (IQWiG), Cologne, Germany
    2. Institute for Medical Biometry, Epidemiology and Informatics (IMBEI), Johannes-Gutenberg-University Mainz, Mainz, Germany
    Search for more papers by this author
  • Ulrich Gehrmann

    1. Department of Medical Biometry, Institute for Quality and Efficiency in Health Care (IQWiG), Cologne, Germany
    Search for more papers by this author

Abstract

The number needed to treat (NNT) is a popular measure to describe the absolute effect of a new treatment compared with a standard treatment or placebo in clinical trials with binary outcome. For use of NNT measures in epidemiology to compare exposed and unexposed subjects, the terms ‘number needed to be exposed’ (NNE) and ‘exposure impact number’ (EIN) have been proposed. Additionally, in the framework of logistic regression a method was derived to perform point and interval estimation of NNT measures with adjustment for confounding by using the adjusted odds ratio (OR approach). In this paper, a new method is proposed which is based upon the average risk difference over the observed confounder values (ARD approach). A decision has to be made, whether the effect of allocating an exposure to unexposed persons or the effect of removing an exposure from exposed persons should be described. We use the term NNE for the first and the term EIN for the second situation. NNE is the average number of unexposed persons needed to be exposed to observe one extra case; EIN is the average number of exposed persons among one case can be attributed to the exposure. By means of simulations it is shown that the ARD approach is better than the OR approach in terms of bias and coverage probability, especially if the confounder distribution is wide. The proposed method is illustrated by application to data of a cohort study investigating the effect of smoking on coronary heart disease. Copyright © 2007 John Wiley & Sons, Ltd.

Ancillary