Subscribe to DSC Newsletter

When a car is said to have a high accident rate, do we consider the possibility that (maybe) the high rate is due to the fact that people who drive the car in question tend to be bad drivers on average? For instance a car purchased or used typically by young males will be more likely to be considered unsafe - not because the car is intrinsically unsafe (it could actually be the safest car), but because the drivers represent the worst segment of the population, in terms of car accidents.

One way to remove the bias is to check accident rates for drivers switching from an unsafe to a safe car. If the accident rate is lower after switching to a safer car, then maybe the issue is with the car, not with the driver (unless the driver improved its driving skills over time, an effect that could be tested as well).

Views: 398

Reply to This

Replies to This Discussion

That's some intersting stuff Vincent. Insurance companies have data on accident rates by age, gender, race, ethnicity, income... of drivers. You propose a good way to check for whether a car is safe/unsafe, or whether its the drivers of the car that are safe/unsafe. Another (and perhaps easier) way though would be to model accident likelihood, using historical data, and all of the above demographics. Once modeled using driver demographics, it would be a matter of adding dummy variables for the car make(s) you want to test. If the coefficient for any 'dummy car make variable' that you test is not statistically different from zero, that would tell you that the safety issue is not with the car, but instead with the drivers of the car. If you find a car dummy coefficient that is in fact significant, the safety of that car would be in question.
I dont know whether this variable should be learned this way by insurance companies.

In my point of view the safety of cars results from its technical capabilities i.e. time to get from a given speed to halt or behavior on rainy streets. These facts are often evaluated by independent testers (i.e. journalists of car magazines). Hence they are given and can be included directly into the model.

Beside: A variable indeed influencing the accident rate is the maximum horsepower of the car. In a country without speed limit this multiplies with the inverse of the age of the male drivers ;)
I agree with you Steffen that there are still cars that are more safe and less safe than other cars due to their technical capabilities. Vincent was trying to explore whether there might be other influencing factors, aside from the car's technical specs/capabilities. With the variable that you suggest (maximum horsepower), you would still want to test whether its the horsepower that causes the car to be unsafe, or if it is the type of person who is drawn to a car with more horsepower. I suspect that you would find maximum horsepower to still be significant in the model which would indicate that there is at least one factor that causes a car to be more or less safe than other cars. Next step would be to see if there were car makes for which the 'car make dummy' is still significant in the model, even after taking max horsepower into account.
Hi, Apart from socio & technical parameter. may be added some unconventional parameter to enhance the reliability & to reduce the dimension of risk of the model may be alcaholic or non alcaholic, wearing shoes or any other, speed of the vehicle, with specs or without specs so on and so forth may not look sound but we have experienced in our crddit scoring model.

Thanks.
Biswajit.
Great points Biswajit. The more information, the better. The only problem with some of the variables (wearing shoes etc...), is that you might have that information from accident reports in some cases, but even if you had that information for a accidents, you wouldn't have it typically for the cars that are not in accidents (which you would need if you wanted to include that info in the model). I'm sure this forum can come up with some interesting ways around that.
There should be no bias. Cars do not "cause" accidents, people do. Cars are rated predominently by crash test data. see ---> http://www.hldi.org/. Of course there are other factors which affect insurance premium, such as age and gender. Your hypotheses to check for bias, by measuring the change to a safer car would be hard to do. Accident rates are relatively low, and you would have to deal with the fact the the driver would probably be moving into another age rating category by the time they switched.

-Ralph Winters
I suppose that what you mean is cars cause very few accidents: in most cases, the accident is due to human errors. It would be interesting to know what's the proportion of car accidents caused by mechanical failures, as opposed to human errors. I suppose it's below 1%, but I'm not sure.
The number of accidents due to mechanical failures are very small (although they can be very severe). I'm not sure whether you can measure this directly. However take a look at the JD Power Reliability survey as well as Consumer Unions reliability reports and you might be able to cull some information from that.

-Ralph Winters
no car would be allowed onto the road without some minimum safety features which should be enough to cause an accident rate low enough to be legally acceptible . given that, if some car has a high accident rate, its either due to some hidden internal fault or some external factor: kind of people who drive, kind of situations driven in, kind of terrains driven in.. so how to know wwhether to pull up a car ffor technical checks or ascribe the accidents to external factors? i guess the answer lies in the fact that safety features included to minimise chances of accidents are not abstract features. they are concrete accessories. we can group together cars by their safety features and run some tests. like, car A and car B has similar safety features, but car A is primarily driven in terrain A and car B in rougher terrain B. if car B has a significantly higher accident rate, then terrain is to blame. complications can be added. now if all cars in a group have similar external factors still only one has siginificantly higher accident rate, then it is perhaps a good idea to ggo for an internal checkup, although it is difficult to pin down exactly similar external factors. a lab experiment is in order.
I like the thought, but you are still biasing yourself to the sample of people who "had car A and now have car B" (or vice versa, you'd have to check going from safe to unsafe too). That's slightly different than an experiment where switching from car A to car B is forced on the subject.

But I do agree that some cars are safer to drive than others simply -- there are lots of factors: visibility, stick/automatic, braking, flipping, etc. Many are measurable by companies, but even those quantified values need to be mapped somehow into a safety space, which of course depends on your driver.
There may be some more sophisticated, albeit obvious factors, as the region. For example some cars are fashionable among some kind of people (young generation, for example) living in the crowded cities of average-sized towns with roads of worse condition than average etc.

Some 15 years ago I met such case; surprisingly the residents were rather rich, and most fashionable cars for youth were also above average standard :-).

RSS

On Data Science Central

© 2019   AnalyticBridge.com is a subsidiary and dedicated channel of Data Science Central LLC   Powered by

Badges  |  Report an Issue  |  Privacy Policy  |  Terms of Service