Adjusted Quasi-least Squares for Analysis of Correlated Binary Data

• Activity Number: 400
• Type: Contributed
• Date/Time: Wednesday, August 10, 2005 : 10:30 AM to 12:20 PM
• Sponsor: ENAR
• Author(s): Justine Shults*+ and Wenguang Sun
• Companies: University of Pennsylvania and University of Pennsylvania
• Keywords: quasi-least squares ; correlated binary data ; generalised estimating equations

Abstract

There has been controversy in the statistical literature regarding the potential and impact of violation of bounds for the correlation in analyses of correlated binary data: It is well-known that the correlation among binary outcomes is constrained by the marginal means, yet statistical approaches such as generalized estimating equations (GEE) do not check that these conditions are satisfied. In this presentation, we first describe when the violation of bounds is likely to be a problem, especially with regard to sample size, value of the correlation, misspecification of the underlying correlation structure, and violation of an assumption of constant correlation across all subjects. We then propose and demonstrate two adjustments to quasi-least squares (QLS) (an approach in the framework of GEE) that can be helpful in overcoming the potential problem of violation of bounds. The first is to modify the QLS estimation procedure so it yields an estimate of correlation that does indeed satisfy the constraints. The second is to allow the correlation to vary according to subject level covariates.