A simulation/case study under the use of robust Poisson and log-binomial model with generalized estimating equation models regarding non-independent data / 中华流行病学杂志

To explore the appropriate method in estimating relative risk (RR)/prevalence ratio (PR) related to non-independent datasets.The simulation datasets generated by computer and case study were analyzed by two generalized estimating equation (GEE) models to investigate and compare the related applicability.Both convergence effects of log-binomial-GEE model and Robust Poisson-GEE model were almost 100％.The estimation results of the two GEE models were both closer to the true value.95％CI coverage of the two GEE models increased along with the reduction of class aggregation or the increase of the number of categories.Robust-Poisson-GEE model seemed to be more stable and steady than the log-binomial-GEE.The two GEE models could correctly evaluate the effects of exposure on the outcome in the case study.Rarely,there appeared problems on convergence of Robust Poisson or log-binomial-GEE model,and the accuracy was high.Both models could be used to estimate the RR/PR on non-independent epidemiological data.

Using log-binomial model for estimating the prevalence ratio / 中华流行病学杂志

[Introduction] To estimate the prevalence ratios, using a log-binomial model with or without continuous covariates. Prevalence ratios for individuals' attitude towards smoking-ban legislation associated with smoking status, estimated by using a log-binomial model were compared with odds ratios estimated by logistic regression model. In the log-binomial modeling, maximum likelihood method was used when there were no continuous covariates and COPY approach was used if the model did not converge, for example due to the existence of continuous covariates. We examined the association between individuals' attitude towards smoking-ban legislation and smoking status in men and women. Prevalence ratio and odds ratio estimation provided similar results for the association in women since smoking was not common. In men however, the odds ratio estimates were markedly larger than the prevalence ratios due to a higher prevalence of outcome. The log-binomial model did not converge when age was included as a continuous covariate and COPY method was used to deal with the situation. All analysis was performed by SAS. Prevalence ratio seemed to better measure the association than odds ratio when prevalence is high. SAS programs were provided to calculate the prevalence ratios with or without continuous covariates in the log-binomial regression analysis.