An article published today on Reuters and elsewhere reads, "Israeli survey finds 3rd Pfizer vaccine dose has similar side effects to 2nd." Buried within this article is the following:

About 0.4% said they suffered from difficulty breathing, and 1% said they sought medical treatment due to one or more side effect.

This seemed quite bad to me and like a worrisome result. I sought information on how many sought medical treatment after the second shot. I could not find this information, but I did find:

only 51 of some 650,000 people to have received the Pfizer shot sought medical attention for symptoms suffered

from a December 2020 article on Israeli vaccination. Comparing the 1% to 51/650000 = 0.008% I found that the current frequency of side effects requiring medical attention was 128x the level found after dose 1. This seemed like a bad sign.

I then sought out more information about side effects post dose 2 in Israel, which I did not find. But instead I looked at the CDC's Advisory Committee on Immunization Practices’ Interim Recommendation for Moderna, and found the following:

The frequency of serious adverse events** observed was low in both the vaccine (1.0%) and placebo (1.0%) recipients

** Serious adverse events are defined as any untoward medical occurrence that results in death, is life-threatening, requires inpatient hospitalization or prolongation of existing hospitalization, or results in persistent disability/incapacity.

1. I can't believe that this was 1%! That seems surprisingly high (for either group). I expect the outside-of-trial data has not been nearly that magnitude.
2. This 1% matches the current Israeli data, and with a more restrictive definition, so the Israeli data no longer seems particularly worrisome in comparison, though I may dig in to this further. In general, I feel somewhat confused by the situation.

Sources: Reuter's article from today - https://www.reuters.com/business/healthcare-pharmaceuticals/israeli-survey-finds-3rd-pfizer-vaccine-dose-has-similar-side-effects-2nd-2021-08-08/

Article from December 2020 - https://www.timesofisrael.com/1-in-1000-israelis-report-mild-side-effects-from-vaccine/

CDC's Advisory Committee on Immunization Practices’ Interim Recommendation for Moderna: https://www.cdc.gov/mmwr/volumes/69/wr/mm695152e1.htm

EDIT: This article's statistics contrast with those of Reuter's, and show data very similar to the 1st shot: https://www.timesofisrael.com/of-600000-israelis-who-received-3rd-dose-fewer-than-50-reported-side-effects/

I recently had what I thought was an inspired idea: a Google Maps for safety. This hypothetical product would allow you to:

1. Route you in such a way that maximizes safety, and/or
2. Route you in such a way that maximizes your safety & time-efficiency trade-off, according to your own input of the valuation of your time and orientation toward safety

First, I wanted to validate that such a tradeoff between safety and efficiency exists. Initial results seemed to validate my prior:

1. The WHO says crashes increase 2.5% for every 1 km/h increase in speed.
2. The Insurance Institute for Highway Safety (IIHS) reports fatalities increase by 8.5% when there is a 5mph increase in speed limit on highways, and 2.8% for the same speed limit increase on other roads.
3. The National Safety Council (NSC) cites speed as a factor in 26% of crashes.

Despite these figures, I felt none of these, on their own, provided sufficient information to analyze the scale of safety gains to be had. The WHO source was outdated and without context (although there was a link to follow for more information that I didn’t see at that time), the IIHS merely talked about increases in speed limits for two types of roads, rather than actual changes in speed that results nor relative safety of the two types of roads, and the NSC provided a merely binary result.

So I went searching for more data.

And I discovered that the US National Highway Traffic Safety Administration (NHTSA) releases a shocking amount of data on every fatal car crash. There’s useful data, such as what type of road the crash happened at, what the nature of the collision was, information on injuries and fatalities, whether alcohol was involved, etc.

(There’s also a surprising amount of information that I expect might make some people uncomfortable. For every crash this data includes VIN number of vehicles involved, driver's height, weight, age, gender, whether they owned the vehicle, driving and criminal history. It also includes the exact time, date, and location of the crash.)

I used the former (useful) information for analysis on this question. Given the initial data found, I figured that one way to approximate the available gains and tradeoffs was to analyze safety-gained from turning on the “Avoid Highways” setting on Google Maps.

After some experimentation and reading others’ thoughts, it became clear that this setting avoids interstates (I-5, I-10, I-15, etc.) but not other types of highways. I used NHTSA data to calculate the number of deaths occurring on interstates vs. on other roads, and found that the Federal High Administration provides data on the number of miles driven in the US per year by type of road. Using these two sources of data, I calculated the number of miles driven per fatality on interstates vs on all other roads (for 2019):

Interstates: ~180 million miles / fatality

All Other Roads: ~104 million miles / fatality

It turns out that interstates appear to be (at least on this metric) safer than non-interstates! This was surprising to me, given the earlier cited results that pointed to speed being dangerous.

