This post links to a Google Sheet containing a quick investigation into the accuracy of Wikipedia's figure for miles per micromort (230) for US drivers, when accounting for preventative behaviors.

The following are the main outcome estimates:

Miles per micromort, no adjustments, in US (2019)91
-- If excluding motorcycles105
-- If excluding motorcycles and pedestrians, pedalcyclists, and other nonoccupants137
-- Amongst passenger vehicle occupants only132
-- Amongst passenger vehicle occupants only, if setting single-car crashes to 0235
-- Amongst passenger vehicle occupants only, if approximating the seatbelt-wearing only rate245
-- Amongst passenger vehicle occupants only, if setting single-car crashes to 0 and approximating the seatbelt-wearing only rate442
-- Amongst passenger vehicle occupants only, if setting single-car crashes to 0 and approximating the seatbelt-wearing only rate and if setting alcohol-impaired, drowsiness-associated, and distraction-associated deaths to 50% of current level (as an approximation of controlling one driver's behavior in two driver crashes)548

This rapid (~1.5 hrs including documentation) investigation was funded by Ruby Bloom via the Bountied Rationality FB group.

New Comment
15 comments, sorted by Click to highlight new comments since: Today at 3:53 AM

The quick examination didn't get into this in the final numbers, but I feel confident that time of day (day vs night) is a big deal. Deaths during the nighttime were roughly the same as daytime deaths, but I'd assume most of the driving happens during the day, and disproportionately deaths are at night, both for visibility or sleepiness reasons.

I would advise people against driving through the middle of the night. Even if you don't feel tired, it's not good to go against your circadian rhythms. 

Driving at night is not just about your own tiredness/circadian rhythm, there are other people driving tired and drunk. 

In my college town, there's a 1/4 mile long plastic-fenced in road, leading to rental houses. Every Thursday-Sunday night, someone will run their car into this fence, leaving broken fence marks the next morning. 

A dialogue between myself and Ruby that may be of interest (shared with permission): 

Ruby: A question: why do you set single-car crashes to zero? 

My response: It seemed you were interested in something like "if you're a safe person, how safe is driving", and I thought single-car accidents may be particular indicators of being 'unsafe' in some ways. I'd be happy to add calculations that include single-car accidents as well. 

Ruby: Maybe, but there are reasons why a safe driver might be a single-car crash too: 
- hit a pot hole, 
- lost traction in bad weather (rain, snow) 
- swerved out of the way of a another car (is that 1 car or 2 car crash?) or out of a pedestrian/animal/whatever. 
- general car malfunction (tire blown, steering, breaks) 

My response: Yeah I think what constitutes a 'safe' driver is pretty unknown, and I wasn't ultimately sure what adjustment to make. A perfectly safe driver, for instance, could arguably prevent each of these examples. Additionally, it's likely an oversimplification to remove all of a single driver's share of distracted and drowsy driving crashes, as there's likely some percentage of those that are unavoidable.


A perfectly safe driver, for instance, could arguably prevent each of these examples. 

A perfectly safe driver, hopefully could. My motivation for wanting to estimate micrmorts was trying to estimate the conversion between microCovids and micromorts for people in the Rationality community. Hence, I'm most interested in the level of safety Rationalists typically have (e.g. seatbelts, not drunk driving) vs population average, so this is pretty good.

If I could, I'd get every Rationalist driver to drive vigilantly ("defensively"), take a car-control/defensive-driving course (e.g. this one), keep their car well-maintained (e.g. tires inflated adequately,  inspected at least once a year). But most people's miles are in Uber/Lyft, which is maybe pretty good since those drivers are mostly (a) very practiced, (b) forced to service their cars more reguarly.

Additionally, it's likely an oversimplification to remove all of a single driver's share of distracted and drowsy driving crashes, as there's likely some percentage of those that are unavoidable.

I wish I could say I never drove when tired. 

My personal experience in Uber/Lyfts (pre-covid, I used them multiple times a week for several years) is that they're probably more dangerous than driving myself (>80% of rides are very safe / normal, but the 1-20% where I think they're driving too fast or recklessly seem like most of the risk). I personally would be happy to pay 10% more to guarantee a safer driver, especially on e.g. a rainy day. I think I probably have more experience driving than most Berkeley EA's though / feel more confident in my driving skills.

But most people's miles are in Uber/Lyft

This is interesting to me, as every time I've looked at Uber/Lyft prices in my area it has seemed a bit high for it to be my go-to option. Can you link me to a good discussion regarding why this is the typical Rationalist choice? (I've read a lot of the sequences, etc. but really don't spend hardly any time on the blog itself)

Nothing fancy. In the Bay Area it's lots of people's choice:

  • Rail (BART) usually won't take you the last mile or two.
  • Car ownership is expensive (just having somewhere to park is either expensive or your car is likely to get broken into), plus many people never learned to drive, and parking when you go places is a pain.
  • The buses are awful.
  • Uber/Lyft aren't that expensive in this area, or weren't when pooling with other random people.
  • People are too lazy to cycle. :P (also bikes getting stolen all the time)

I think I underestimated how much of the Rationalist community was in the bay area. That fact alone resolves most of my confusion, thank you.

