Recent research has shown that there exists large heterogeneity in car-following behaviour such that different car-following models best describe different drivers' behaviour. A literature review reveals that current approaches to calibrate and compare different models for one driver do not take the complexity of the models into account or are only able to compare a specific set of models. This contribution applies Bayesian techniques to the calibration of car-following models. The Bayesian framework promotes models that fit the data well but punishes models with a high complexity, resulting in a measure called the evidence. This evidence quantifies how probable each model is to be the model that best describes the car-following behaviour of a single driver. It can be computed for any car-following model. When considered over multiple drivers, the evidences can be used to describe the heterogeneity of the driving population. In an experiment seven different car-following models are calibrated and compared using a data set that was collected with a helicopter. The results indicate that multi-leader models better describe the car-following models even if their higher complexity is accounted for, and that for the description of microscopic driving behaviour the reaction time is essential; models without a reaction time perform significantly worse.