I have a data set of car accidents. It includes information on accidents (like date and time), vehicles (like make and year), and drivers (like age and gender).The goal is to estimate whether a new feature of the car can help improve safety.
I have another database where I can find information on non-accident cars along with drivers' characteristics and driving history.
The treatment is the presence of the new feature of the car, which only depends on car features. The dependent variable is whether the car has an accident. In order to match non-accident cars to accident cars, which depend on accident features (like date, time and weather), car features and drivers features.
What matching strategy is appropriate in this case? There is a clear structure of the data, i.e., cars are clusters, and drivers are units within respective clusters. do I need to account for how drivers self-select into those clusters? This is likely to depend on the treatment, i.e., whether the car has the new feature or not.