Largely missing from discussions around what course of action we should take during the COVID-19 pandemic, unfortunately, were statistical lives, their assumed value in dollars, and the quantitative tradeoff between the costs of decisions and the cost in DALY/QALYs. I really think that should be a regular exercise in any discussion of politics.
Happy to see this on HN, it's a fascinating topic.
I don't disagree, but it's a false dichotomy: we don't have to choose between the statistical view and boots on the ground. All forms of useful knowledge should be sought. Human brains are heavily biased towards anecdotes and stories, as opposed to cold calculated statistics... So that's the baseline we're starting from.
> I don't disagree, but it's a false dichotomy: we don't have to choose between the statistical view and boots on the ground. All forms of useful knowledge should be sought. Human brains are heavily biased towards anecdotes and stories, as opposed to cold calculated statistics...
Precisely my point. We can't do those statistics without narrowing the discussion to a finite set of dimensions that *may* be representative.
The more one works toward the doctorate, the more one realizes the squishiness of efforts to generalize from the nifty p-value to the population at large.
Especially where life-altering decisions are involved.
Those ascribing to themselves vast authorities are far too insulated from negative consequences. More skin in the game, please.
Isn't the current political drama entirely be aide one party focuses entirely on ground truth and anecdotes and presents no coherent connection or policy positions?
After a skim, I will say I don't understand why predictable or unpredictable does not explicitly address variance. That's my confusion.
The point may be how we differently value lives with or without faces. With reference to Schelling: "In many situations, people are less inclined to bear a particular cost or exert a particular effort to protect statistical lives than to protect the same number of identified lives."
It's a interesting paper with ties to lots of classical problems in clinical decision making, risk analysis, and statistical modeling. I've not seen it framed in quite this way.
I'm not sure "unpredictable" is quite the right term but maybe?
The problem is when you have very high impact events that are very low probability, the probability assigned to the event in the individual case (a factory explosion) is essentially zero, even though in aggregate at the group level you know these occur. Moreover, these very low-base rate probabilities are difficult to estimate accurately for various reasons, so the probability in a given case might actually be 10^-4 or 10^-100. So you end up deciding something probably won't happen, but at a societal level it will, because the probabilities are underestimated and there's lots of cases like this.
I think they'd say variance is relevant, but at the level of actual model uncertainty attached to an estimated probability or something?
The stats is a little underdeveloped, but the papers I've seen often are, except in bayesian low-base rate modeling papers, which tend to ignore the individual versus group prediction, model uncertainty, and utility combination.
I can't make sense of what they're trying to say here - they keep using the phrases "identified lives" and "statistical lives", but don't define them - they are talking about actual human lives here, right?
From the paper's definitions, "predictable" or "unpredictable" is from the POV of an agent whose action will probably cause some effect (predictable) or might cause some effect (unpredictable). The affected lives are identified - eg names and faces - or statistical - anonymous members of some population. The paper takes the ethical POV that from a policy perspective all lives are equivalent, albeit the loss of different lives can have different consequence in terms of liability.
In terms of the relationship between the agent and the potentially affected, neither empathy or group identity is mentioned in the paper, but that would be interesting to look at elsewhere. Hypothetically, at what point does "the other" that you treating anonymously and categorically become specific and individual to you?
Existing ethical discussion considers the differences in care for identified versus statistical lives. However, there has been little attention to the different degrees of care that are taken for different kinds of statistical lives. Here we argue that for a given number of statistical lives at stake, there will sometimes be different, and usually greater, care taken to protect predictable statistical lives, in which the number of lives that will be lost can be predicted fairly accurately, than for unpredictable statistical lives, where the lives are at stake because of a low-probability event, such that most likely no one will be affected by the decision but with low probability some lives will be at stake. One reason for this difference is the statistical challenge of estimating low probabilities, and in particular the tendency of common approaches to underestimate these probabilities. Another is the existence of rational incentives to treat unpredictable risks as if the probabilities were lower than they are. Some of these factors apply outside the pure economic context, to institutions, individuals, and governments. We argue that there is no ethical reason to treat unpredictable statistical lives differently from predictable statistical lives. Moreover, lives that are unpredictable from the perspective of an individual agent may become predictable when aggregated to the level of a societal decision. Underprotection of unpredictable statistical lives is a form of market failure that may need to be corrected by altering regulation, introducing compulsory liability insurance, or other social policies.
"Look at this child. She will die, unless you donate now" - an identified life. Humans are very sensitive to these.
"If we fund this drug, outlook for patients who get this type of cancer (Z cases per year) will improve from 50% to 80% survival over 5 years" - These are "more predictable" statistical lives. We don't know exactly who will be saved but we think we have a good handle on how many people it will be.
"If we don't put this control in the plant, we believe there's a 0.1% chance of an accident per year, which could cause up to 100,000 deaths in the surrounding area" - these are "less predictable" statistical lives. We know the impact is large, but we don't _really_ know how large. We know the probability is quite low, but we don't have an exact figure for that, either.
Numerically, if 100 people were dropping dead next to the plant per year that would be a pretty big deal, but instead this (in spherical cow terms) identical issue remains on the risk register, for now. Most likely, no one will come to any harm.
