I’d like to hear more EA discussion on discount rates. Much of policy analysis involves unilaterally discounting future benefits. For example, an economist might say “Let’s value eating one apple today the same as eating four apples ten years from now.” The professionals I’ve spoken with who do this sort of analysis say that discounting is justified because it’s a natural part of human decision-making. Psychologically, it’s pretty clear that most people make decisions given much greater weight to instant or near-term gratification.
However, I’d be more likely to put this under a ‘cognitive bias’ than a ‘terminal value.’ I think most people, upon deep reflection, would realize that eating an apple is eating an apple just the same no matter when it happens.
Removing the (utility) discounting in policy analysis seems like it could do a lot of good for future people, who matter the just the same as we do. Under modern methods, we choose small amounts of short-term good in exchange for really bad outcomes decades or centuries from now.
Does anyone disagree? If not, how tractable is this goal?
There’s quite a bit of internal discussion on this at CEA.
There are several reasons for discounting. Some of them are quite correctly applied in social policy contexts, whereas some are not applicable (as the case you highlight, which is often called ‘pure rate of time preference’). They are also sometimes misapplied.
I do think that helping to make sure that discounting is done correctly according to context is an important goal, and this is something that the Global Priorities Project may push for. But trying to remove discounting altogether in analysis may harm future people rather than help them.
This paper by Dasgupta has some good discussion of the different purposes of discounting (but I wouldn’t take too much from its discussion of eta).
Obviously, if I’m going to die unless I eat that apple in the next ten minutes, the apple has extremely high value now and zero after 10 minutes.
Extending that idea, you are integrating across all the probabilities that the apple will become useless or reduce in value between now and when you’re going to get it.
Why is it exponential? Maybe stretching a bit, but I would guess that the apple changes in value according to a poisson process 1 where the dominating force is “the apple becomes useless to you”.
Obviously, if I’m going to die unless I eat that apple in the next ten minutes, the apple has extremely high value now and zero after 10 minutes.
Right, but this is a special case and not an argument for a general discount rate.
Extending that idea, you are integrating across all the probabilities that the apple will become useless or reduce in value between now and when you’re going to get it.
Yes, heuristic-based discount rates like this (e.g. due to general uncertainty) are helpful and should be applied when necessary. But that’s different from just a utility discount rate (i.e. future identical events are just less valuable).
Why is it exponential? Maybe stretching a bit, but I would guess that the apple changes in value according to a poisson process [1] where the dominating force is “the apple becomes useless to you”.
There are instrumental effects of saving a life earlier or later in time. It’s not clear to me which should be better (and this may change over time), but it seems quite plausible that there should be a small (I’d guess well under 1% p.a.) positive or negative discount rate on this.
Right. On the other hand people later in time generally have higher productivity, so perhaps they’d be able to achieve more. This could be a bigger or smaller effect (although if forced to guess I’d marginally prefer the life now to one later).
We can also consider how lives saved now will save more future lives, leading to more achievement with even higher productivity. It seems like it might be turtles all the way down. Figuring out flow-through effects of present lives saved versus the discount rate value of future lives seems difficult.
The turtles all the way down problem is something which crops up in many cases looking at growth.
The basic way to deal with it is usually to sidestep it: so rather than cash out in some terminal units (like number of lives saved through history), convert everything into a unit we can get a grip on (e.g. as good as saving how many lives in 2014). Of course that can be dependent on hard-to-estimate figures, but they’re at least empirical figures which relate to near-term consequences.
I’m not sure exactly what you guys mean by turtles all the way down but I have some relevant links. Do you mean that growth continues indefinitely? Nick Beckstead has argued that it should not. A related concept is the question of whether we should be always favouring saving future lives rather than consuming resources. Seth Baum has argued that we should not.
