Estimations of relative risk: Simulation comparisons of popular methods in SAS and R.

摘要

Relative risk (RR)/Prevalence Ratio (PR) is an important measurement of the association between the outcome and the exposure factor. Compared with the odds ratio (OR), RR/PR is more intuitive and easier to explain to the general audience. But due to the constrained parameter space which may result in convergence problems in estimating RR/PR, it is not as widely used as the OR. Recently, there appears a lot of discussion on how to solve the non-convergence problem of estimating RR/PR in the log binomial model, which motivates the simulation study in this thesis. Specifically, we compare the most popular methods of estimating RR/PR, which include the crude method, the conversion method, the Inverse-Probability-of-Treatment Weight (IPTW) method, COPY method and the Non-Linear Programming (NLP) method. We explore the robustness of each method by intensive simulation studies. The factors we investigate include different sample sizes, different values of RR/PR, different P0 for the conversion method, different choices of C values for the COPY method, and the initial values for the non-linear programming method. We make recommendation of the best performed method based on the accuracy of estimation, length of computational time, and easiness of programming. Finally the recommended method is applied to the NHANES data, which reveals the relationship between the Vigorous Physical Activity and adult obesity.