My intuition is that decision makers feel the risk (to them) of being blamed for unlikely events is quite low. The odds of a given tail risk manifesting during any given leader's tenure are also correspondingly low. The main exception seems to be terrorism; political leaders seem to be very sensitive to that.
Sometimes I think human risk estimates are primarily driven by how easily we can imagine the risk occurring to us, and the vividness of that imagining.
Most people seem to prefer the very diffuse but predictable risk of particulate emissions from coal plants over the minute but less predictable tail risk of a nuclear plant next door (i.e, the exact opposite preference to that described in the paper).
IMHO this can largely be explained at a population level by the fact that there are lots of movies about reactor meltdowns, but (as far as I know) no movies which show a particle emitted from a coal plant entering a person's lungs and their ensuing battle with cancer.
16 comments
[ 2.5 ms ] story [ 45.4 ms ] threadHappy to see this on HN, it's a fascinating topic.
Distant officials in high offices making Olympian decisions seem to produce mixed results on their best days.
Precisely my point. We can't do those statistics without narrowing the discussion to a finite set of dimensions that *may* be representative.
The more one works toward the doctorate, the more one realizes the squishiness of efforts to generalize from the nifty p-value to the population at large.
Especially where life-altering decisions are involved.
Those ascribing to themselves vast authorities are far too insulated from negative consequences. More skin in the game, please.
...
ie.: You don’t get your political base engaged and mobilized with statistical rigour. But objectively, if you could, it would make things better.
After a skim, I will say I don't understand why predictable or unpredictable does not explicitly address variance. That's my confusion.
The point may be how we differently value lives with or without faces. With reference to Schelling: "In many situations, people are less inclined to bear a particular cost or exert a particular effort to protect statistical lives than to protect the same number of identified lives."
I'm not sure "unpredictable" is quite the right term but maybe?
The problem is when you have very high impact events that are very low probability, the probability assigned to the event in the individual case (a factory explosion) is essentially zero, even though in aggregate at the group level you know these occur. Moreover, these very low-base rate probabilities are difficult to estimate accurately for various reasons, so the probability in a given case might actually be 10^-4 or 10^-100. So you end up deciding something probably won't happen, but at a societal level it will, because the probabilities are underestimated and there's lots of cases like this.
I think they'd say variance is relevant, but at the level of actual model uncertainty attached to an estimated probability or something?
The stats is a little underdeveloped, but the papers I've seen often are, except in bayesian low-base rate modeling papers, which tend to ignore the individual versus group prediction, model uncertainty, and utility combination.
In terms of the relationship between the agent and the potentially affected, neither empathy or group identity is mentioned in the paper, but that would be interesting to look at elsewhere. Hypothetically, at what point does "the other" that you treating anonymously and categorically become specific and individual to you?
Existing ethical discussion considers the differences in care for identified versus statistical lives. However, there has been little attention to the different degrees of care that are taken for different kinds of statistical lives. Here we argue that for a given number of statistical lives at stake, there will sometimes be different, and usually greater, care taken to protect predictable statistical lives, in which the number of lives that will be lost can be predicted fairly accurately, than for unpredictable statistical lives, where the lives are at stake because of a low-probability event, such that most likely no one will be affected by the decision but with low probability some lives will be at stake. One reason for this difference is the statistical challenge of estimating low probabilities, and in particular the tendency of common approaches to underestimate these probabilities. Another is the existence of rational incentives to treat unpredictable risks as if the probabilities were lower than they are. Some of these factors apply outside the pure economic context, to institutions, individuals, and governments. We argue that there is no ethical reason to treat unpredictable statistical lives differently from predictable statistical lives. Moreover, lives that are unpredictable from the perspective of an individual agent may become predictable when aggregated to the level of a societal decision. Underprotection of unpredictable statistical lives is a form of market failure that may need to be corrected by altering regulation, introducing compulsory liability insurance, or other social policies.
"If we fund this drug, outlook for patients who get this type of cancer (Z cases per year) will improve from 50% to 80% survival over 5 years" - These are "more predictable" statistical lives. We don't know exactly who will be saved but we think we have a good handle on how many people it will be.
"If we don't put this control in the plant, we believe there's a 0.1% chance of an accident per year, which could cause up to 100,000 deaths in the surrounding area" - these are "less predictable" statistical lives. We know the impact is large, but we don't _really_ know how large. We know the probability is quite low, but we don't have an exact figure for that, either.
Numerically, if 100 people were dropping dead next to the plant per year that would be a pretty big deal, but instead this (in spherical cow terms) identical issue remains on the risk register, for now. Most likely, no one will come to any harm.
My intuition is that decision makers feel the risk (to them) of being blamed for unlikely events is quite low. The odds of a given tail risk manifesting during any given leader's tenure are also correspondingly low. The main exception seems to be terrorism; political leaders seem to be very sensitive to that.
Sometimes I think human risk estimates are primarily driven by how easily we can imagine the risk occurring to us, and the vividness of that imagining.
Most people seem to prefer the very diffuse but predictable risk of particulate emissions from coal plants over the minute but less predictable tail risk of a nuclear plant next door (i.e, the exact opposite preference to that described in the paper).
IMHO this can largely be explained at a population level by the fact that there are lots of movies about reactor meltdowns, but (as far as I know) no movies which show a particle emitted from a coal plant entering a person's lungs and their ensuing battle with cancer.