Quantifying Risk

post by dpandey · 2021-10-12T14:41:47.401Z · LW · GW · 3 comments

Epistemic Status: Providing a jumping-off point for further discussion, not intended to be a final authority by any means.

With COVID, a lot of people in the rationality community have had to fundamentally reassess what kind of risks they're willing to handle. We've used microCOVID extensively, tried to figure out what our chances of severe illness or death are, and decided on what precautions to take relative to that.  

Regardless, it's incredibly difficult to fundamentally assess risk or the tradeoffs of risk. The EPA values a statistical life at $7.4 million based on surveys of people asking what they valued tiny risks to their life to be. So should we measure risk in dollars, trying to quantify risk using the universal currency of utility, and figuring out exactly what the potential impact of that risk is? I'd argue against it. 

My moral values don't prohibit valuing life in dollars, nor would I want them to, but I think that in general it's pretty socially frowned upon to use monetary values as a way to measure statistical life; moreover, we get into questions of what trying to quantify life using money means, and generally I think people would value their own lives very differently from $7.4 million dollars if they were ever in that situation. 

Microdeaths might work better as a statistical measuring tool. But I think that the best way to quantify any risk might be to relate it directly to a risk that people take every day and generally know is pretty dangerous—driving. Already, driving is considered a good comparison point; when trying to evaluate COVID risk recently, I've consistently compared my chances of dying at any given activity to my chances of dying on the car ride there. (The car ride usually wins, often by an order of magnitude or so). So I think the best way to evaluate risk might actually be in miles, where each mile is equivalent to roughly 12 nanodeaths, or a 12/1,000,000,000 chance of dying.

What does this mean for risks like COVID? Well, combining CDC data with the Economist's calculator from March, there was a 0.1% chance of death for an unvaccinated 35-year-old male who got COVID, which did not change substantially with Delta. Vaccination lowered that risk by a further (pessimistically) 92%, so a fully vaccinated 35-year-old male has around 8,000 microdeaths by getting COVID, or 666,666 miles of driving. Thus, a single microCOVID for this fully vaccinated person is "equivalent" in death risk to around 2/3 of a mile.

We don't have a good understanding of the risks of a lot of activities yet, and this calculation does not take into account things like hospitalization risk and contagion to more vulnerable people, so take any calculation you do with a grain of salt. But hopefully, comparisons like this will make it easier to visualize the actual risk of something, and be more or less careful as necessary.


Comments sorted by top scores.

comment by davidad · 2021-10-12T18:26:33.489Z · LW(p) · GW(p)

The key problem with valuing life in dollars, "the universal currency of utility," is that utility is not affinely related to money, especially in extreme cases. However, utility is affine in probability, even in extreme cases.

That's why micromorts work, but they only work as a one-dimensional sufficient statistic when the only outcomes under consideration are "death" and "status quo".

In cases like infectious disease, where we have outcomes like "hospitalization for some time" and "transmission to others" (and I would add to your list "long-term sequelae" and "flu-like illness") whose proportional impact on utility for different people can vary widely, a one-dimensional sufficient statistic can't fully faithfully be used to compare kinds of risk with different profiles of such outcomes (e.g. COVID risk to vehicle accident risk). This approach can be used to compare different sources of COVID-profile risk, and of course that's what a microCOVID is, but pretty much anything distinct from COVID has a different profile of ways it could impact utility; even the Delta variant arguably has a different profile from wild-type.

All that said, as long as we're clear that we're only considering fatal and status-quo outcomes, I do like the idea of an alternately scaled unit, a mortmile := 0.012 micromorts.

Replies from: dpandey
comment by dpandey · 2021-10-12T19:45:14.475Z · LW(p) · GW(p)

Yeah, I think that that's a good point about the one-dimensionality of any unit of measure used to assess risk. It might be possible to effectively start measuring in quality-adjusted life minutes or hours, but that quite quickly becomes a massive headache to calculate, even if it's more accurate to the actual impact on people. I think that using a unit like the mortmile is a good way to effectively make back-of-the-envelope calculations to assess the degree of risk and quickly understand just how risky something is, especially when differences are measured in orders of magnitude (as they usually are).

comment by cistrane · 2021-10-15T11:10:24.889Z · LW(p) · GW(p)

There are between 30 and 40 thousand traffic fatalities per year in US. That means, if we use your numbers, that collectively Americans drive about 3 trillion miles a year. Fascinating, never thought to sum all the miles we drive.