Do you have a car? Most rationalists in Berkeley don't, so unless it's nearby or you're doing one of a small set of public-transit compatible routes, Uber/Lyft are the only option.

Also note that the amortized cost of owning cars vs Lyft are moderately comparable, and if you’re not doing that much tracking Lyft / Uber comes out a lot cheaper

Unfortunately, a car is an unavoidable cost for me, I expect that is a large part of the difference.

For sure if you have a car it doesn’t make nearly as much sense to Uber places

I do have a car, but I don't even live in the bay area and didn't realize how many of you were in Berkeley. Makes sense now.

This is a very helpful analysis. I was independently undertaking a similar analysis, and it's nice to have this for comparison. I hadn't thought to exclude pedestrians, pedecyclists, and other non-occupants, nor of excluding single-vehicle crashes.

I think a some important pieces are missing from this analysis, as follows.

1) The final number, 548, is the number of miles that I must drive to accrue one micromort for all passenger vehicle occupants. But I am more interested in how many miles I have to drive to accrue one micromort for myself. The average (mean) car has 1.5 occupants. Assume that the average (mean) crash of interest involves two cars, and deaths are distributed roughly equally between each person involved in a crash, total of 3 people on average. So you would need to drive or ride 3 times as many miles as claimed by Josh's analysis to accrue one micromort for yourself. A micromort would accrue to any other passengers in your vehicle as well.

(If excluding all accidents with one car, then the assumption of the average accident including only 2 cars is actually off, as some accidents will include more than 2 cars.) (source for average passengers per vehicle) -- starting point for further honing this factor

2) If a passenger car gets in a crash with a larger vehicle such as a semi truck or a bus, likely the occupant of the passenger car will be injured much worse than the occupants of the larger vehicle. (I didn't look this up but it seems reasonable.) This would bias the results in the opposite direction from excluding motorcycles and pedestrians. More generally, we should be aware of cherry-picking which risk factors we remove from the analysis.

3) Deaths per 100M motor vehicle miles driven stayed approximately constant in a range of 1.0-1.2 per 100 million miles from 2009-2019 but went up dramatically in 2020 to 1.37 deaths per 100M miles (73 miles per micromort) and has stayed at a higher level or even increased in 2021 based on preliminary data.

4) Taking out 50% of deaths due to alcohol impairment, so as to account only for the other driver being drunk, seems fine. But I'm less confident that the average lesswrong reader does not drive distracted or drowsy. This seems like an area where we should be careful about being too confident due to self enhancement bias. On the other hand, there are other ways to drive more safely than the average driver. One of the most important of these is maintaining a safe following distance.

5) Depending on the context, we might be more interested in the micromorts per mile of interstate highway type driving in particular. This risk is about half the mortality risk per mile as compared with all driving.  (assuming "motorways" has a similar meaning to interstate-type highways)

This figure is relevant, for instance, in answering a question such as "What is my risk of highway death from taking a 1000-mile road trip, as compared with my risk of death from covid over the course of the same vacation?"

Josh's analysis mutiplies the initial figure of 91 miles per micromort by a factor of roughly 6. 

Jost  includes factors of approximately 
A) 1.51 for excluding motorcycles, bikes, and pedestrians
B) 1.78 for excluding single-car crashes
C) 1.86 for seatbelt wearing
D) 1.24 for drunk, distracted, and drowsy driving. This is dominated by drunk driving

I would reduce factor A, as I think it's balanced by cases where a passenger car hits a truck. More research on this could be helpful. For now, I'd say to get rid of half of it, make it 1.25
I would get rid of factor B, but replace it with a discretionary safe driving factor (see below)
I would keep factor C
I would keep most of factor D, maybe reduce it to 1.2.

Add (or rather multiply) in the following factors

E) Factor of 3 to account for only micromorts accruing to one traveler, not to co-passengers or occupants of other vehicles. This is based on an average of 3 people per crash, a figure that would benefit from further research.

F) Factor to account for being a safer than average driver (beyond the effect of wearing a seatbelt and not driving drunk). I think that it might be reasonable for especially safe drivers to use a factor of 2 or 3 here (which would be comparable to eliminating single-car crashes plus more). A factor of 10 seems like it would be too much. For myself as driver, I think I would be conservative and keep it at 1, absent further analysis. I am safer than average in some of my driving practices but less safe than average in other practices and skills.

One way to get a start at estimating this factor might be to look at your auto insurance rate and compare it to the average rate for comparable coverage of a comparable vehicle in your state. Insurance companies are in the business of rating risk after all.

G) Factor of 2 if considering only interstate highway driving (or other divided limited-access highways).

I would start with 73 miles per micromort, using the rate from 2020. Overall this gives me

73*1.25*1.86*1.2*3 = 611

611 miles per micromort for each vehicle occupant, or 1222 miles per micromort considering only interstate highway driving, to be adjusted further by a factor of up to 3 depending on one's beliefs about one's relative safety as a driver.

So a micromort is one millionth of a fatality, so there has been one fatality every 91 million miles