Statement of purpose The estimation of odds ratios (OR) is widely used in cohort or cross-sectional injury studies as well as in other public health fields. We compared the performance of Methods to estimate risk ratios (RR) directly using simulated data and data from an ongoing injury cohort study.
Methods/approach We used data from a cohort of collision-involved motorcyclists in California. The binary outcome was head injury and the binary predictor was the use of a novelty motorcycle helmet. We estimated RRs using log-binomial regression, maximum likelihood (ML) Cox regression, and Mantel-Haenszel (MH) estimation, and we estimated ORs using ML logistic regression. The log-binomial regression was fitted using (1) a Poisson procedure with robust errors, (2) a ML general linear models (GLM) procedure, and (3) an iterated re-weighted least squares (IRLS) GLM procedure.
Results Of the 8,086 motorcyclists in our cohort study, 319 were using a novelty helmet and 7,767 were using other types of helmet at the time of collision. The probability of head injury was 0.351 among novelty helmet users and 0.138 among users of other helmets. All procedures for log-binomial regression with robust errors gave the same results (RR 2.537, 95% CI 2.163–2.974). Point and interval estimates closely matched the Cox regression and MH estimates. The logistic regression model produced an OR of 3.368 (95% CI 2.653–4.277).
Conclusions Methods for estimating RRs worked well in this dataset. All RR Methods estimated a head injury RR of 2.5 for motorcyclists wearing novelty helmets, compared with other types of helmets. Logistic regression produced an OR of 3.5. Further analysis will be performed on simulated data sets.
Significance/contributions The long-standing practice of estimating ORs with cohort or cross-sectional data may not be justifiable, given the available RR Methods, and the improved interpretability and the desirable collapsibility of RRs over ORs.