Hacker News Reader

> game of assigning quantitative value to lives is almost never going to help you make better decisions

Quality adjusted life years is part of the analytical basis of funding of interventions in the UK's national health system (NHS).

Does this change your position?

Not really. National healthcare systems are maybe the one reason why I said "almost never" and not "never". When your problem is literally and directly, not indirectly, about assigning finite resources where lives are on the line, that's the cases where a model that relates the two is most likely to be useful. And even then most healthcare systems don't even try.

When the effects of sacrificing lives are more disconnected, the model becomes less useful. In particular, I don't think these sorts of models are useful for making decisions about how to respond to COVID, which is what this article is about.

It doesn't change mine! This came into vogue in the early 70s when they wanted to cut funding as much as possible to make the operation more efficient while still providing a high level of care, according to these "metrics" like QLAY, which are fundamentally broken (as are and have been all attempts to reduce to metrics). It managed to operate without them with a great deal of satisfaction from both the providers and the patients until this came in, and then things got really shitty while improving along the metrics so it was nobody's fault because it was just "science".

> But it's going to be some complex function.

Or you can pick a number from Wikipedia [0]. Somewhere between $50 thousand and $10 million.

I agree with what I suspect is your central point; which is when it comes to COVID-19 there are other more important uncertainties than the value cost of a life. But it is useful for an order-of-magnitude check to see if the response is overblown.

[0] https://en.wikipedia.org/wiki/Value_of_life#Estimates_of_the...

I'm not sure it's useful even as an order of magnitude check, because there's so much uncertainty about how many lives different policies cost. People who believe a response is too loose inevitably also think it kills more people than their opponents; those who believe a response is too strict inevitably think it saves fewer.

Honestly, this looks to me like a word salad -- I can't pull anything out of this other than "models are useless, so don't bother", which is a non-answer.

No, models are useful.

However, models of "value of life vs. other-thing" are almost always useless.

Why? Because you don't care about that question. You care about other questions.

And, crucially, the models you use to answer those other questions tend to have so much uncertainty that your model of "value of life vs. other-thing" turns out to not really contribute anything useful to the analysis.

If I ask for a policy that maximizes the value of some function $f(x) = \theta_1 * x^n + \theta_2 * c$, for example, then will you focus your effort on modeling $n$ or $c$? What if I tell you that the best you can do is get $n$ down to an interval $[-100, 100]$? Will you then go on to think a bunch about what to do with $c$, especially if you know that whatever $c$ is it's probably between -1 and 1?

My suggestion is that, in most cases, "value of life vs. other-thing" is the $c$ in this example.

In the case of COVID, you want all your smart modelers thinking about how to minimize the chance of protected exponential growth within any region. If you do that, you have the right answer. The exact trade-off between lives and other things doesn't matter because, with covid, __preventing localized protected exponential growth is always the right answer__. The "lives vs other-thing" question is a complete distraction in the decision space.

> I can't pull anything out of this other than "models are useless, so don't bother", which is a non-answer.

Is it? Models are always modeling something! Modeling the effect of reddit meme spread on covid-19 spread is probably useless. That doesn't mean cluster SIR models are useless.

Don't model things that don't help you make decisions.

Sometimes it's fun to do things that are pointless.

> However, models of "value of life vs. other-thing" are almost always useless. > Why? Because you don't care about that question. You care about other questions.

The “value of a statistical life” (or to use a better term, “value of a prevented fatality”), is weighed against other things in a large number of public policy choices. For example: environmental pollution, biodiversity, quality of landscapes, noise pollution, economic activity. It’s also important to aim for reasonable levels of prevention of different risk categories: food safety, road safety, air transport safety, medical safety in hospitals, industrial safety, etc. If public policy (via regulatory obligations for instance) uses very different implicit values for a prevented fatality, it means that more statistical lives could have been saved by allocating prevention effort in a different way.

I think their point is something like what's discussed here:

https://news.ycombinator.com/item?id=22915837

I'd distill it as:

1. Every model should be presented alongside a discussion of whether the fitting process is well-posed[1]

2. If the fitting process is ill-posed, the model cannot be used to guide policy

3. If the fitting process is well-posed, but only via Bayesian regularization of some kind, before being used to guide policy it should be checked for sensitivity to choice of regularization, and the regularization scrutinized in exhaustive detail

4. Compositions of models must be checked for well-posedness

https://en.wikipedia.org/wiki/Well-posed_problem

Wow, thank you so much.