I heard a story recently in which someone said if they wanted to kill somebody (and get away with it) they would do it with their car. They cited various statistics about the number of auto-related deaths (including car-on-pedestrians) coupled with additional stats about the number of drivers actually sentenced to any sort of crime...blah, blah, blah.

My question is this: Is it statistically feasible to demonstrate that cars ARE (statistically speaking) actually used as weapons to commit murder?

In other words, I realize it may not be possible to demonstrate that any single car-on-pedestrian 'accident' was actually an attempted/committed murder. Rather, I'm wondering if a method might be imagined in which it could be demonstrated that some percentage of those 'accidents' are actually, in all likelihood, not accidents at all...

  • 2
    $\begingroup$ Was this on the Freakonomics podcast? $\endgroup$
    – Steve S
    Commented Sep 8, 2014 at 7:09
  • $\begingroup$ I can't easily imagine how: the difference between murder and killing is 'malice aforethought' (a mental or motivational state); murder and manslaughter are also distinguished by type of intention. Neither distinctions seem very amenable to statistical analysis. $\endgroup$ Commented Sep 8, 2014 at 9:03
  • $\begingroup$ I've adjusted the title to make it reflect your actual question. $\endgroup$ Commented Sep 8, 2014 at 9:06
  • $\begingroup$ Given how often people are run over on dirveways, I think it will be very hard. $\endgroup$ Commented Sep 8, 2014 at 9:11
  • $\begingroup$ While not specifically a statistical method, depending on where the accident happened, it's entirely possible for forensic science to determine that an accident could have been murder. It's not possible for all accidents, but depending on markings on the floor that indicate speed changes, eye witness reports of specific maneuvers, and other crime scene evidence, you can determine that certain incidents weren't actually accidents, but intentional attacks. Also, in a number or legal jurisdictions, there are people that have been convicted of manslaughter for hitting someone with their vehicle. $\endgroup$
    – Nzall
    Commented Sep 8, 2014 at 14:54

1 Answer 1


This may be a long-shot (practically speaking), but if you could get your hands on the (victim, driver) pairs and had a decent social network search engine, you could calculate the "degrees of separation" between the driver and victim and then construct a null distribution of "degrees of separation" by assuming random assignment of driver and victim from the local population where the accident occurred (e.g., everyone within typical commuting distance). This would correct for the "small town" effect, where everyone has close ties to everyone else.

The key hypothesis is: do the actual driver/victim pairs have fewer degrees of separation than the population at large? If so, it means that either (a) close acquaintances are somehow "synched" in their movements about town [e.g., demographic stratification] (b), at least some of the incidents appear to involve an unusually large number of close acquaintances.

Another approach would be to do logistic regression with "degrees of separation" as the variable, and "probability of accident/victim pariing" on the y axis. A strongly increasing function would suggest a "closeness" effect.

You would need to corroborate this by seeing if any of the "high relation" pairs actually resulted in a homicide trial and compare it to the overall rate of homicide indictments for pedestrian collisions.

  • 8
    $\begingroup$ Alternately, do the actual driver/victim pairs have the equivalent degrees of separation as known murderer/victim pairs? $\endgroup$
    – Alexis
    Commented Sep 8, 2014 at 5:54
  • $\begingroup$ @Alexis Great suggestion! My only concern is the "dilution" effect...most pedestrian hits are probably not premeditated (i.e., they really are accidents), so I doubt the overall test of equality of means for separation would show they are similar to that for the class of murders. However, your suggestion would be very useful if we envision the population of driver/victim pairs as a mixture of what are essentially murders and actual accidents. Then we could perform inference on the mixing parameter :-) Thanks for the great suggestion!! $\endgroup$
    – user31668
    Commented Sep 8, 2014 at 12:54
  • $\begingroup$ Two points. Foremost: your concern is a great example of why to consider what effect size is large enough to be relevant while using a priori power analysis to plan sample sizes. Next: did you notice my subtle insinuation of an hypothesis appropriate to equivalence testing (and relevance testing). $\endgroup$
    – Alexis
    Commented Sep 8, 2014 at 14:40
  • $\begingroup$ @Alexis points well taken! Thanks for clarifying...I missed your insinuation $\endgroup$
    – user31668
    Commented Sep 8, 2014 at 15:15
  • $\begingroup$ "Everyone within commuting distance" likely isn't a good proxy for "everyone on the route". Just the mere fact of visiting a friend is going to cause you to drive in close proximity to their neighbors. Any sort of by-invitation event is going to have a very high concentration of closely acquainted people driving in proximity. $\endgroup$
    – Ben Voigt
    Commented Sep 8, 2014 at 16:44

Not the answer you're looking for? Browse other questions tagged or ask your own question.