‘Turtles all the way down’ is a silly metaphor from philosophy, and cosmology, representing a difficult premise, similar to Schrodinger’s Cat. It refers to the problem of the ‘prime mover’, or ‘first cause’, in the universe, e.g., who created God?, what happened before the Big Bang?, etc. The idea is that it’s so absurd to figure out what the absolute origin of everything is that the world might as well rest on the back of a turtle, who itself sits upon an infinite pile of turtles below it.
The analogy isn’t perfect, I admit. What I meant is this:
There’s a trade-off between saving lives in the present due to the flow-through effects they’ll have in terms of saving lives in the future, and just saving a greater number of lives in the far future.
Time is so indiscrete, and the world so full of variables, that I can’t think of how to solve to problem of how much do we neglect saving lives at one point in the present or near-future, for the purpose of saving lives in the future further ahead.
Finding the perfect slice(s) of time to focus upon seems like trying to get to the bottom of an infinite stack of turtles to me.
Yes, I’m familiar with ‘turtles all the way down’ in general. For this question of finding the ideal time-slice to focus on, it’s the second link (Seth Baum’s post) that is relevant. He addresses the issue of an indefinitely postponed splurge—the idea that you might always have to wait before consuming goods by countering that we are consuming all of the time, just by staying alive.
That’s a philosophical counter but I could also just give a more practical one—there are plenty of other people who will fuel consumption. If you think that the future is neglected, then you don’t need to have an exact plan for when consumption should occur in order to invest in it.
Fair question. I was meaning something like growth continues indefinitely.
If I wanted a careful statement I’d say it wasn’t turtles all the way down (as Nick Beckstead argues), but that it’s turtles down as far as we can see. For many practical purposes these are indistinguishable in terms of raising problems we need new methods for thinking about—though it does kill some arguments which try to use the tail of the “all the way down” assumption.
In a similar vein, infinity can often be a good working approximation for very large finite numbers—but if you treat that literally and start trying to play Hotel Infinity tricks, you get in trouble.
[Discount Rates]
I’d like to hear more EA discussion on discount rates. Much of policy analysis involves unilaterally discounting future benefits. For example, an economist might say “Let’s value eating one apple today the same as eating four apples ten years from now.” The professionals I’ve spoken with who do this sort of analysis say that discounting is justified because it’s a natural part of human decision-making. Psychologically, it’s pretty clear that most people make decisions given much greater weight to instant or near-term gratification.
However, I’d be more likely to put this under a ‘cognitive bias’ than a ‘terminal value.’ I think most people, upon deep reflection, would realize that eating an apple is eating an apple just the same no matter when it happens.
Removing the (utility) discounting in policy analysis seems like it could do a lot of good for future people, who matter the just the same as we do. Under modern methods, we choose small amounts of short-term good in exchange for really bad outcomes decades or centuries from now.
Does anyone disagree? If not, how tractable is this goal?
Edit: Here’s a good piece on the topic: http://www.givingwhatwecan.org/blog/2013-04-04/was-tutankhamun-a-billion-times-more-important-than-you
There’s quite a bit of internal discussion on this at CEA.
There are several reasons for discounting. Some of them are quite correctly applied in social policy contexts, whereas some are not applicable (as the case you highlight, which is often called ‘pure rate of time preference’). They are also sometimes misapplied.
I do think that helping to make sure that discounting is done correctly according to context is an important goal, and this is something that the Global Priorities Project may push for. But trying to remove discounting altogether in analysis may harm future people rather than help them.
This paper by Dasgupta has some good discussion of the different purposes of discounting (but I wouldn’t take too much from its discussion of eta).
In addition to upvoting, I want to mention that strikes me as something very worthwhile for the Global Priorities Project to try.
Obviously, if I’m going to die unless I eat that apple in the next ten minutes, the apple has extremely high value now and zero after 10 minutes.
Extending that idea, you are integrating across all the probabilities that the apple will become useless or reduce in value between now and when you’re going to get it.
