Background Researchers have used case-control study design to examine the influence of road infrastructure on bicycle-related injuries. This design compares cases (road sites or individual cyclists with an injury) with controls (sites or individuals with no injury) to identify risk factors for cyclists. Unlike randomised controlled trials, these non-randomised observational studies, such as case-control, are prone to biases. There has been no assessment of the methodological quality of such studies in the context of road safety.

Objective We aimed to assess the risk of bias in case-control studies investigating infrastructure-related risk factors for cyclists.

Methods We accessed studies from a recently conducted Evidence Gap Map of all effectiveness studies related to road safety interventions. We included case-control (and case-crossover) studies that identified road infrastructure risk factors for cycling injuries. We examined the risk of bias using “A Cochrane Risk of Bias Assessment Tool: for Non-Randomised Studies of Interventions” (ACROBAT-NRSI). The tool includes seven domains of risk-of-bias with domain-specific signalling questions (3–5 questions per domain). For example, in the domain, ‘bias due to confounding’, a signalling question is structured as ‘Did the authors use an appropriate analysis method that adjusted for all the critically important confounding domains?’. Next, we judged the domain-specific risk-of-bias based on the responses to these questions.

Results We identified and reviewed nine eligible studies, conducted between 2012 and 2021. Domains with the highest risk of bias were “bias due to confounding” because of inadequate control for bicycle volume as cycling exposure and “bias in measurement of outcomes” because of the use of underreported injury data. Among the studies, eight sufficiently described data collection of outcome that is replicable, five reported proportion of observations with missing data for each infrastructure characteristic, two were based on a published study protocol, two collected infrastructure data blinded to outcome status, and none reported sample size calculations. Across all the studies, controls were selected independent of the outcome of interest and the infrastructure characteristics.

Conclusions We found a high prevalence of potential sources of bias in the existing studies. We discuss methodological implications for future studies.