Generally said, I plan to validate a specific indicator X (derived from driving behaviour) which should be linked to traffic safety (accidents) on level of road sections. Validation should statistically prove this relationship. I can think of various ways (i.e. possible study designs), but I would appreciate your opinions and advices.
Option 1 - I can define, based on theory or experience, two conditions when X is safe and when is unsafe. This can be compared to accidents / no accidents.
Option 2 - I can collect X on many sections and link it with accident history.
Option 3 - I can choose the critical sections (based on accidents) and collect X there - thus I should obtain "unsafe" values of X.
Option 4 - naturalistic driving study = studying the drivers behaviour visually. When they behave "unsafely", values of X will be "unsafe".
These are only my theories - please feel free to comment! I will welcome discussion, references, etc. Thank you.
I like options 1 and 2. I would discard options 3 and 4. In an ideal world I would add data from a test track. This could be a driving course where people are presented with different problems and you can correlate response to those problems with indicator X. This could also be a driving simulator -- there are plenty of driving video games. Maybe you could get someone to create a different version of a driver's education driving simulator to focus on this issue. The beauty of the test course and videogame options is that you can control for all the other variables. Angle of the sun, wind, wet roads, car type, and other traffic can all be adjusted while the risk of injury is controlled. That said, it is not exactly realistic, and the nature of a test will tend to make drivers more alert than the typical stressed zombie driving home from work while chatting on their cell phone and wondering what is for dinner.
Your option 3 does not give a context for interpreting the "unsafe" values of X. I am not quite clear on option 4, but it sounded like it would have the same issue.
Option 1 would a validation exercise indeed. You can do this very simply by correlating the model and real outcome. There are several statistical methods depending on what outcomes you are comparing. One example I have is EuroRAP star rating vs. crash rates - two different variables, not directly comparable.
If you went with Option 2, e.g. a statistical modeling (accident prediction model), then you can use various residual plots as long as you got an independent sample not used in modelling. There are more sophisticated methods than that of course.
Option 3 could be modified by adding data from 'safe' sites. Then you are looking cross-tab analysis, the predecessor of statistical models. See answer for Option 2.
Option 3 - you are looking at proxy measures. As far as I understand this subject, it is the next frontier in road safety. Great for developing understanding of combined human, vehicle and road factors involved in crashes, i.e. everything that current statistical models do not do. The near future.
I like more the second approach (but I think you can combine it also with the first). If I would have to study the effect of the X factor on traffic safety having different (pre-defined) road sections levels based on experimental data, I would try an ANOVA analysis which would allow you to discriminate the effect size (strength) across levels.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis