2.3. Statistical analysis

We used log-binomial regression to estimate fatality risk ratios for each helmet type compared with full-face helmets. We included rider age and sex as covariates. The following covariates were considered individually and retained if their overall p was 0.15 or less or if they altered any helmet type coefficient by 10% or more: elevated blood alcohol content, motorcycle speed, operator vs passenger status, motorcyclist culpability, motorcycle type, motorcycle brand, collision type, state highway collision occurrence, and weekend occurrence. These variables were selected for consideration because they seemed likely to be associated with both helmet type and fatality risk. We used directed acyclic graphs to facilitate the identification of potential confounders. Operator status, motorcycle brand, and motorcycle type were rejected; all other variables were included in the model. We modeled age and motorcycle speed as continuous, quadratic, and categorical. Each of these models produced nearly identical helmet risk ratios. The categorical model was selected to facilitate interpretation. We considered effect measure modification by including product terms between the helmet type variable and selected covariates. We used standard regression diagnostics to check for overly influential data points. All data management and analysis was implemented with Stata 14 software (StataCorp, 2014).