Characterization and reduction of variability in selection based on effect-size using association measures in cohort study of heterogeneous diseases

Epidemiological studies employ pairwise measures of association to quantify dependencies among diseases and exposures. The reliable use of these measures to draw conclusions about the underlying association strengths requires that the measures have no undue dependencies that systematically distort their values. These conditions are particularly relevant when multiple pairwise associations are compared and ranked in a cohort study, as in the case of construction of disease networks. Following an empirical approach and using disease diagnoses data from a large cohort of 5.5 million patients as a test set, we develop a comprehensive methodology to characterize the variability of measures in selection of disease pairs based on effect-size. Specifically, we define putative bias variables, and examine the distribution of the measure scores conditioned upon it. This procedure reveals systematic bias in widely used measures such as relative risk and correlation coefficient. In addition, we devise a novel measure family using a stochastic model for differential rate of development of diseases. We demonstrate marked reduction in above-mentioned bias with an appropriate choice from this measure family.