I decided that I’d do more validation of this result if this was surprising to most people, but wouldn’t perform more validation if this wasn’t. Asking around, it looks like this result is not surprising to most:

From Effective Altruism Polls:

From EA Corner Discord:

And from the LessWrong Slack:

So first of all, good job community, on seemingly being calibrated. Second, I followed my earlier plan and did not look further into this result given that it was aligned with most people's priors. And finally, I do think this makes the expected value of a Google Maps for safety significantly lower than my prior.

Assuming this result would hold through further validation, there are still ways that a Google Maps for safety could be beneficial. A few examples of this:

1. Seeing if there are other road-type routing rules that would provide safer outcomes.
2. Using more specific data, such as crash reports by road, to identify particularly dangerous roads / intersections and avoid them.
3. There seem to be some behavioral economics-like results with road safety that could be leveraged during route design. For example, apparently roads with narrower lanes are safer than roads with wide lanes, presumably because narrower lanes have the effect of people driving more slowly, while having a lower effect on increased accident rate.
4. Digging further into data on factors that contribute to crashes (alcohol, weather, distraction, evening, etc.) could reveal patterns that provide clues as to the safer route by situation.

I think this could be a really cool app to have, and I’d support its development if someone were to take it on, but it seems like a big project. I was sad and surprised to find that the potential quick win of turning on the “avoid highways” option is seemingly not a win at all (although there exist confounders and further validation would be beneficial).

Would anyone like the domain alignai.org ? Otherwise I'll probably let it expire (bought for a previous org, which doesn't want it).

EDIT (7/27/23) After very preliminary research, I now think "telling people to ride in SUVs or vans instead of sedans" may turn out to be worthwhile.

As I’m working on derisking research, I’m particularly aware of what I think of as “whales”... risks or opportunities that are much larger in scale than most other things I’ll likely investigate.

There are some things that I consider to be widely-known whales, such as diet and exercise.

There are others that I consider to be more neglected, and also less certain to be large scale (based on my priors). Air quality is the best example of this sort of whale, though 3-8 other potential risks or interventions are on my mind as candidates for this, and I won’t be surprised to discover a couple whales that did not seem to be so prior to investigation.

I thought that road safety and driving was a widely-known whale. Based on a preliminary investigation (more on what this means), I now tentatively think it is not.

This preliminary analysis yielded an expected ~17 days of lost life as a result of driving for an average 30 year old in the US over the next 10 years.

I’m not sure how many of these 17 days an intervention could capture. I suspect most likely readers of what I’d write already grab the low-hanging fruit of e.g. not driving while impaired and wearing a seatbelt. So it does not seem probable that I would discover an intervention that alleviated even 30% (~5 days) of risk. Furthermore, I suspect most interventions in this space could have large inconvenience or time costs, causing greater reduction in the expected gain of my research in this space.

While this analysis does neglect loss of QALDs due to injury, which I don’t know the scale of, I predict they are unlikely to greatly affect this conclusion.

The 10 year timeframe may seem odd to some. But if we assume that self-driving cars of a certain ability level will greatly increase the safety of vehicle travel, which I personally believe, then 10 years may be even longer than the relevant window for investigation. Metaculus predicts L3 autonomous vehicles by the end of 2022, L4 autonomous vehicles by the end of 2024, and L5 autonomous vehicles by mid 2031. It’s not entirely clear to me at which of these stages most of the safety benefits are likely to occur, nor how long widespread use will take after these are first available, but it does seem to me as though the dangers of car travel, at least for most people who are likely to read my content, will not persist long into the future.

I have some context for effect sizes I think I’m likely to find with various interventions. I have preliminary estimates for interventions affecting air quality & nuclear risk, and more certain estimates for interventions on smoke detectors and HPV vaccination. With that context, road safety does not seem to particularly differentiate itself from much else I expect to investigate. With this discovery that road safety does not seem to be a ‘whale’, I tentatively think I will not further investigate it in the near future.

Very quick, likely highly flawed and inaccurate Fermi on the worthwhileness of riding in SUVs or vans instead of sedans:

~50% safer per mile maybe

~20% more expensive to buy / rent / rideshare

~1 day per year gained in expectation for a 35 yo

~$10k spent on vehicle travel per year per person

Cost: ~$2k / day of life gained

Worthwhile if valuing a year at: ~$600k or greater