Why is it exponential? Maybe stretching a bit, but I would guess that the apple changes in value according to a poisson process 1 where the dominating force is “the apple becomes useless to you”.
Right, but this is a special case and not an argument for a general discount rate.
Yes, heuristic-based discount rates like this (e.g. due to general uncertainty) are helpful and should be applied when necessary. But that’s different from just a utility discount rate (i.e. future identical events are just less valuable).
Sure :)
Many assets have compounding value (e.g., interest) that comes from owning things earlier. But I don’t think human life is one of those things.
There are instrumental effects of saving a life earlier or later in time. It’s not clear to me which should be better (and this may change over time), but it seems quite plausible that there should be a small (I’d guess well under 1% p.a.) positive or negative discount rate on this.
It’s worth pointing out that lives saved now are in a better position to save more lives (c.f., flow-through effects).
Right. On the other hand people later in time generally have higher productivity, so perhaps they’d be able to achieve more. This could be a bigger or smaller effect (although if forced to guess I’d marginally prefer the life now to one later).
We can also consider how lives saved now will save more future lives, leading to more achievement with even higher productivity. It seems like it might be turtles all the way down. Figuring out flow-through effects of present lives saved versus the discount rate value of future lives seems difficult.
The turtles all the way down problem is something which crops up in many cases looking at growth.
The basic way to deal with it is usually to sidestep it: so rather than cash out in some terminal units (like number of lives saved through history), convert everything into a unit we can get a grip on (e.g. as good as saving how many lives in 2014). Of course that can be dependent on hard-to-estimate figures, but they’re at least empirical figures which relate to near-term consequences.
I’m not sure exactly what you guys mean by turtles all the way down but I have some relevant links. Do you mean that growth continues indefinitely? Nick Beckstead has argued that it should not. A related concept is the question of whether we should be always favouring saving future lives rather than consuming resources. Seth Baum has argued that we should not.
‘Turtles all the way down’ is a silly metaphor from philosophy, and cosmology, representing a difficult premise, similar to Schrodinger’s Cat. It refers to the problem of the ‘prime mover’, or ‘first cause’, in the universe, e.g., who created God?, what happened before the Big Bang?, etc. The idea is that it’s so absurd to figure out what the absolute origin of everything is that the world might as well rest on the back of a turtle, who itself sits upon an infinite pile of turtles below it.
The analogy isn’t perfect, I admit. What I meant is this:
There’s a trade-off between saving lives in the present due to the flow-through effects they’ll have in terms of saving lives in the future, and just saving a greater number of lives in the far future.
Time is so indiscrete, and the world so full of variables, that I can’t think of how to solve to problem of how much do we neglect saving lives at one point in the present or near-future, for the purpose of saving lives in the future further ahead.
Finding the perfect slice(s) of time to focus upon seems like trying to get to the bottom of an infinite stack of turtles to me.
Yes, I’m familiar with ‘turtles all the way down’ in general. For this question of finding the ideal time-slice to focus on, it’s the second link (Seth Baum’s post) that is relevant. He addresses the issue of an indefinitely postponed splurge—the idea that you might always have to wait before consuming goods by countering that we are consuming all of the time, just by staying alive.
That’s a philosophical counter but I could also just give a more practical one—there are plenty of other people who will fuel consumption. If you think that the future is neglected, then you don’t need to have an exact plan for when consumption should occur in order to invest in it.
Thanks, noted. That makes sense.
Fair question. I was meaning something like growth continues indefinitely.
If I wanted a careful statement I’d say it wasn’t turtles all the way down (as Nick Beckstead argues), but that it’s turtles down as far as we can see. For many practical purposes these are indistinguishable in terms of raising problems we need new methods for thinking about—though it does kill some arguments which try to use the tail of the “all the way down” assumption.
In a similar vein, infinity can often be a good working approximation for very large finite numbers—but if you treat that literally and start trying to play Hotel Infinity tricks, you get in trouble.