Police crash reports are often the main source for official data in many countries. However, with the exception of fatal crashes, crashes are often underreported in a biased manner. Consequently, the countermeasures adopted according to them may be inefficient. In the case of bicycle crashes, this bias is most acute and it probably varies across countries, with some of them being more prone to reporting accidents to police than others. Assessing if this bias occurs and the size of it can be of great importance for evaluating the risks associated with bicycling. This study utilized data collected in the COST TU1101 action "Towards safer bicycling through optimization of bicycle helmets and usage". The data came from an online survey that included questions related to bicyclists' attitudes, behaviour, cycling habits, accidents, and patterns of use of helmets. The survey was filled by 8655 bicyclists from 30 different countries. After applying various exclusion factors, 7015 questionnaires filled by adult cyclists from 17 countries, each with at least 100 valid responses, remained in our sample. The results showed that across all countries, an average of only 10% of all crashes were reported to the police, with a wide range among countries: from a minimum of 0.0% (Israel) and 2.6% (Croatia) to a maximum of a 35.0% (Germany). Some factors associated with the reporting levels were type of crash, type of vehicle involved, and injury severity. No relation was found between the likelihood of reporting and the cyclist's gender, age, educational level, marital status, being a parent, use of helmet, and type of bicycle. The significant under-reporting - including injury crashes that do not lead to hospitalization - justifies the use of self-report survey data for assessment of bicycling crash patterns as they relate to (1) crash risk issues such as location, infrastructure, cyclists' characteristics, and use of helmet and (2) strategic approaches to bicycle crash prevention and injury reduction.

The use of logistic regression models in data analysis and machine learning has expanded in recent years and has become the primary preference of researchers in risk assessment studies across a wide range of scientific fields. From the assessment of credit risk in financial institutions to the estimation of risk factors for traffic accidents or the identification of etiological factors for chronic diseases. All logistic models are natural extensions of the simple binary model, and their interpretation is based on it. Using the data of a cross-sectional study on the risk factors of traffic collisions, the two main extended models of logistic techniques, multinomial and ordinal logistic regression, are presented in the article in detail. Emphasis is placed on the use of ordinal regression since the outcome variable of the collision data is defined as ordinal measurement reflecting a latent continuous scale.