Log-Binomial Model and Its Comparison with Suitable Models

• Activity Number: 137
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Shailendra Banerjee*+
• Companies: Centers for Disease Control
• Keywords: log-binomial ; common event ; out-of-bounds ; convergence ; modified Poisson

Abstract

Log-binomial model is a generalized linear model with binomial errors and a log link function. It follows from this model that the risk ratio or relative risk is simply the exponentiation of the product of the coefficient times the difference in covariate values. So, when the event is common, this model is appropriate for estimating relative risk. But this model gives narrower confidence interval, generates out-of-bounds probabilities for events and has convergence problems. In this study, we used several alternate models like logistic regression, Poisson regression, modified Poisson on a prostate cancer dataset with a probability of occurrence 0.38. Both Poisson and log-binomial models estimate relative risk for the risk factor compared to odds ratio for the logistic regression. The estimates for risks are larger with wider confidence intervals with logistic regression compared to that for log-binomial and modified Poisson regression models. Log-binomial model resulted in out-of-bounds probabilities and non-convergent situations. The models will be investigated further and their applications will be further verified with simulated and more observed data.