What Econ is 'particularly EA-relevant?'

Use-cases: defining value, meta-ethics

Use-cases: Empirical and policy

David Reinstein

"Why do this, why try to generate a best take on 'my welfare function'" without having gone through the literature (much, only listening to podcasts etc)?

1. Uninformed theorizing may bring a new perspective which would be constrained by reading

2. Narrating "figuring out things" may bring insights to others (help newbies learn about population ethics, give experts a sense of how an economist would take this on)

Simple example, I argue it is consistent with RW intuitions (expand)

- Prob (1-p) of 'small world S' with NL = 100 million people indexed by i

- Prob p of 'large world L' with NL = 100 billion people indexed by i

As a 'baseline', think of p=1/2 for intuition; assume that if you didn't act, this would be its value.

Ignore the 'non-identity' issue.

Each individual has a value function vi(W),

call it the vector VW across all individuals in world W

For simplicity, maybe we assume that every individual (in a particular world W) gets the same v_i(S), but this can be affected by further arguments, discussed below

For now assume all value/happiness is positive, and/or we have a clear way of weighing negatives and positives (as well as among positive states)

You can 'donate' (or take costly actions) by choosing some vector g1, g2, g3 each of which may potentially affect either:

1. the values each individual gets (happiness whatever) in state S

2. the values each individual gets in state L

3. the probability of a 'large world L', probability p

Justifications

Aside: Why not include 'extinction' or 0 population?

• I think this conjures non-utilitarian, non 'add up across people' values

• We already know there will be at least some people ... alive today

I want a 'consequentialist utilitarian welfare function':

I think I should care about outcomes (probability-weighted perhaps) mediated through how the people themselves value their own states. I think I should act so that my actions do best to maximize this. I don't want the 'thing I am maximizing' to depend on my actions or on how I think about my actions or 'why' I chose something.

"Deciding between"

1. 'Average utilitarian' (or some function like that)

2. Total utilitarian

3. Some representation of 'person affecting' (but it's hard to achieve that with a simple SWF

   1. Considering allowing a function of both the VW vector of happinesses and the population in a state VW ... but that does seem a bit of A FUDGE

Simple cases:

I. Suppose you can only affect the probability of a large world, but not welfare, you can affect p

• Here the 'average utility' model seems appropriate for a person with PAV; you don't care which world exists, so you won't invest in p

• True, if you 'could' affect happinesses, it does seem weird and unfair that you are valuing each person in the Small world so much more ... but as you can't, this doesn't guide you to an 'unjustifiable decision'

Possible variation: you can affect both p and VS but not VL ... would probably preserve this

II. Suppose you can only affect the happinesses, but not the probability of each state

• Here the 'total utilitarian' model proscribes the 'right action'. For the same cost, you would improve the life of someone in world S instead of someone in world L, but you would only value this at approximately 1/p = 2x as much in our example ... only because each person in world S is about twice as likely to exist as each person in world L

• But it feels weird to a PAV, because with this SWF, if you could increase the probability of world L, you would do so. The actual value function puts (in our example) 1000 times more value on world L

"If you can affect both" ... how to reconcile it?

1. A 'correction term' to either the average or total model that implies that 'increasing p should have little or no value'

2. A 'value of my actions (and outcomes?)' that does treat the components of my decision (g0, g1, g2) differently insofar as these affect p versus the value functions

Reading syllabi/curated/outlines

Rhys-Bernard: reading syllabus - An introduction to global priorities research for economists

Trammel: "Economic theory and Global Prioritization'

"Web books" and content

Reinstein (Exeter Msc) Microeconomics - touches on EA and related 'decision theory' fundamentals; want to adapt

Other lists of syllabi and resources

David Rhys-Bernard curated Pablo Stafforini's list (below), adding to it further.

"Economics+ for EA+ and vice/versa"

Who is this for and why

1. EA's and fellow-travelers:

• The principles and practice of effective altruism are closely connected to the principles of Economics. EA's engage economic concepts, arguments, and theoretical and empirical "results" throughout their work. However, these principles are sometimes misunderstood or mis-stated, sometimes alluded to but not considered in depth. The terminology itself can sometimes be muddled between EA, rationalist, decision theory, computer science, and academic economics communities. This leads to confusion and barriers to engagement.

• I also get the sense that EA's are somewhat over-optimistic and over-broad about 'what Economics tells us' (possible example: GiveWell on 'log utility'); often a result that holds under only very-specific conditions is stated as a generally accepted truth. On the other hand

• Some useful areas of Economics seem neglected in EA, e.g., (my own impressions):

  • "Basic" supply-demand-production systems, aggregation, empirical general and partial equilibrium models ... are very relevant to considering policy interventions, especially in animal welfare.

    • Aside: This has been neglected by academic economists because its seen as non-deep, and neglected by private-sector/government economists because there is no established 'animal welfare policy audience'

    • Some progress may be happening: see the

  • Some decision-theoretic concepts and models of preferences... with interesting implications for 'social preferences' and 'communicating preferences to machines'

  • Non-utilitarian preference relations like 'lexicographic' ... seems neglected in the discussion of population ethics (it's OK to have 'non-continuity', perhaps)

  • Social Welfare Functions and Social Choice

  • Comparability of 'utility' across individuals and time, revealed preference

  • Economics (esp. field and natural experiments, revealed preferences) of eliciting risk and time preferences

  • Some work in 'the Economics of other-regarding behavior' (charitable giving and 'crowding out', consumer altruism...) (But I do think some parts of this literature are stuck in a trap of confusion about the application of models like 'warm glow').

2. Economists & co.

• Understanding how individuals and firms make choices, how these aggregate in market (and other settings) is the original 'what' of Economics. The original 'why do we care about this' is 'to understand what will achieve the best outcomes for humanity (and perhaps beyond). But IMO it has been somewhat waylaid by the desire to demonstrate cleverness and rigorous extensions of existing models. It has also been distracted by parochialism: to the extent there is a 'policy audience', it is typically the US Government.

  • Economists will benefit from an approach that returns to the 'global welfare' question first and foremost, and benefit from engaging with the EA community, which is practically trying to achieve these goals

• This 'integration' will bring in interesting concepts from Philosophy and Decision Science that Economists may have neglected

• EA questions and goals provide a new research agenda and the opportunity to apply core concepts and tools from Economics in ways that may be more directly relevant than the more mainstream ('fix the US economy'...) targets. Some brainstorms on this (needs clarification)...

  • Preference axioms (transitivity etc), VnM Axioms, impossibility theorems etc:

    • Application to 'social preferences' and 'aggregated social preferences'

    • Time discounting: weigh present/future

    • Population ethics; weigh definite/possible individuals

    • Uncertainty and 'preference over outcomes versus over impact

  • Defining preferences and constraints: tools for aligning AI?

  • The 'aggregation from individual optimization problems' may be unreliable for predicting chaotic human systems, but more relevant for 'aligning AI'

  • Measuring and assessing 'tradeoffs between income gains at different levels' (e.g., for GiveWell and GiveDirectly) with different empirical and theory-driven approaches

  • Aggregating social welfare functions and other social preference formulations with epistemic and moral uncertainty

  • Implications of GE models for animal welfare interventions; a new set of value measurements and possible interventions; not just the 'market failures' approach

Making this work: implementation

Opportunities, threats, important issues